University of Dundee

DOCTOR OF PHILOSOPHY

Road Pothole Detection Method Using Built-in Sensors In Smartphone

Zhang, Dalong

Award date:
2017

Awarding institution:
University of Dundee

Link to publication

General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners
and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.

• Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
• You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal
Take down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately
and investigate your claim.

Download date: 09. Feb. 2025

 1

UNIVERSITY OF DUNDEE

Road Pothole Detection Method

Using Built-in Sensors In
Smartphone

by Dalong Zhang

An initial submission of thesis for examination for the degree of

Doctor of Philosophy

Division of Physics
School of Engineering, Physics & Mathematics
University of Dundee

2017
 2

Contents
List of Figures ................................................................................................................ 6

List of Tables ............................................................................................................... 10

Abbreviations ............................................................................................................... 11

Acknowledgements ...................................................................................................... 12

Declaration of authorship ............................................................................................. 13

Abstract ........................................................................................................................ 14

Chapter 1. Introduction ............................................................................................ 15

1.1 The research project ...................................................................................... 15

1.2 Background ................................................................................................... 15

1.3 This thesis...................................................................................................... 16

Chapter 2. Literature review .................................................................................... 18

2.1 Pothole detection method .............................................................................. 18

2.1.1 Contact measurement ............................................................................. 18

2.1.2 Non-contact measurement ..................................................................... 21

2.1.3 Sub-conclusion....................................................................................... 22

2.2 Modelling ...................................................................................................... 23

2.2.1 Vehicle modelling .................................................................................. 23

2.2.1.1 2-DOFs suspension system of 1/4 vehicle ......................................... 25

2.2.1.2 4-DOFs suspension system of 1/2 vehicle ......................................... 26
2.2.1.3 7-DOFs suspension system of the whole vehicle ............................... 29
2.2.2 Tyre modelling ....................................................................................... 30

2.2.2.1 General theoretical model of tyre dynamic ........................................ 31

2.2.2.2 Semi-empirical model of tyre dynamic .............................................. 31
2.3 Signal Processing .......................................................................................... 32

2.3.1 Coordinate axis transformation .............................................................. 33

2.3.2 De-noise ................................................................................................. 33

2.3.2.1 Kalman filter ...................................................................................... 34

2.3.2.2 FFT filter ............................................................................................ 34
 3

2.3.2.3 Wavelet............................................................................................... 35
2.4 Conclusion..................................................................................................... 36

Chapter 3. Detection system design ........................................................................ 38

3.1 Detection method requirement analysis ........................................................ 38

3.2 Detection method .......................................................................................... 38

3.3 Demand analysis of each module .................................................................. 40

3.3.1 CPU ........................................................................................................ 40

3.3.2 Sensor..................................................................................................... 40

3.3.3 Memory .................................................................................................. 42

3.3.4 Network.................................................................................................. 43

3.3.5 Battery .................................................................................................... 44

3.3.6 UI ........................................................................................................... 44

3.3.7 Size......................................................................................................... 44

3.3.8 Conclusion ............................................................................................. 44

3.4 Classification and evaluation of embedded systems ..................................... 45

3.4.1 Sun SPOT............................................................................................... 45

3.4.2 Raspberry Pi ........................................................................................... 45

3.4.3 NUC ....................................................................................................... 46

3.4.4 Mobile phone ......................................................................................... 46

3.4.4.1 The development history of mobile phone......................................... 46

3.4.4.1.1 The development of mobile phone network: from 1G to 4G ...... 46
3.4.4.1.2 Development of hardware of smart-phone .................................. 47
3.4.4.1.3 Development of Software on Smart-phone ................................. 53
3.4.4.2 Features of mobile phone ................................................................... 56
3.5 Conclusion..................................................................................................... 57

Chapter 4. Experiment design ................................................................................. 59

4.1 Introduction to road for experiment .............................................................. 59

4.2 Vehicles to be tested ...................................................................................... 65

 4

4.3 Collection software ....................................................................................... 66

4.4 Experiment scheme ....................................................................................... 68

4.5 Summary of chapter ...................................................................................... 75

Chapter 5. Modelling and simulation ...................................................................... 77

5.1 Simulation of suspension system .................................................................. 77

5.1.1 2-DOFs suspension system of 1/4 vehicle ............................................. 77

5.1.2 4-DOFs suspension system of 1/2 vehicle ............................................. 80

5.1.3 7-DOFs suspension system of the whole vehicle .................................. 84

5.1.4 Simulate suspension using MATLAB/Simulink ................................... 85

5.2 Tyre simulation ............................................................................................. 90

5.3 Conclusion..................................................................................................... 93

Chapter 6. Data processing ...................................................................................... 95

6.1 Coordinate correction .................................................................................... 95

6.2 Convert Acceleration Data into Distance ...................................................... 99

6.2.1 Time domain integral method ................................................................ 99

6.2.2 Frequency domain integral method ..................................................... 100

6.2.3 Discussion ............................................................................................ 100

6.3 De-noising ................................................................................................... 108

6.3.1 FFT ....................................................................................................... 109

6.3.2 Wavelet de-noise .................................................................................. 112

6.3.2.1 Application of wavelet ..................................................................... 112

6.3.2.2 Applying wavelets to reduce noise for signal. ................................. 112
6.3.2.3 Discussion ........................................................................................ 115
6.4 Conclusion................................................................................................... 115

Chapter 7. Data Processing and Discussion .......................................................... 117

7.1 Data processing ........................................................................................... 117

7.2 Comparison between actual measurement of Tay Road Bridge pothole

position and data processing result......................................................................... 140
 5

7.3 Forth Road Bridge Data Processing result .................................................. 143

7.4 Real pothole detect test ............................................................................... 144

7.5 Discussion ................................................................................................... 152

Chapter 8. Conclusion and Future Work ............................................................... 154

8.1 Conclusion and Prospects............................................................................ 154

8.1.1 Conclusion ........................................................................................... 154

8.1.2 Future work and considerations ........................................................... 155

8.2 Design of Operating Mode .......................................................................... 156

8.3 Future works on real pothole detection methods ........................................ 159

Reference ................................................................................................................... 161

Appendix A: Detailed deduction process of bounce height h without considering energy

loss in tyre simulation ................................................................................................ 164

Appendix B: Detailed deduction process of bounce height h with considering energy

loss in tyre simulation ................................................................................................ 166

Appendix C: Deduction process of PCA ................................................................... 167

Appendix D: Time domain integral method .............................................................. 168

Appendix E: Frequency domain integral method ...................................................... 169

Appendix F: DFT and FFT ........................................................................................ 170

Appendix G: MATLAB source code of FFT noise reduction processing procedure 171

Appendix H: Deduction process of wavelet .............................................................. 172

Appendix I: Tay Road Bridge Data processing: Route 2 to 5 ................................... 174

Appendix J: Real road data process: Route 2 to Route 4 ........................................... 199

Appendix K: Publications .......................................................................................... 214

 6

List of Figures
Figure 2-1 Accumulation test vehicle .......................................................................... 19
Figure 2-2 2-DOFs suspension system of 1/4 vehicle ................................................. 26
Figure 2-3 4-DOFs suspension system of 1/2 vehicle (pitch model)........................... 27
Figure 2-4 4-DOFs suspension system of 1/2 vehicle (roll model) ............................. 28
Figure 2-5 7-DOFs suspension system of the whole vehicle ....................................... 29
Figure 3-1 Typical gyroscope ...................................................................................... 41
Figure 3-2 Microstructure of gyroscope chips based on MEMS (Micro-
electromechanical Systems, is a mini system that integrates optical systems, drive
components, mechanical components and electronic control system in a single unit)
technology .................................................................................................................... 42
Figure 4-1 Dundee Tay Road Bridge ........................................................................... 59
Figure 4-2 3D model of Tay Road Bridge ................................................................... 60
Figure 4-3 Typical expansion joint on Tay Road Bridge ............................................ 60
Figure 4-4 Forth Road Bridge ...................................................................................... 61
Figure 4-5 3D model of Forth Road Bridge................................................................. 62
Figure 4-6 Three sections from south to north of Forth Road Bridge: S1 (408m), S2
( 1006m) and S3 (408m) .............................................................................................. 63
Figure 4-7 Parameters measured and imported into Matlab ........................................ 67
Figure 4-8 Axes direction defined by accelerometer and gyroscope ........................... 69
Figure 4-9 1145642 points collected from Dundee to Edinburgh on 24/09/2012 with
sample rate of 200Hz. Sample time about 5728 seconds............................................. 70
Figure 4-10 Zoom in of previous diagram, containing about 6100 points, sampled in
about 30 seconds .......................................................................................................... 71
Figure 4-11 Further zoom in of previous diagram. About 3100 points measured in 15
seconds ......................................................................................................................... 71
Figure 4-12 1145642 points from Dundee to Edinburgh, measured on 24/09/2012 with
sample frequency of 200Hz. Sample time about 5728 seconds................................... 72
Figure 4-13 Zoom in of 4-5, containing about 25000 points, or 125 seconds ............. 73
Figure 4-14 Dundee to Edinburgh, 24/09/2012. Google maps can only display a small
part of geometry information of kml file, so that the length displayed in this screenshot
is 6.91 miles. ................................................................................................................ 74
Figure 4-15 Sample of speed data, 5613 points from Dundee to Edinburgh, collected
on 24/09/2012 .............................................................................................................. 75
 7

Figure 5-1 2-DOFs simulation model of suspension system of 1/4 vehicle in

MATLAB/SIMULINK ................................................................................................ 78
Figure 5-2 Use sine wave as the road input excitation, x-axis represents time (measured
by second), and y-axis represents displacement of tyre in the vertical direction
(measured by meter). ................................................................................................... 79
Figure 5-3 Output wave of 2-DOF vehicle model ....................................................... 79
Figure 5-4 4-DOFs simulation model of suspension system of 1/2 vehicle in
MATLAB/SIMULINK. (a) is the graphical block diagram, (b) shows the detail of the
grey block of Figure (a), (c) shows the detail of the yellow block of Figure (a). ........ 82
Figure 5-5 Use sin wave as the road input excitation .................................................. 83
Figure 5-6 Vertical distance output wave of 4-DOF vehicle model ............................ 83
Figure 5-7 7-DOFs simulation model of suspension system of the whole vehicle in
MATLAB/SIMULINK ................................................................................................ 84
Figure 5-8 Parameter subsystem of 7-DOFs simulation model of suspension system of
the................................................................................................................................. 84
Figure 5-9 A sample suspension model designed in simulink using simdriveline ...... 85
Figure 5-10 Sub-design of Figure 5-9, the front and rear suspension ......................... 86
Figure 5-11 step input of road, height changes to 0.01 at time point 7 ....................... 87
Figure 5-12 The vertical displacement of vehicle after it passes the step change of road.
height=0.01m ............................................................................................................... 87
Figure 5-13 Relationship between height of step change of road, displacement of
vehicle body and maximum displacement of vehicle body ......................................... 89
Figure 5-14 Tyre with center O passes pothole AB with width of w. The velocity of
tyre before passing is v, the velocity after it bounces up from road is v’ which can be
decomposed to a component of vx’ horizontally, and a component of vy’ vertically. . 90
Figure 6-1 Vehicle coordinate system ......................................................................... 95
Figure 6-2 Nature coordinate system ........................................................................... 96
Figure 6-3 Image of equation 6-15 ......................................................................... 101
Figure 6-4 Integral result of equation 6-15 ................................................................ 102
Figure 6-5 Double integral result of equation 6-15 ................................................... 103
Figure 6-6 Integral result of equation 6-15 without DC component ......................... 104
Figure 6-7 Second integral result of equation 6-15 without DC component ............. 105
Figure 6-8 Integral result of equation 6-15 using frequency domain method ........... 106
 8

Figure 6-9 Second integral result of equation 6-15 using frequency domain method
.................................................................................................................................... 106
Figure 6-10 Original input signal ............................................................................... 110
Figure 6-11 Input signal with random noise .............................................................. 111
Figure 6-12 Noise reduction result ............................................................................ 112
Figure 6-13 Steps in the process of one-dimensional signal reduces noise by using
wavelets...................................................................................................................... 113
Figure 6-14 Signal decomposing using wavelet ........................................................ 114
Figure 6-15 Noise reduction using wavelet decomposing ......................................... 115
Figure 7-1 Initialisation process................................................................................. 119
Figure 7-2 Common process ...................................................................................... 121
Figure 7-3 Pothole detect process .............................................................................. 123
Figure 7-4 The acceleration data of Route 1 .............................................................. 124
Figure 7-5 De-noise data on the three axes with wavelet of Route 1 ........................ 126
Figure 7-6 Baseline of acceleration data of Route 1 .................................................. 127
Figure 7-7 AC component of accelerometer data of Route 1 .................................... 128
Figure 7-8 Use distance as the abscissa in Route 1 ................................................... 129
Figure 7-9 Gyroscope data of Route 1 ....................................................................... 130
Figure 7-10 Denoise the gyroscope data of Route 1 .................................................. 131
Figure 7-11 Baseline of gyroscope data of Route 1 ................................................... 132
Figure 7-12 AC component of gyroscope data of Route 1 ........................................ 133
Figure 7-13 Conduct distance correction with Dist2 as the abscissa ......................... 134
Figure 7-14 GPS speed data from 517th second to 666th second. X axis is time (second),
y axis is velocity (m/s) ............................................................................................... 135
Figure 7-15 Square calculation applied to every point GPS speed data. X axis is points,
y axis is square of velocity ......................................................................................... 136
Figure 7-16 Threshold and Acc1zDenoiseAC. X axis is points, y axis is amplitude (m/s²)
.................................................................................................................................... 137
Figure 7-17 Normal expansion joint style of No. 1-29 and 34-42 ............................. 141
Figure 7-18 Abnormal expansion joint style of No. 30, 31, 32, 33 ........................... 142
Figure 7-19 Tay Road Bridge result vs 3D model ..................................................... 143
Figure 7-20 Forth Road Bridge result vs 3D model .................................................. 144
Figure 7-21 Road for test in Camperdown Country park (shown as the red route)... 145
Figure 7-22 Three acceleration data of Route 1......................................................... 147
 9

Figure 7-23 Wavelet de-noised result of Z axis of Route 1 ....................................... 148

Figure 7-24 Baseline of Z axis of Route 1 ................................................................. 149
Figure 7-25 Z axis of Route without baseline ............................................................ 149
Figure 7-26 De-noised Z signal with threadhold. The x axis is distance................... 150
 10

List of Tables
Table 2-1 Comparison of road test methods ................................................................ 19
Table 2-2 Subsystems of the vehicle with their elements ............................................ 25
Table 3-1 Technology and the highest speed of various types of network .................. 43
Table 3-2 Power Dissipation and Precisions of Acceleration Sensors and Gyroscopes:
Samsung Galaxy Note 1 (N-7000) and iPhone 5 ......................................................... 51
Table 3-3 Milestone Events for the Development of Smart-phones............................ 53
Table 3-4 Comparison of development platform of three mainstream smartphones... 54
Table 3-5 Number of software in App Store of three mainstream smartphones ......... 54
Table 3-6 Typical software on smartphones making use of sensors/GPS/camera ...... 56
Table 4-1 Two detection vehicles used in experiments ............................................... 66
Table 4-2 Differences between detection devices ....................................................... 68
Table 5-1 vertical displacement of vehicle body with various step change of road (unit:
m) ................................................................................................................................. 89
Table 6-1 Error obtained by calculating two integral methods.................................. 108
Table 7-1 Tay Road Bridge pothole result ................................................................. 139
Table 7-2 Pothole ID and the interval distance between potholes ............................. 141
Table 7-3 Distances between each pothole and the start point .................................. 146
Table 7-4 Detection result of Route 1 ........................................................................ 151
Table 7-5 The detection result of all 4 routes ............................................................ 152
 11

Abbreviations

GPS Global Positioning System

VBI Vehicle-Bridge Interaction
IRI International Roughness Index
APDS Automatic Pavement-Distress-Survey System
PCES Pavement Cracks Evaluation System
DOF Degrees of Freedom
FFT Fast Fourier Transform
IIR Infinite Impulse Response
FIR Finite Impulse Response
WT Wavelet Transform
DWT Discrete Wavelet Transform
PCA Principal Component Analysis
MEMS Micro-Electro-Mechanical System
GSM Global System for Mobile communication
FDMA Frequency Division Multiple Access
TDMA Time Division Multiple Access
GPRS General Packet Radio Service
EDGE Enhanced Data rates for GSM Evolution
HSPA High Speed Packet Access
WCDMA Wideband Code Division Multiple Access
NUC Next Unit of Computing
SATNAV Satellite Navigation
VR Virtual Reality
 12

Acknowledgements

I would never have been able to finish my thesis without the guidance of my PHD

supervisors Prof. Allan Gillespie and Dr Charles Main. It has always been a great

pleasure to work with them in all the past years.

I would like to thank them for giving me the opportunity to undertake this project and

to work in a professional research environment.

Furthermore, I would like to thank my parents, my wife and my daughter for all their

support and help.

Finally, I would like to thank all members of the Materials and Photonics Systems

(MAPS) group at the University of Dundee for providing welcoming, professional and

productive surroundings for the work of this project. In particular, I wish to express my

gratitude toward Dr. Fu Yu for searching and applying of my PHD project. I thank you

all who have encouraged me in everything I do throughout my time in Dundee.

 13

Declaration of authorship
I declare that the thesis entitled

Road Pothole Detection Method Using Built-in Sensors In

Smartphone

and the work presented in it are my own. I confirm that:

 this work was done wholly while in candidature for research degree at this

University;

 where any part of this thesis has previously been submitted for a degree or any

other qualification at this University or any other institution, this has been

clearly stated;

 where I have quoted from the work of others, the source is always given. With

the exception of such quotations, this thesis is entirely my own work;

 I have acknowledged all main sources of help;

 where the thesis is based on work done by myself jointly with others, I have

made clear exactly what was done by others and what I have contributed

myself.

Dalong Zhang
August 2015
 14

Abstract

In this thesis, the smartphone is installed on vehicle as a tool to detect potholes on the
road. The data collected and pre-processed by smartphone is then analysed to obtain
the result.
The history and status of pothole detection study is firstly discussed. The traditional
pothole detection method has two disadvantages: 1, low efficiency; 2, limited detection
area.
To work out a new method, the suspension models and tyre models are studied. Also
the axis-correction and de-noise in signal process discussed for data analysis. Based on
these, the requirement of detection system is analysed, and the features of embedded
systems are discussed, which leads to the conclusion that smartphone is the best
hardware for pothole detection.
Then the smartphone (3 mobiles are chosen: Samsung Note 1, Nexus 7 and iPhone 5)
is fixed on vehicle (Nissan Micra K11 and Saab 93) as experiment platform. The Tay
Road Bridge is chosen as experiment road. The software Sensor Insider Pro and Sensor
Data are chosen to collect data.
Simulations about vehicle and tyre are done by MATLAB Simulink. From the
simulation the relationship between the bouncing height and the speed of vehicle is
obtained, and is used as threshold in pothole detection.
The experimental data from acceleration sensor and gyroscope is processed by
MATLAB using axis-correction and wavelet transform. Together with velocity data
gathered by GPS and threshold (calculated from speed), the position of pothole is
obtained. Comparing with road, the detection has good accuracy, which proves the
feasibility of this pothole detection method.
At the end of thesis, other potential field of application of smartphone is discussed.
 15

Chapter 1. Introduction

1.1 The research project

This research project puts forward a new method for pothole detection, in which a
smart-phone is installed on a running vehicle and the data is collected from smart-phone
sensors - such as acceleration sensors or gyroscopes - and the GPS built in the smart-
phone. By processing the collected data, the potholes on the road on which the vehicle
traverse through can be located.
On the basis of looking back the development of smart-phones, the thesis analyses the
feasibility and advantages of pothole detection using smart-phones.
Through simulating the suspension model of vehicles, I find out the relationship
between pothole and the output signal of sensors, and then put forward a method to
determine pothole by variable threshold value related to the vehicle running speed.
Taking Tay Road Bridge as my experimental subject, I designed and implemented
experiment to collect the data, which is then processed through a wavelet-based noise-
reduction process and produces the final results.
In addition, I compare and discuss the differences between results of acceleration
sensors and that of gyroscopes. At last, I initially discuss other areas that smart-phones
may be applicant on, including fault inspection of vehicles and pulse detection of
human body, with some case studies included.

1.2 Background
There are two methods to detect road pothole. The first method is public reporting,
which is economic in manpower and material resources, because the roads having larger
traffic flow may draw more attentions. However there may be misreport. Furthermore,
the roads having poor traffic flow may draw a little care and have less information
available. The second method is more common, which is to use a special pothole–
detection vehicle running on the road to detect pothole and record its location (Lin et
al., 2008). This kind of vehicle may record its vertical vibrations when its wheels run
through the road, or scan and analyse the road by using a vehicle-mounted laser scanner
or an ultrasonic scanner (Huston et al., 2000, Wang et al., 2011), or by using a camera
to form images under visible light which is more advanced (Lin and Yayu, 2010). Each
of these detection methods has advantages and flaws. Generally speaking, the direct-
contact detection method has lower error rate but it works slower so the detection
efficient is lower, and it can only detect potholes in the locations where the wheels
 16

contacted. The indirect detection method works faster but its error rate is higher. At
present most of the techniques are still in experimental stage or the proof stage.
No matter what kind of detection-vehicle is used, it is demanded that the vehicle
traverse through each inch of the road so as to collect and obtain complete data.
Moreover, the road pothole varies with time. A complex road network requires a long
time to detect, thus only constantly updated data are practical, which is a huge job
requiring lots of manpower and materials. Under such conditions it is hard to improve
the detection efficiency. Meanwhile, with the development of the road network, more
and more roads need to be detected; thus the expenditure on this rises year by year.
According to figures, the annual expenditure on road pothole repair is up to $1.3 billion
in USA (Cheng et al., 1998). At the same time, the loss caused by pothole is up to $4.8
billion. Therefore a new pothole detection method is in demand, which shall be high-
efficient, low cost, reliable and easy to be widely promoted and have precise output
results. This is the main direction discussed in this thesis.
It takes time to detect and repair of potholes. Generally, the length of repair time
depends on repair device (and its level of automation), manpower and size of pothole.
For small potholes, simple repair is enough; however, large pothole may require
retreading of a long section of road, which consumes long period of time. The length
of detection time, on the other hand, depends on number of detection vehicle and length
of road network, increase the number of detection vehicles surely will increase
efficiency of detection, unfortunately the increase of length of road - which is always
increasing - decrease the efficiency.

1.3 This thesis

The following questions are discussed in this thesis:
1. In which way the traditional methods can be improved? Is there a new method that
can exceed the traditional methods?
2. What equipments are required?
3. Which physical measurements should be kept?
4. What algorithm is required to analyse the data and obtain the result?
5. The new method should be proved to meet our goal.
Based on the study of previous work, a method of measuring road pothole based on
smartphone built in sensor is proposed in this thesis. The smartphone is fixed inside the
 17

vehicle, and the vibration data of the car body is obtained and processed by collecting
the data of the built-in accelerator. The noise data and baseline drift are removed by
signal processing of the vibration data, and the signals caused by road pothole are
reserved. The vehicle body and tyre modules are designed and simulated, and the
relationship between vehicle speed, pothole width and vibration amplitude is obtained,
and the dynamic threshold based on the speed is proposed. Dundee Tay Road Bridge is
used to simulate pothole on expansion joint. The experimental results show high
accuracy and little misjudgement.
The following is the structure of the paper:
Chapter 1 is this chapter.
Chapter 2 summarizes and analyzes the previous work.
Chapter 3 studied the data acquisition equipment, and designed the experimental
method.
Chapter 4 and 5 establish the dynamic model of the common vehicle suspension and
tyre, and give the simulation results.
Chapter 6 designs the methods and steps of data processing.
Chapter 7 data on the experiment are processed, and the results are discussed.
Chapter 8 discusses the previous work and describe the future work on real pothole
detection methods.
 18

Chapter 2. Literature review

In Chapter 1, five questions are brought out for my aid of study. First question: In which
way the traditional methods can be improved? Is there a new method that can exceed
the traditional methods? This question is answered in Section 2.1 with a browse of
previous work of pothole detection methods. With the method chosen, the question two:
equipment is discussed in Section 2.2 including various vehicle models. At last in
Section 2.3, literatures about the question three and four are surveyed. The question
five is not in the scope of this chapter and will be discussed in Chapter 7.

2.1 Pothole detection method

2.1.1 Contact measurement

In the contact measurement, the sensor directly contacts the road, or is installed on
another device that has direct contact with the road.
The International Roughness Index is a standard roughness test index defined by The
World Bank (Sayers, 1998). The standard calculating program of IRI was published in
its No.46 report, 1982, in which a one-fourth simulation (similar to a single-wheel
trailer) is used. During the mobile is driving on the known section with the speed of
80km/h, the accumulated displacement of suspension system within a certain distance
is calculated.
Then after the standard deviation of the flatness measurement device is determined by
the continuum flatness instrument, the standard deviation of flatness in every 100
meters is calculated from the displacement of every 10 centimetres.
The existing test methods are mainly special instruments or testing vehicles. The most
typical one is the flat monitor assembled with test vehicle (Lin et al., 2008). A 1/4 model
is used by IRI for the measurement, in which the vehicle runs with speed of 80km/s and
the cumulative vertical displacement of suspension is kept as output.
 19

Movement

Axle Speed

Figure 2-1 Accumulation test vehicle

Table 2-1 Comparison of road test methods

Type Complexity Continuous or Efficiency Measurement

Intermittent Index

3-meter ruler simple intermittent low Maximum

clearance section

Continuous relatively continuous high Standard

flatness complex differential cross

tester section

Accumulatio relatively continuous high VBI

n tester complex

Pavement complex continuous high IRI

inspection

car

Smart-phones, i.e. mobile phones with modern intelligent operating system, appear in
recent ten years (such as the first one, Moto A6188 in 1999), while the smart-phones
having built-in sensors such as acceleration sensors appears much later, such as Sony
ERSSION W580C in 2007. Therefore the idea to detect pothole with smart-phones
emerges in recent years. In 2008, researchers of MIT (Eriksson et al., 2008) detect road
 20

potholes by adopting special equipment: P2, which includes a GPS and a three-axis
acceleration sensor set on a vehicle to collect signals. In 2009, Swin adopted wavelet
analysis to process the flatness data of road. In the same year, Spanish researchers used
a mobile sensor to record the data of the vehicle so as to implement detection in the site
of the accident (Cadenas et al., 2009). In this project, the researchers simulated an
Android mobile phone which can automatically contact with the emergency call centre
using webservice when an accident occurs. In 2011, Latvia’s researchers announced a
method to implement real-time detection of pothole with Android smart-phones with
built-in acceleration sensors (Mednis et al., 2011, Strazdins et al., 2011). The searchers
used four kinds of smart-phones including Samsung and HTC to carry out the
experiment, gathering Z-axis data and obtaining results with the correct rate of 90%. In
the same year, Italian researchers (Bujari et al., 2012) announced a method to
implement road crossing recognition by using three-axis acceleration data of smart-
phones with high-pass filtering and pattern recognition. In 2012, Indian researchers
installed Android smart-phone Google Nexus S and HTC Wildfire on the Suzuki
Access 125 vehicle to obtain brake data so as to evaluate the road traffic condition
(Bhoraskar et al., 2012). In addition, researchers of Taiwan also installed a HTC
Diamond mobile phone (Windows Mobile operating system) and an external GPS on a
motorcycle to inspect the road flatness (IRI) with a three-axis acceleration sensor (Tai
et al., 2010). Similarly, the researchers of Microsoft Research India installed a HP iPAQ
hw6965 mobile phone (Windows Mobile operating system) and an external three-axis
acceleration sensor on the model vehicle to inspect brake, whistle and bump (Mohan et
al., 2008).
All those work indicate that using built-in sensors of smart-phones to detect potholes is
feasible. However, limited by the age, (no gyroscope is used in any of above researches,
only acceleration sensor and GPS (Mathur et al., 2010)) almost all of the above-
mentioned methods adopted three-axis acceleration sensors (built-in or external) in the
testing platform. In addition, only several papers involved simple high-pass filtering in
noise-reduction process of the data (Promwongsa et al., 2014), and other papers
involved nothing about this. Thus, the inspection results were not ideal. Moreover, there
is no analysis on vehicle characteristics in any research, while the relationship between
feature of road surface and built-in sensor can only be obtained by analysing the vehicle
characteristics, which means that the data process lacks theory support and can only be
done on experimental level. Furthermore, no research considers the difference in result
 21

brought by difference of vehicles. The applicability of these methods is not wide

enough. There will be better results if researchers go into a further research.

2.1.2 Non-contact measurement

In the contact measurement method, the contact measurement method can only provide
information of part of road contacted with the tire. To obtain the data of whole road,
the detection vehicle has to run repeatedly on the same road.
The non-contact measurement method uses imaging technology to complete testing.
There are two sub-systems: Image acquisition subsystem and Image processing
subsystem. The Image acquisition subsystem captures the image of road by using video
cameras fixed on testing vehicles, and then transform the image to digital data and store
it (Lin and Yayu, 2010). The Image processing subsystem analyse the digital data by
manual identity or machine vision recognition and get position of potholes, and stores
the result (Georgopoulos et al., 1995).
The image of road is normally obtained by visible light cameras (Nejad et al., 2011).
However ultrasonic, infrared or laser scanning technology may also be used (Huston et
al., 2000, Wang et al., 2011). In general, the grey-scale of image is kept and used for
analysis (Acosta et al., 1992), because the change of light source may also change the
colour of images and finally affect the analysis result if colour of image is kept. To
maintain the stability of the light into the camera, an active light source is commonly
installed on the testing vehicle (Perttunen et al., 2011).
The Image processing can be divided to several modules such as pre-processing,
segmentation, feature extraction, classification, decision making and evaluation (Hou
et al., 2007). The purpose of pre-processing is to identify and normalise areas of
different brightness. The feature extraction obtains the geometry and statistical
parameters of the image, or identify the feature of the image by image transformation.
Image classification divides the image categories and provides service for decision
making and evaluation.
To detect minor changes on the road, the camera should have high enough resolution,
and fast enough imaging speed. At the same time, the speed of vehicle cannot be too
high, otherwise it is a great challenge to the image acquisition and storage system.
The French road management department LCPC developed the road surface damaged
Dolly (GERPHO) in early 1970. The GERPHO works at night with special lighting
system to take photograph of road, so that it can avoid the effects of the environment
 22

light source. The cost of GERPHO photography is high. In addition, it requires

operators to interactively work together and cannot process detection work
automatically.
Another non-contact measurement system, which is called Automatic Pavement-
Distress-Survey System (APDS), was released in 1990 by Fukuhara in Japan (Fukuhara
et al., 1990). It works at night with laser scanning technology, with the maximum
detection speed is 10 km/h. APDS stores data on tapes, and processes data with 512
MC86020 microprocessors. The disadvantage of this system is that it is too big and
complex, and can only work at night.
Association of MHM in United States developed the automated road image analysis
(ARIATM) in the late 1980’s. ARIATM uses 2 cameras to obtain image of road surface.
It takes 15 minutes to process 1.609 km road (Mohajeri et al., 1991).
Between late 1980’s and early 90’s, the evaluation system of pavement cracks (PCES)
was developed by United States Earth Technology Corporation. This non-contact
measurement system uses line-scan camera which is more advanced than the traditional
surface-scan to gain road data, but requires special data acquisition card and high
intensity lighting. Due to the immature image processing technique and high cost of
data acquisition card and data processing software, no funding was provided to further
this study.
In 1995, a detection system named Large-Area Laser Scanner with Holographic
Detector Optics for Real-Time Recognition of Cracks in Road Surfaces (CREHOS) was
published by Dr. Max Monti of tress analysis laboratory (IMAC) of Swiss Federal
Institute of technology (EPFL). CREHOS scans road with a laser beam, and collects
and pre-processes analog data (Monti et al., 1995).
In late 1990’s the automatic road analyser ARAN was developed by Canadian
Roadware Company. Its camera has high intensity, high speed flash, which frees the
quality of video from effect of environment brightness, shadows of trees or the position
of the sun. According to the introduction of the company, it is the only commercially
available automatic road analyser in the world. But in fact ARAN is not fully automatic,
and still requires manual intervention to work.

2.1.3 Sub-conclusion
Both contact and non-contact measurement methods have advantages and
disadvantages. Contact measurement method has high accuracy and stability, and is
 23

suitable for all kinds of terrain and conditions. But its measuring field is very narrow,
which makes it inefficient.
Non-contact measurement method can scan a large area at a time by using image
processing method, thus it is faster. But it has higher error rate. Plus it is easily
influenced by the illumination of the surrounding environment and is not so stable. The
accuracy of non-contact measurement method is very difficult to be increased, unless a
revolutionary new algorithm appears.
However, it is possible to keep the advantage of contact measurement method and
increase its efficiency by using new measurement technologies. Therefore in this study,
the contact measurement method is used. The new measurement technology is
described in Chapter 3.

2.2 Modelling

2.2.1 Vehicle modelling

When the vehicle is running on the road, the vibration caused by the unevenness of the
road, pothole and speed bump will spread innerly to the vehicle through wheels,
suspension and vehicle body. In the experimental scheme, as the smart-phone with
built-in sensor is fixed in the centre of the vehicle, the acceleration shown on sensor is
considered as the acceleration of the vehicle. Assuming that the vehicle is running
forward at a constant speed, the measured value in the vertical direction of the
acceleration sensor shall have a certain relationship with roughness degree of the road,
which is determined by the response characteristics of the structure of the vehicle.
Therefore to research the vibration of the vehicle when it traverses through the pothole,
the structure of the vehicle shall be analysed (Dupont et al., 2008). Meanwhile, in regard
to signal processing, the vehicle can be regarded as a system and the road condition is
an input signal for the system, so the vibration detected by the smartphone actually is a
response of the vehicle system to signals of pothole on the road which is an output
signal (Li et al, 2008).
Further, as a whole system, the vehicle is made up of various subsystems such as tyres
(Elmadany, 2012), suspension system, damper, steering system, engine/power
transmission system and vehicle body (as shown in the Table 2-2) (Wu et al., 2013, Xie
et al., 2009). Each subsystem is approximate to having no internal relative motion and
they are connected with hinges, springs, dampers or rubber elements (Hunt, 1991). In
modelling procedure, to research the integrity property of the vehicle system, each
 24

subsystem is studied first, so that models for subsystems are established with range of
their physical parameters estimated. At the same time, subsystems that are non-relative
to the project are ignored. In addition, subsystems that have no relative motion (or no
significant motion) are combined together as one bigger subsystem to improve the
efficiency and accuracy of model (Lu et al., 2013). On the other hand, simplify some
less important subsystems and combine multiple subsystems between which there is no
relative motion or the relative motion is not important to improve the processing
efficiency, in the meantime cutting the loss of accuracy.
 25

Table 2-2 Subsystems of the vehicle with their elements

Engine Cylinder block, crankshaft connecting rod mechanism, air valve,

intake system, exhaust system, fuel supply system, ignition system,
cooling system, lubricating system, starting system
Chassis Transmission system clutch, transmission, drive shaft, drive axle
Driving system frame, front axle, drive axle, wheel housing
(including steering wheels and driving
wheels), suspension
Steering system steering and steering mechanism
Brake device power supply device, control device,
transmission device and brake
car body Front sheet metal parts, cab, car carriage
Electrical equipment Battery, engine starting system, ignition system, lighting system and
signal device

Actually, for this research project, the core study object is the interaction force between
vehicle and road. So adopt the most typical digital simulation method which is
establishing digital vehicle simulation model, to get numerical results of interaction
force between vehicle and road via simulation analysis with regard to the signal output
of pothole on the road.
Degrees of freedom (DOF) is the number of independent parameters that define
vehicle’s configuration. It is the number of parameters that determine the state of the
vehicle system (Zhang, 2010).

2.2.1.1 2-DOFs suspension system of 1/4 vehicle

This model focuses on the degree of freedom of vertical motion of sprung mass and un-
sprung mass. It studies the effect of stiffness and damping of suspension and tires (Yan
and Zhou, 2011, Aleksander and Iljoong, 1992).
 26

2
2

1
1

Figure 2-2 2-DOFs suspension system of 1/4 vehicle

Based on the Lagrangian formula, the summary of pressure on tyre from vehicle and
gravity of tyre should balance the support from ground on tyre; at the same time, the
gravity of vehicle should balance with support from suspension and spring.
Corresponding dynamics equation is as followed:
m1 z1  c ( z1  z2 )  k2 ( z1  z 2 )  k1 ( z1  q )  0
 (2.1)
 m2 z2  c ( z2  z1 )  k2 ( z2  z1 )  0

In which m1 is mass of tire, m2 is mass of 1/4 vehicle body, k1 is rigidity of tire,

k2 is rigidity of suspension spring, c is damp of suspension, z1 is vertical
displacement of tire, z2 is vertical displacement of vehicle body, q is vertical
displacement of ground.
With only two degrees of freedom, the 1/4 vehicle model can only show the vertical
motion of one tire. A more complex model is needed, which should contain at least one
front tire and one rear tire. So it is necessary to study the 1/2 vehicle model which has
four degrees of freedom.

2.2.1.2 4-DOFs suspension system of 1/2 vehicle

As there are two front tires and two rear tires on each vehicle, there are two types of 1/2
vehicle model: front/rear model - which is also known as Pitch model, and left/right
model – which is also known as Roll model (Gao et al., 2007).
 27

First, the Pitch model is analysed. The force analysis of Pitch model is shown in the
following diagram (Chen et al., 2012):

’ ’

Figure 2-3 4-DOFs suspension system of 1/2 vehicle (pitch model)

Thus the dynamic equation of 1/2 vehicle can be derived as follows:

m1 z1  k11 ( z1  q1 )  k12 ( z1  z1 )  c1 ( z1  z1 )
m z  k ( z  q )  k ( z  z )  c ( z  z )
 2 2 21 2 2 22 2 2 2 2 2
 (2.2)
m z  k ( z  z  )  k ( z  z  )  c ( z  z  
1  c2 ( z 2  z2 )
)
 3 3 12 1 1 22 2 2 1 1

 J   -  k12  z1  z1   c1  z1  z1   l f   k22  z2  z2   c2  z2  z2   lr


In which m1 , m2 are mass of tire, m3 is mass of 1/2 vehicle body, k11 , k21 are

rigidity of tire, k12 , k 22 are rigidity of suspension spring, c1 , c2 are damp of

suspension. z1 , z2 are vertical displacement of tire, z1 , z2 are vertical

displacement of vehicle body, J is moment of inertia of 1/2 vehicle body,  is the

rotary angle of vehicle body at the centre of gravity, q1 , q2 are vertical displacement

of ground, l f , lr are the distances of the suspension locations, with reference to the

centre of gravity of the vehicle body.

The Roll model of 1/2 vehicle is similar with the Pitch model. The force analysis is
shown in the following diagram:
 28

’ ’

Figure 2-4 4-DOFs suspension system of 1/2 vehicle (roll model)

So the dynamics equation is:

m1 z1  k11 ( z1  q1 )  k12 ( z1  z1 )  c1 ( z1  z1 )
m z  k ( z  q )  k ( z  z )  c ( z  z )
 4 4 41 4 4 42 4 4 4 4 4
 (2.3)
   
m2 z2  k12 ( z1  z1 )  k42 ( z4  z4 )  c1 ( z1  z1 )  c4 ( z4  z4 )
 I  -  k12  z1  z1   c1  z1  z1   lL   k42  z4  z4   c4  z4  z4   lR

In which m1 , m4 are mass of tire, m2 is mass of 1/2 vehicle body, k11 , k41 are

rigidity of tire, k12 , k 42 are rigidity of suspension spring, c1 , c4 are damp of

suspension. z1 , z4 are vertical displacement of tire, z1 , z4 are vertical

displacement of vehicle body, I is moment of inertia of 1/2 vehicle body,  is the
rotary angle of vehicle body at the centre of gravity, q1 , q4 are vertical displacement

of ground, lL , l R are the distances of the suspension locations, with reference to the
centre of gravity of the vehicle body.
In the research project, the two front wheels are assumed to traverse through the pothole
simultaneously, and so are the two rear wheels, which doesn’t cause the vehicle to
swing left and right. Thus my study focuses on the analysis of the pitch angle of the
 29

vehicle. So the Pitch model of 1/2 vehicle is used in my study, no swing is considered.
Moreover, in real life, the vehicle may be bounced up from road, which means the
displacement of vehicle in the vertical direction should be considered.
In the contact measurement, when the vehicle passes through the pothole, from the side
direction view, the tyres firstly contact with the falling edge of pothole, and then contact
with the rising edge of pothole, as shown in figure 5-14. This is exactly the same as the
processing of the vehicle passing through the expansion joint of bridge, and can be
analysed using the dynamic response of the 1/2 vehicle model. Therefore, in the
experiment in this thesis, the expansion joints on Dundee Tay Road Bridge are used as
the experimental object.

2.2.1.3 7-DOFs suspension system of the whole vehicle

The force analysis of the whole vehicle is shown in Figure 2-5 (Guan et al., 2010):

Figure 2-5 7-DOFs suspension system of the whole vehicle

The dynamic equation is as follows (Fan et al., 2007):

 30

m1 z1  k11 (q1  z1 )  k12 ( z1  z1 )  c1 ( z1  z1 )

m z  k (q  z )  k ( z  z )  c ( z  z )
 2 2 21 2 2 22 2 2 2 2 2

m3 z3  k31 (q3  z3 )  k32 ( z3  z3 )  c3 ( z3  z3 )


m4 z4  k41 (q4  z4 )  k42 ( z4  z4 )  c4 ( z4  z4 )
mz  k12 ( z1  z1 )  k22 ( z2  z2 )  k32 ( z3  z3 )  k42 ( z4  z4 )

  c1 ( z1  z1 )  c2 ( z2  z2 )  c3 ( z3  z3 )  c4 ( z4  z4 )

 J x  -  k32  z3  z3   c3  z3  z3   k42  z4  z4   c4  z4  z4   c (2.4)

   k12  z1  z1   c1  z1  z1   k22  z2  z2   c2  z2  z2   d

  ( f d 3  f d 4 )c  ( Fr 3  Fr 4 )c （f d 1  f d 2 )d   Fr1  Fr 2  d

 J y  -  k12  z1  z1   c1  z1  z1   k42  z4  z4   c4  z4  z4   a
   k22  z2  z2   c2  z2  z2   k32  z3  z3   c3  z3  z3   b

  ( f d 1  f d 1 )a  ( Fr 3  Fr 4 )a （f d 2  f d 3 )b   Fr 2  Fr 3  b

All the parameters in the equation have the same meaning as in chapter 2.2.1.2.
For this research project, the whole vehicle model is much more complex than 1/2
vehicle model without any improvement. So the 1/2 vehicle Pitch model is chosen for
this study.

2.2.2 Tyre modelling

As the part of a vehicle that has direct contact with roads, tyres not only support vehicle
weight, but also work as a cushion against the impact which is caused when the vehicle
travels on surface of roads, provide sufficient adhesion for driving and braking, and
brings adequate steering and directional stability. Apart from the aerodynamic force
and gravity, near all the other forces and torques are generated by the rolling of tyre on
the surface of ground. Therefore, it is necessary to research the mechanical
characteristics of tyres.
The movement of a vehicle depends on the tyre force, such as longitudinal braking force
and the driving force, lateral force and external tilt force, back to the positive torque
and rolling torque etc., all these forces are functions of slip rate, sideslip angle, camber
angle, vertical load, road friction coefficient and vehicle velocity (Furuichi and Hideo,
1978). Therefore, the establishment of accurate mathematical model of tire mechanics,
in essence, is how to express the functions' relationship.
 31

2.2.2.1 General theoretical model of tyre dynamic

In 1940, Fromn derived the tyre cornering model for the first time by simplifying the
tyre as “beam”. In this way the relationship between steady state sideslip angle and
lateral force is explained by the lateral deformation of the crown of tyre.
Fiala improved the tyre cornering model in 1954. He assumed that the beam layer or
buffer layer was lateral translation deformation and bending deformation with effect of
lateral force. By simplifying the beam layer or buffer layer as a elastic supporting
"beam" affected by lateral concentration force, the relationship among lateral force,
aligning torque, slip angle and camber angle were obtained (Barone, 1977).
Experiments proved that Fiala’s model is more precise with lateral force, while larger
error is with aligning torque.
In 1966 Pacejka studied the static and dynamic mechanism of tyre and introduced the
string model of finite width provided with tread elements, in which the tyre was
considered as many flexible supported stretched parallel strings joined together by
lateral cord (Pacejka, 1979).
Sharp brought out spoke type model in the middle 80s. He considered the tyre as
identical radial spokes which were flexible and were connected with the hub. The
regular deformation of spokes leaded to hysteresis losses. Assume there is no slippage，
the circumferential and lateral deformation of spokes when entering the area touching
with ground can be calculated using kinematics. The force upon the spoke and force
behind the spoke can then be calculated with elastic property of spokes. Finally the total
force and torque can be calculated by summarising all the forces and torques inside the
area touching with ground. The feature of this model is less number of parameters but
less preciseness (Takayama et al., 1983).

2.2.2.2 Semi-empirical model of tyre dynamic

Major error is likely to happen when pure theoretical model of tyre is used to simulate
the dynamic of whole vehicle (Scavuzzo et al., 1993). It is convenient to use
experimental data, however, due to the variety of the road, the limited number of
experiments cannot cover the dynamic of tyre on all roads with various vertical load.
Thus the semi-empirical model is brought out to resolve the problem.
Fiala derived the dimensionless equation from Fromn’s simplified tyre model in 1954.
In 1988 Zeostal et al. describe semi-empirical model in the following form:
 32

C
a1 3  a2 2 
 FZ
F   FZ 3 2
(2.5)
a1  a3  a4  1
In the equation, C stands for the cornering stiffness of tyre, which is the slope of the
curve at the origin of the lateral force. The drawback of this model is the number of
variable is too large.
In 1987, Pacejka et al. established the Magic formula based on large number of
experiments (Bakker et al., 1987). The Magic formula can describe the feature of forces
and torques on tyre under pure slip, pure cornering or pure brake, pure drive conditions
with preciseness and convenience. It is a practical model. The equation is:
Y ( X )  D sin{C arctan[ BX  E ( BX  arctan( BX ))]} (2.6)
In which X represents the slip angle and Y represents side force; or X represents
longitudinal slip and Y represents longitudinal force (in this equation one symbol may
represent various physical quantities). C governs the shape of the curve (shape factor),
D is peak factor, and E is curvature factor.

Generally, the tyre model is the combination of experimental value of variables and
theoretical calculations. Each model has its own application scope. If a model is used
for an impropriate condition, the preciseness of calculation can be greatly reduced. In
this study, the hitting between the tyre and the edge of pothole is too strong, and may
be out of range of most models. Thus the model must be carefully chosen so that it can
provide enough accuracy.
This project only study the uniform linear motion of the vehicle. Therefore, this paper
focuses on the pressure and the supporting force in the vertical direction.

2.3 Signal Processing

The output of sensor is normally a weak analogue signal, which requires amplification,
filtering, A/D convention before being process by computers. The purpose of signal
processing is to filter the redundant content of signal, especially noise and interfere, and
to transform the signal to a form that is easy for analysis.
The study of signal processing started in 60s of 20th century. The simulation of circuit
and filter designing done by Bell Laboratory and MIT laid the foundation of digital
filter. In the middle 60s, the Fast Fourier Transform (FFT) is invented to increase the
speed of spectral decomposition to more than 100 times, which enabled the PC to do
 33

spectral decomposition. Thus it is practical to process signal with digital filtering and
Fast Fourier Transform, which becomes the core of digital signal processing. The term
‘Digital Signal Processing’ began to be used in scientific and technology fields.

2.3.1 Coordinate axis transformation

The coordinate axis transformation is an operation that transfer the signal from one
coordinate system to another, which can be done by general transformations such as
translation, rotation or scaling, or mapping from one system such as Cartesian
coordinates to another, such as polar coordinates (Jiali et al., 2008). The mapping
relation needs to be predetermined, otherwise the mapping cannot be processed.
However, in this project, the relation between the local coordinate system of the sensor
and the global coordinate system may change at any time. Because there are relative
movement (and relative rotation) among the ground, the vehicle and detection sensor,
three coordinate systems are required. The movement and rotation is represented as
transformations from one coordinate to another. Therefore the mapping relation cannot
be predetermined. A statistics method is used to determine the mapping relation
between sensor coordinate system and global coordinate system.

2.3.2 De-noise
The digital signal filter is a discrete time system that processing digital signals in order
to obtain expecting response characteristic. It process the digital signal transferred from
analog signal.
Theoretically the digital signal filter can implement any filtering that can be represented
by mathematical algorithm. The two main limitation of digital signal filer is its
processing speed and cost. The digital signal filter cannot run faster than its internal
circuit. But as the cost of IC continue to decrease, the digital signal filter becomes more
and more common and an important part of everyday life as radios, mobile phones and
stereo.
The digital signal filter is assembled by basic digital circuits such as register, time,
adder and multiplier. With the development of integrated circuits, its performance
continues to increase and cost continues to decrease. Therefore the application of the
digital signal filter is wilder. According to the characteristic of digital filter, it can be
classified to linear and non-linear, causal and non-causal, infinite impulse response (IIR)
and finite impulse response (FIR) and so on. Among all the digital signal filters, the
linear time invariant digital filter is the most basic type.
 34

Because the digital filter can make use of time delay unit, so it can introduce a degree
of non-causality and become more flexible and powerful than traditional digital filters.
Comparing with IIR filer, FIR filter has the advantage of easy implementation and
system stability, it has been widely used (Huan, 2005).

2.3.2.1 Kalman filter

Kalman filter is an effective recursive filter (autoregression filter). It can estimate the
dynamic of a system from incomplete and noisy measurements.
An example of Kalman filter is to predict the position and velocity of the object from a
set of limited observations (possibly biased) of the object including noise. It can be
found in many engineering applications, such as radar and computer vision. Meanwhile,
Kalman filter is also an important topic in control theory and control engineering.
For the radar, people are interested in its ability to track targets. However, the
measurements of position, velocity and acceleration may contain noise at any time.
Kalman filter can obtain a good estimation of position of target by de-noise with
dynamic of target. This estimation can be the estimation of current quantities (filtering),
of future quantities (forcasting), or of past quantities (interpolation or smooth).
This filtering method is named after its inventor Rudolph E. Kalman. But actually Peter
Swerling proposed a similar algorithm earlier.
Kalman filter is firstly implemented by Stanley Schmidt. Kalman found out that his
filter is useful to predict the orbit in Project Apollo when he visited NASA’s Ames
Research Centre. Later, the guidance computer on Apollo used Kalman filter. Papers
about Kalman filter is written by Swerling (1958), Kalman (1960) and Kalman and
Bucy (1961).
Kalman filter now has many different implementations. The form proposed by Kalman
is called simple Kalman filer. Apart from it, there are variants such as Schmidt’s
extended filter, information filter and the square root filter which is developed by
Bierman Thornton. The most common application of Kalman filter is phase-locked loop,
which exists in radars, computers and any video or communication equipment (Gao et
al., 2004).

2.3.2.2 FFT filter

Fast Fourier Transform (FFT) is the fast algorithm of discrete Fourier transform. The
FFT can also be used to compute the inverse transform of the discrete Fourier transform.
 35

FFT has a wide range of applications, such as digital signal processing, computing the
multiplication of large integers, solving partial differential equations, and so on.
The most common FFT algorithm is Cooley-Tukey algorithm. Using divide and
conquer (D&C), this approach recursively breaks the discrete Fourier transform of
length N=N1N2 down to a sequence of smaller discrete Fourier transforms, in which
the number of smaller transforms is N2 and length of each smaller transform is N1.
This method is well known after published in an algorithm for the machine calculation
of complex Fourier series (J. W. Cooley and J. W. Tukey, 1965). However, later it is
found out that these two authors just re-invented the algorithm brought out by Carl
Friedrich Gauss in 1805. Actually this algorithm was brought out for several times with
various forms in history.
The most common application of Cooley-Tukey algorithm is to break down a DFT with
length of N to two DFT with length of N/2. This application is usable only for DFT
with length of power of 2, thus it is called radix-2 decimation-in-time (DIT) FFT. As
indicated by Causs, Cooley and Tukey, Cooley-Tukey algorithm can be used for not
only DFT of any length (mixed-radix FFT), but also other variants such as the split-
radix FFT. Although the Cooley-Tukey algorithm is a recursive method, in many
traditional implementations it is rewritten in a non-recursive form. Besides, because
Cooley-Tukey algorithm breaks a DFT to several smaller DFTs, it can be combined
with any other DFT algorithm.

2.3.2.3 Wavelet
Wavelet analysis is also called wavelet transform (WT). The wavelet analysis is a fast
developing branch of applied mathematics based on the foundation work done by Y.
Meyer, S. Mallat and I. Daubechies. The appearance of wavelet analysis is a milestone
in history of signal analysis development, and is being wildly used in multiple field,
including molecular dynamics, ab initio calculations, astrophysics, density-matrix
localization, geophysics, optics, turbulence and quantum mechanics. DWT (Discrete
Wavelet Transform) is also used in many other area such as image processing, blood
pressure, heart rate and ECG analysis, DNA analysis, protein analysis, meteorology,
General signal processing, speech recognition, computer graphics and multifractal
analysis. Theoretically wavelet analysis can be used instead of Fourier Transform. In
this section, the history and development of wavelet analysis is discussed.
The Fourier analysis is one of the most widely used mathematical analysis methods
 36

since the analysis theory of heat conduction was published by Fourier in 1822. The
Fourier analysis is a spectral analysis, which can reveal the spectrum of a signal. But
the Fourier analysis has its own disadvantage. The Fourier coefficient is the weighted
average of signal f ( x ) on the whole time-domain, with which is impossible to
represent the local feature of f ( x ) on time-domain. However the local feature of
f ( x ) on time-domain is important in both theoretical and practical application.
The idea of scaling and translation was firstly used to compose wavelet orthogonal basis
by Alfred Haar, who gave out the construction of the Harr wavelet in 1910. The 70’s is
important for the development of the wavelet analysis, when the birth of wavelet
analysis was prepared by the establishment of Calderon-type reproducing formula, the
atomic decomposition for Hardy spaces and study of unconditional basis. In 1982, a
basis is firstly constructed by J. O. Stormberg, which is similar with modern wavelet
basis.
The first real wavelet basis was constructed by Y. Meyer in 1986. After that Lemarie
and Battle independently constructed wavelet function with exponential decay. S.
Mallat introduced the concept of multi-resolution analysis (MRA), with which the
previously proposed wavelet functions were uniformed. In 1988 orthogonal wavelet
basis with limited subset was constructed by Daubechies. The single orthogonal
wavelet was constructed based on spline function by Jintai Cui and Jianzhong Wang in
1990. They also discussed the best localization properties of scaling function and
wavelet function. In 1994, multi-wavelet theory was founded by Goodman based on
MRA. He also gave out the sample of construction of multi-resolution wavelet. In
recent years, the high-resolution wavelet theory has been attracting the attention of
researchers, with many of its topic being studied (Hesami et al., 2009, Yang et al., 2007).
In my study, the wavelet analysis is used in two aspects:
1 – To separate the data to the summation of a group of signals, in which I may find the
key feature signal;
2 – To filter the noise and enhance the useful signal.

2.4 Conclusion
Based on discussions in this chapter, the contact measurement is chosen as my pothole
detection method, which will be analysed in detail in Chapter 3 and carried out in
Chapter 4. The 4-DOFs suspension system of 1/2 vehicle is chosen as the suspension
model in my study, and energy conservation will be used in tyre model. Both
 37

suspension model and tyre model will be discussed in Chapter 5. For the collected data,
the principal component analysis (PCA) is chosen to process axis-correction, and
wavelet analysis is chosen to process de-noised. The data process is discussed in
Chapter 6.
 38

Chapter 3. Detection system design

Based on the discussion of the Chapter 2, the contact measurement method is chosen to
be used in my study.
In this chapter, considering the disadvantage of traditional contact measurement
methods, a new contact measurement method is proposed. Firstly the requirement
feature of each module is analysed in Section 3.1, and then the detection method -
especially how it overcomes the disadvantage of traditional contact measurement
methods – is brought out in Section 3.2. In Section 3.3, the hardware requirement is
discussed based on Section 3.1 and 3.2. At last the data collection is designed to be
carried out in Chapter 4.

3.1 Detection method requirement analysis

In my design, the measurement device should be deployed and applicable in large scale,
so that it can be used by different owners and therefore can be used to detect the same
road repeatedly. In this way the disadvantage that can only cover a small part of road is
resolved. Secondly, to improve the efficiency of the whole detecting system, the device
should be able to easily store and send data to a server. To reduce the workload of the
server, this front-end device should have capacity of data pre-processing. At last, it
should be universally, so that it is stable and can be used on different types of road.

3.2 Detection method

The objective of this project is to design or search for a device that is fixed on the
vehicle, and the device is mounted as many vehicles as possible.
When the vehicle traverses through a pothole, the vibrations caused by the pothole will
spread to the vehicle body through wheels, dampers and suspension and lead to the
vibration of the vehicle body. When a device with built-in acceleration sensors is fixed
on the vehicle, (The ‘fixed’ here means to install device firmly on equipment such as
mobile holder, so that the vibration of device is synchronised with vehicle.) the
vibration of the device can be considered as a synchronization vibration. Thus when
vibration of device is detected through its acceleration sensors or other sensors, the
vibration of the vehicle can be obtained, which means the pothole is detected.
Combining with the GPS built-in the device, the location of the vehicle can be
determined, which means that the location of the pothole is determined.
The device may store the data from the sensors in its internal storage or send the data
to the system data server via wireless network. The system reads and processes the data
 39

from the device. Combining with the GPS built-in the device, the location of the vehicle
can be determined, which means that the location of the pothole is determined. The
process can be fulfilled with computing capability of the device or other calculating
devices (e.g. a work station connected with the data server). Although the real-time
calculation by device can reduce data flow and calculation workload on the work station,
the power dissipation and power consumption of the device will increase, which
reduces its stand-by time. Thus the work station is the better device with respect to the
measurement of pothole.
When a lot of vehicles with devices run on the road to collect data and send the collected
data to the central terminal, the operation efficiency of the whole system will be greatly
increased because each vehicle can be considered as a detection vehicle, which means
that there are lots of inspecting vehicles involved. With plenty of vehicles installed the
detection device, at any time there are some of the vehicles testing and sending data on
the road. As long as there are such kind of vehicles running on the road, the data of that
road is always being collected. The efficiency of running many vehicles (with detection
device) on the road is much higher than single detection vehicle.
The advantage of running many vehicles with detection device is not only higher
efficiency. Due to the noise of the road and vehicle body, the contact measurement
method can only detect the part of road that has contact with tyre of vehicle, every
single detection device may bring errors in measurements. However by revising the
information of multiple detection vehicles, more precise information can be obtained.
Thus it can improve the accuracy of detection.
The collection devices will be installed in the vehicle and as predicted, it requires many
devices to meet the needs in the future. And the specific demands are as follows:
1. Intelligent system: OS shall be of sufficient intelligence and openness; there is mature
development tool to develop and debug the application software needed conveniently;
2. Well-established hardware: the hardware shall include CPU, mainboard, storage,
battery, communication interface and sensor, constituting a complete embedded system;
3. Low cost: it has large quantity demand and it is needed to control the cost within
acceptable and reasonable range.
The detection device can detect the displacement of the vehicle vertical direction, in
order to infer pothole. That is to say the sensor in the detection device should includes
a module to digitise and record distance data. For example, distance data can be
converted to voltage or current, then the voltage or current can be converted to digital
 40

signal by A/D converter. In this way the distance is digitised. This job should be applied
to acceleration or gyroscope type sensor. The detection device collects data from front-
end sensors, then pre-processes and packages as local data, then waits to send to server
to process or sends real-time data to server if there is wireless fidelity (Wi-Fi)
connection. The server collects data from all the detection devices, analyses and
presents result to users.

3.3 Demand analysis of each module

In the Section 3.3, development, feature and requirements of each module of detection
device is discussed.

3.3.1 CPU
The work of CPU is to assign work load to modules of detection device, and do
calculation for data pre-processing. Hence the CPU should satisfy following
requirements:
1 the processing speed of CPU should be high enough to pre-process the collected data
when the vehicle is running with a reasonable speed;
2 the power consumption of CPU should be in proper range, so that the detection device
can keep working whenever the vehicle is on the road;
3 the cost of CPU should be low enough so that CPU can be large-scale deployed.

3.3.2 Sensor
By using flexural electrodes and circuit, the acceleration sensor can measure
acceleration and convert linear acceleration signals to electrical signals as output (Ravi
et al., 2005). For example, some laptops have built-in acceleration sensors to monitor
whether the notebook is dropped. And if that happens, the laptop automatically turns
off hard disks to prevent serious damage.
Accelerometers measure the information of the transformation motion in three
directions, which is its linear acceleration (Xiaoya et al., 2010); Gyroscopes, however,
measure the information of its rotating movement, which is its angular motion.
Gyroscopes also able to measure the flatness of the road. When vehicles travels on the
road and encounters a pothole, the tilt between four tires will cause the rotation of the
vehicle body, which will have a response on the gyroscope. In this way the gyroscope
can be used to measure the flatness of the road information.
 41

Gyroscopes are not sensitive on vehicle noise (for example, noise from the engine).
Engine noise may cause the vehicle body to shake, and the acceleration sensor is unable
to distinguish the difference between the vibrations caused by the roads and the
vibration caused by engine noise. Thus the rate of miscarriage by acceleration sensor is
increased. But the engine noise has none or a minimal impact on the rotation motion of
vehicle body, so that the gyroscope may have more accuracy measurements of the
flatness of the road.

Figure 3-1 Typical gyroscope

 42

Figure 3-2 Microstructure of gyroscope chips based on MEMS (Micro-electromechanical

Systems, is a mini system that integrates optical systems, drive components, mechanical
components and electronic control system in a single unit) technology

With lower cost, higher sensitivity, higher accuracy volume miniaturization but lower
power consumption, the accelerometer and gyroscope is more and more popular, and
applications continue to appear. For instance, a new application: Apple will use
microphone and accelerometer to reduce vibration noise of iPhone (United States patent
publication number US20120286943 A1, http://appft1.uspto.gov/netacgi/nph-
Parser?Sect1=PTO2&Sect2=HITOFF&u=%2Fnetahtml%2FPTO%2Fsearch-
adv.html&r=17&p=1&f=G&l=50&d=PG01&S1=(apple.AS.+AND+20121115.PD.)&
OS=an/apple+and+pd/11/15/2012&RS=(AN/apple+AND+PD/20121115)).

3.3.3 Memory
Memory stores software and data for the detection device. First of all, enough memory
supports the detection device to work normally; secondly, the access speed of memory
should be fast enough, so that the overall speed of the system will not be slowed down;
third, the power consumption of the memory should not too high, otherwise it will
 43

shorten the working time of the detection device as a battery-powered system; finally,
same as CPU, the memory should have a low enough cost to be large-scale deployed.

3.3.4 Network
The network of detection device refers to the communication module that connects to
wireless network. Strictly speaking, the network itself is not a part of the data
processing system. However in Section 3.3.4 the wireless wide area network
development and situation are discussed, because one of the requirements of the
detection device is that it can real-time sends collected data to server through wireless
network (Wi-Fi), it is a necessary part of the detection device.
Table 3-1 Technology and the highest speed of various types of network

Type Major Tech Max download speed Max upload speed

1G analog FDMA / /
2G digital TDMA (GSM) GSM: 14.4Kbit/s GSM: 14.4 Kbit/s
GPRS: 53.6 Kbit/s GPRS: 26.8 Kbit/s
EDGE: 217.6 Kbit/s EDGE: 108.8
Kbit/s
3G digital UMTS UMTS: 384 Kbit/s UMTS: 128Kbit/s
(WCDMA/TD- HSPA: 7.2 Mbit/s HSPA: 3.6Mbit/s
SCDMA/CDM HSPA+: 14.4 - HSPA+: 5.76 –
A2000) 168.8Mbit/s 23.0Mbit/s
4G digital WiMAX/LTE LTE: 100Mbit/s LTE: 50Mbit/s
LTE-A: 1Gbit/s LTE-A: 500Mbit/s
Network is becoming more and more important for people's daily life. According to a
survey made by O2 and Samsung together shows that nowadays people spend about 2
hours in average to use mobile phones every day. Half of this time is spent to access
the Internet, play music and play games, and only 12 minutes are used to make phone
calls. (http://www.telegraph.co.uk/technology/mobile-phones/9365085/Smartphones-
hardly-used-for-calls.html)
In UK, 4G license is far cheaper than 3G license (as shown in Table 3-1). In 2000 the
United Kingdom finance obtained 22.5 billion pounds from the spectrum auction, while
the price in 2012 is only 1.3 billion pounds. This suggests that the 4G technique costs
less to get start with, which makes it easier to spread. This means that in the future, the
spreading speed of 4G is probably faster than 3G.
 44

Therefore a conclusion can be reached: High speed, stable, affordable wireless network
has been laying the communication transmission basis for distributed data collection.

3.3.5 Battery
Generally the power consumption of built-in gyroscope is higher than the accelerometer.
The typical current consumption of accelerometer is 0.25mA for k3dh (An
accelerometer sensor IC comes from STMicroelectronics corporation, used in many
smart mobilephones such as Samsung I9100), and the typical current consumption of
gyroscope is 6.1mA for k3g ( A gyroscope sensor IC comes from STMicroelectronics
corporation). However, they are almost negligible compared to the whole machine
power consumption. According to the energy consumption of the whole system, the
energy consumer includes CPU, sensor, memory, network modules, as well as other
captive hardware. The battery should have high energy density, can be repeatedly
charge and discharge, and can keep safety under impact.

3.3.6 UI
Some embedded systems have user interface (UI) and some do not. Whether an
embedded system has UI depends on the requirements of the system. If the UI is
necessary, it should provide clear instruction for user, so that even those who have never
used the system can easily learn to use it. The UI should also be hard to crash.

3.3.7 Size
Internal space of vehicle is precious. A large number of vehicles have a long design
period in order to have a large internal space. Therefore, as an on-board, the detection
device should have a minimal size and weight. The universal adaptability of the
detection device should also be taken into account. It may take a wide range of research
and experiments to get good results.

3.3.8 Conclusion
In summary, the requirements of the detection device is: it should be an embedded
system, which is made up of CPU, sensors, memory, network module, battery etc. On
this embedded system, software can be run to complete assigned work.
 45

3.4 Classification and evaluation of embedded systems

3.4.1 Sun SPOT

At initial stage, the project planned to use Sun SPOT, an embedded system with various
kinds of sensors mounted inside (three-axis acceleration sensor, temperature sensor and
light-sensation sensor) which was promoted by Sun Company in 2008. It includes ARM
920T CPU with 180MHz, 512KB RAM and 4MB Flash on-board memory, USB
interface and 2.4GHz IEEE 802.15.4 wireless interface for communication as well as
750mAh rechargeable built-in battery.
Sun SPOT is developed by Sun Company and it certainly adopts Java as development
environment, which guarantees the high efficiency of development and easy operation.
In addition, Sun Company provides a large quantity of technical support.
However this embedded system has disadvantages as well. The price of Sun SPOT
reaches $400 per set. Due to such high cost, it is only used for prototype verification
and difficult for wide-scale promotion. Besides, it has poor universality. 180MHz CPU
(upgraded to 400MHz then), with low performance, cannot conduct data processing
and calculation with high requirement. It is not provided with long-distance wireless
data transmission module (such as GPRS or 3G module). And with 4MB Flash memory,
it cannot conduct continuous data collection for a long term with high sampling
frequency (such as 100Hz), which restricts its application to most extent.

3.4.2 Raspberry Pi
During this experiment, Raspberry Pi is newly promoted with support of Raspberry Pi
Foundation in 2012., With the dimension (67.6*30 mm) larger than Sun SPOT
(includes BCM2836 CPU with 700MHz (ARM architecture as well), 256MB or
512MB memory, SD card for extension and I/O such as Ethernet, USB, HDMI, RCA.
For OS, adopt Linux developed by Open Source (other OS is available for operation as
well)), the Raspberry Pi has more mature and richer development tools and supporting
software than those of Sun SPOT. What’s more, the hardware configuration is higher
than that of Sun SPOT (the hardware has been improved much only with four years)
and it can even run programs designed for PC with high performance graphic
processing unit. In addition, its price is only $25/35 (Raspberry Pi has two versions, the
difference is that the latter one has more USB and Ethernet interface).
 46

The mainboard of Raspberry Pi itself does not include the sensor, battery or wireless
communication module. In fact, it is much more similar to the host of a PC. And its
standard USB interface and Linux OS contributes to easy extension.
However, if Raspberry Pi is applied in the project, it is needed to add the sensor, battery
and long-distance wireless data transmission module (such as GPRS or 3G module),
which will produce some expense. Even so, Raspberry Pi still has the cost performance
much higher than Sun SPOT developed four years ago.

3.4.3 NUC
What’s more, the companies such as Intel have been promoting mini computers, namely
NUC (next unit of computing) in recent years. Strictly speaking, NUC is not an
embedded system. It has good versatility, and it can be used instead of embedded
system under most conditions. Therefore it is discussed together with other embedded
systems in this thesis. NUC is a host with dimension only 10*10*5cm internally
mounted with ultra-low voltage CPU ranging from Celeron to i5, mainboard, memory,
2.5 inch hard disk (notepad mechanical or SSD hard disk) or mSATA interface SSD
hard disk, integrated display card and other general devices. And it provides such
common interfaces as mini PCI-E, USB, HDMI, Wi-Fi and Bluetooth as well. As it is
a platform of x86 and operates windows operating systems (other operating systems are
available as well), it has advantages such as universality to operate most kinds of Win-
Tel software, small dimension and low power consumption (to tens of watts). However,
it is not popularized currently mainly because of the high price: the price of CPU and
mainboard is 100-300 pound; in addition the internal storage and hard disk shall be
purchased; if this system is used for this project, modules such as sensor and 3G are
needed to be purchased as well.

3.4.4 Mobile phone

There are a number of manufacturers and a variety of types of mobile phone. To
understand the feature of mobile phone, one has to know about their history and
development.

3.4.4.1 The development history of mobile phone

3.4.4.1.1 The development of mobile phone network: from 1G to 4G

In 1999, Internet was rapidly popularizing around the whole world, so more and more
users generated the demand for mobile network. Under such circumstances, the
 47

shortcomings of 2G network (including GPRS (so-called 2.5G) and EDGE (so-called

2.75G)), such as slow network speed, unstable connection and lost connection when
change service zone, are exposed. Therefore, it is a milestone of the development of
mobile communication network that 3G network such as CDMA2000, WCDMA and
TD-SCDMA are approved in the same year.
3G network witnesses an extremely rapid development. Under differences of
application environment, the actual access speed of 3G network varies from 300k to
2Mbps. The more advanced technology, HSDPA, or so-called 3.5G network, have
theoretical download speed of 14Mbps and upload speed of 5.8Mbps, which is a huge
improvement comparing with the speed of 2G network (dozens kbps to over 100 kbps).
As a result the market and users have higher requirements to the performance of the
mobile phone itself.

3.4.4.1.2 Development of hardware of smart-phone

The hardware has experienced a rapid development along with OS and software.
1. CPU: In early days the mobile phones are mainly used for voice communication and
do not have much additional functions, therefore the design of mobile phones does not
focus on the computing speed but how to reduce the power dissipation as much as
possible so as to extend the stand-by time, which requires the assistance of better battery
processing technologies and power management program. However, under the same
technological level and the same architecture, the higher processing speed means higher
frequency of system clock, which means that the increase of power dissipation is
inevitable. At that time, a reasonable reduce of the processing speed of CPU is
acceptable in order to reduce the power dissipation and extend stand-by time.
However, with the arrival of 3G era, the users need a stronger CPU to process more
powerful programs, which immensely promote and propel the development of CPU.
Based on the research experience on desktop CPU, the CPU clock of smart-phones
rapidly increases to GHz from dozens of MHz, with the number of core increases from
one to two, four and even eight. Meanwhile, the GPU responsible for graphics
acceleration has been developed. One thing should be noted that so far the CPU for the
majority of the smart-phones is the SoC (System on Chip) architecture based on ARM.
Compared to the desktop CPU (Typical Intel x86 series), smart-phone CPU requires
lower power dissipation, lower cost and is more compact, which is more suitable for
the embedded system.
 48

Although the CPU of embedded system varies from the architecture of x86, we still can
roughly compare their performances. In consideration of the development area of
desktop CPU in recent years, Moore’s Law (the observation that : when the price does
not change, the number of transistors in an integrated circuit doubles approximately
every 18 months as well as the performance. In other words, the computer performance
that can be bought with each dollar doubles every 18 months. The Law reveals the
development speed of the information technology) is not as sure as it was in the past,
and it shows the trend of declining, while the development of embedded CPU is still
rapid.
2. User interface (UI): The earliest mobile phones had small screens that can only
display monochrome. Those mobile phones before Moto GC87C in 1998 even adopted
LED digital tube for display. This is partly limited by the manufacturing technologies,
techniques and cost, partly because it is hard to process and display rich colours for the
low speed CPU. In fact, the basic function of making calls has no demand for a strong
CPU. However, since the beginning of the 3G era, smart-phones have become the
personal digital assistant, on which a large number of software requires advanced
hardware to support advanced requirements such as high resolution display with large
number of colours. On one hand, the UI becomes more and more easy to use with
introduction of touch screen and gesture-control; on the other hand, high-quality UI
consumes more resources, which promotes the development of hardware. With
advanced UI, the smart-phones are easy to use not only by the professional users, but
also by the senior, children and less-educated people.
3. Memory: the memory capacity increases to gigabytes from hundreds kilobytes or
several megabytes in the early stage. The access speed increases and power dissipation
reduced along while the memory capacity increases.
4. Input/output device (I/O): has upgraded from infrared way and series ports to Wi-Fi
and USB ports, with the speed increased greatly.
5. Sensors: In early days, microphone is the only sensor on the mobile phone, which is
used to receive voice signals. Later, more sensors are added to the mobile phone to
support more functionality. The acceleration sensors are used to detect the change of
gravity to verify the change of mobile phones’ attitude so as to automatically adjust
display. Since then the developers have found that the acceleration sensors can be used
for more purposes. A typical example is using acceleration sensor to control the data
transmission between two mobile phones. For example, in the famous social contact
 49

software Bump, simply by a slight touch of two mobile phones the two owners can
mutually link as connections, which is a typical model of using acceleration sensors
and GPS Location Based Service. Another example is that, it can be used to control
mobile phone games through tilting the devices rightwards or leftwards, which is
particularly helpful in racing games. However the acceleration sensors cannot
completely determine the gesture of the devices because it cannot measure the angular
acceleration. Thus gyroscopes are equipped to some smart-phones to provide more
precise output in the combination of the acceleration sensors, and used in more
applications.
The development trend of sensors is diversification and high technology. Currently the
mainstream smart-phones adopt various sensors such as acceleration sensors and light
sensors which can detect the change of illumination intensity and adjust the brightness
of screen so as to improve the display performance. Some high-end smart-phones even
adopt gyroscopes, temperature meters, barometers which is used to measure altitude,
magnetometers, and distance sensors which can automatically shut down the touch
screen to avoid mistake-touch when the user makes calls with his face near the touch
screen .The trend of research reflects in miniaturization, low power dissipation and high
precision. Originally, the built-in acceleration sensors of mobile phones are used for
recognizing the gravity direction so as to automatically adjust the screen display
direction with the rotation of the mobile phones when the users rotate their phones,
which provides convenience for the users. When this function is adopted in the software,
the situation is full of variety. A lot of functions can be realized when the acceleration
sensors are used to detect device vibration (Xu et al., 2012). A typical example is to use
acceleration sensors in many mobile games as one of the input units. Thus the users can
control the roles (such as racing bicycles) in the games only by shaking the mobile
phone rightwards or leftwards to adjust the mobile phone’s relative inclination angle to
the ground, which brings more enjoyment of the games. In fact it happens to be one of
the most distinctive functions of Nintendo’s historical game device, Wii. Besides in
games, shaking mobile phone also can be used as one of the input orders for users.
Since then adding connections by shaking mobile phones at the same time has appeared
in many social contact software. In the past, when two smartphone users stand next to
each other, if they hope to add each other to contact list, instead of manually type
contact information into phone, which takes time and may contain misspelling words,
they can open blue tooth and find each other, then click ‘Connect’ and input password,
 50

then the match process is completed. Now with built-in sensors, Location Based Service
of GPS and Internet data connection, this process is highly simplified: the users just
need to stand together, open a social app and then shake smartphone at the same time.
In addition, the characteristics of acceleration sensors have been used in other creative
functions. For example, if the sensor detects that the mobile phone is picked up and the
attitude changed from horizontal to vertical, the incoming call will be answered
automatically; and if the sensor detects that the mobile phone is turned to face down,
the incoming call will be refused automatically.
 51

Table 3-2 Power Dissipation and Precisions of Acceleration Sensors and Gyroscopes: Samsung
Galaxy Note 1 (N-7000) and iPhone 5

Samsung Note N7000 Apple iPhone 5

Accelerometer Gyroscope Accelerometer Gyroscope
Model ST LIS3DH ST L3G4200D ST LIS331DLH ST L3G4200DH
Measurem ±2g/±4g/±8g/± ±250/±500/±200 ±2g/±4g/±8g ±250/±500/±200
ent range 16g 0 dps 0 dps
Frequency 1Hz - 5kHz 100 - 800Hz 0.5Hz - 1kHz 100 - 800Hz
range
Precision 16 bit 16 bit 16 bit 16 bit
Minimum 1mA 1.5mA 10mA 1.5mA
operating
current
size 3x3x1 mm 4x4x1 mm 3x3x1 mm 4x4x1 mm
Accelerometers and gyroscopes are more and more widely used on mobile. The first
built-in accelerometer is introduced to iPhone to provide gravity sensor, so that phone
can accommodate to different gestures of the users. After that there are more and more
applications of accelerometers on mobile, resulting the development of a lot of software.
Iphone4 is the first mobile that introduces the gyroscope, which enables mobiles to
sense their motion. Since then, an increasing number of mobile phones, including the
iPhone series, Samsung's high-end Android phones, and Nokia lumia920 (Windows
Mobile series) also have built-in gyroscope. It can be predicted that there will be more
mobile phones in the future with built-in gyroscope. The built-in gyroscope is one of
the development directions of mobile-built-in sensors in the future.
6. GPS: Strictly speaking, GPS is not a traditional-type sensor, but for the system that
calls the GPS, if the GPS is considered as a black box and how it works is not involved,
the GPS has no difference with the other sensors such as acceleration sensors and
gyroscopes. It can output the information parameters such as three-dimensional
coordinate of the current location and the current movement speed, as other sensors
output information about acceleration, altitude, temperature etc.
However, the GPS receiver installed inside the smart-phones is only a minor part of the
huge system of GPS. Apart from the receiver, the whole GPS system includes 24 GPS
satellites (21 main satellites and 3 standby satellites) and maintenance stations on
 52

ground. Whether the GPS receiver can output correct data totally depends on whether
the GPS satellite sends correct signals. GPS signals include carrier wave, ranging code
and navigation message, based on which the GPS receiver can work out the current
location and movement speed. From this point, the actual working mode of the built-in
GPS is different from other sensors.
Before the emergence of smart-phones, due to the limitation of algorithm and hardware
technologies the main issues for GPS receiver are the large power dissipation, slow
computing speed (cold start time can be longer than 1 minute, and the satellite-search
speed is slow), high cost and poor precision. The precision of civil communication
signals is about 100 meter, which is because of the interference to civil communication
signals implemented by American government before the year 2000. A lot of
manufacturers produce independent GPS receivers, which works as accessories of
smart-phones or PDAs using Bluetooth as communication method. However, the
Bluetooth itself consumes so much power that deteriorate the working time of GPS
receivers. With the creation of more efficient algorithm and the improvement of
processing techniques, the production cost of GPS receiver drops, which enables the
GPS to become one of the standard configurations of smart-phones. The precision of
GPS receiver is about 10 meter and can totally meet the demand of vehicle navigation.
Quite a number of users need smart-phones as their SATNAV (satellite navigation)
when they drive on the highways. Therefore for these users a new function to use GPS
does not bring more burdens to endurance of the battery (of the smart-phone).
The precision of speed detection of GPS is high. Generally, the precision of GPS to
measure geo-location is within 10 meters. If two measurements are taken at regular
interval to calculate the speed according to the distance between two points, the error
will be rather large. But in fact, the speed detection method of GPS is to detect the
Doppler frequency shift of signals, with which the error is generally around 0.5km/h.
With such a high precision the requirements for speed detection of vehicles can totally
be met.
The combination of Network, PDA and sensors will broaden the applications of smart-
phones. It is how to find out new application fields, fully develop the functions of smart-
phones and finish the jobs which cannot be realized or need more cost to be realized in
the past that will be the research focus in the future.
In conclusion, with the rapid development of technologies in recent years, all conditions
necessary for this research project are complete and mature.
 53

Table 3-3 Milestone Events for the Development of Smart-phones

Year Event Company/Product

1999 The first smartphone Motorola A6188
2000 The first 3G smartphone Panasonic P2101V
2005 The first smartphone with built- Mio DigiWalker A700
in GPS
2007 The first smartphone with built- Sony ERSSION W580C
in acceleration sensor
2008 The first iPhone Apple iPhone 1
2008 The first Android smartphone HTC G1
2010 The first 4G smartphone HTC EVO
Almost all smartphones can be perfectly used as a SATNAV today. More and more
navigation software appear, providing many functions that traditional SATNAV cannot,
such as display of real-time traffic jam and road condition update.

3.4.4.1.3 Development of Software on Smart-phone

Smart-phones should have a relatively complete OS. In the early stage of the
development of PDA, the design of OS stems partly from desktop OS such as Linux
and WinCE. This kind of OS has relatively mature architecture. Meanwhile, it makes it
easier to transfer the existing desktop software to PDA’s OS and at a lower cost. In
2008 the App Store greatly aroused the developers’ interest in iOS software
programming. In 2009 the emergence of Android reduced the cost of embedding OS in
smart-phones for the manufacturers; as a result an increasing number of users can afford
their smart-phones
A mature and excellent OS can not only facilitate its users but also give the developers
much more facility and activity to develop the corresponding application software in
an easy way eagerly. Facility means that an excellent OS should have software
developing tools which are mature, powerful and easy to use for developers, such as
xcode, jdk/eclipse, Visual Studio; and activity means that the Network & Store/Market-
based release mode has changed the software release mode in the past, greatly boosting
the developer’s positivity of software development.
 54

Table 3-4 Comparison of development platform of three mainstream smartphones

Operation System Programming Latest Version Developer’s OS

Platform
iOS xcode 6.0 OS X
Android jdk/eclipse 4.4 multi-platform
Windows Phone Visual Studio 2013(12.0) Windows

Table 3-5 Number of software in App Store of three mainstream smartphones

OS Name Quality of software

Android App Store 1.2 million
iOS Google Play 1.2 million
Windows Phone Store 0.3 million
IPhone app store appears and enables any developer to use the development tools
provided by Apple to develop and upload his own application to the app store. At the
same time, it provides thousands of applications for users to download, no matter where
they are or what time it is. This model greatly stimulated the enthusiasm of developers.
Up to now, there are more than 1 million applications in the app store. Since then, the
similar mode appears on Android and Windows phone platform as well, enables users
to find, download and use the software more conveniently.
As shown in the data issued by IDC, the total sales volume of mobile phones globally
reached 1,821,800,000 in 2013 including 1,004,200,000 smart phones, increasing by
38.4% compared to the sales volume in 2012. It shows that the smart phones have been
replacing traditional functional phones gradually, becoming the useful tools for
consumers. Besides, apart from the basic functions of making and receiving phone calls
and messages, the phones have three additional functions as powerful processing ability,
multi physical-quantity measuring sensors and Internet wireless access whenever and
wherever possible, which makes smart phones as the pocket information processing
devices unprecedentedly and revolutionarily. Depending on these, an increasing
number of innovative functional applications have been designed and realized. And
some are applied, increasing work efficiency greatly (such as Bump) and some even
make the functions unrealizable in the past possible (such as Nokia Here). Under such
condition as perfect and mature hardware, it is how to develop new better functions
with more innovation for them that becomes the most urgent problem today.
55
 56

Table 3-6 Typical software on smartphones making use of sensors/GPS/camera

Name of software Description Modules used

Bump Users only have to collide the Acceleration
mobile phones with each other to sensor, GPS
exchange contact information,
photo and other goals
VR software and virtual reality, enhanced reality acceleration sensor,
games and gaming compass,
gyroscope, camera
Nokia HERE Digital map that can provide acceleration sensor,
information such as nearby gyroscope, GPS,
shops, point of interests and gas compass，camera
station.
Pothole detection Detecting potholes in a certain acceleration sensor,
software（pothole city GPS
Hunter/Agent/De
ctector/Radar/Stre
et Bump）
MOVES Keep record of how many acceleration sensor,
kilometres the user walks each GPS
day, with how many steps, which
transport is used, the position the
user stops and so on.
‘Getting taxi’ Passengers can release needs of GPS
software taxi on smartphone and get in
touch directly with taxi drivers.
The efficiency of taxi is
improved.

3.4.4.2 Features of mobile phone

The first feature of smartphone is strong function of its hardware. Due to the variety of
user needs, various hardware are designed to be integrated together, including high-
 57

speed CPU and numerous sensors. The former provides high processing speed, and the
latter enables the smartphone to collect data in a variety of ways.
The second feature of smart phone is customised software. All smartphones have a fully
functional operating system. On top of the operating system, a variety of development
tools can be used to develop software. At the same time, uses of smartphones can
choose to install the various software at any time according to their needs. Hence the
smartphone is flexible and can perform a variety of tasks under a variety of conditions.
Smartphone has high performance-price ratio. Due to the large user demand, various
manufactures are competitive, which leads to the low profitability of the product.
Therefore the smartphone can be purchased at low price, which makes smartphone
suitable for large-scale deployment.
Generally smartphones are equipped with a large touch screen and easy to use user
interface (UI), which is suitable for all types of users, even those who have not
previously used smartphone. So that the usage of software on smartphone is easy.
Both the operating system and hardware of smartphone have gone through several
generations of development. Hence they are matured with not many bugs on software
or hardware, which brings good stability and low failure rate to the smartphone.
Smartphone has been more and more widely used. For example, using smartphones as
data acquisition and processing device to treat children amblyopia (Tao Yin, Dalong
Zhang, et al. 2017), or using smartphone cameras to take pictures of human faces for
fatigue detection (Chen Jia, Dalong Zhang, et al. 2017), have received good effect.

3.5 Conclusion
Mobile phone is the most suitable equipment for the project.
The target of the project is to put the phone inside the vehicle and fix it to the greatest
extent; collect data with sensor mounted inside the phone; conduct analysis and process
for the data to obtain the information on roads or vehicles.
In order to achieve the above aims, we need to take into the following factors:

1. Realizability: use the smart phone as detection tool and the measurable physical
quantities are only those measured and resulted by the sensors mounted inside the phone,
just referring to the acceleration sensor, gyroscope and GPS mentioned in last chapter;
2. Easy operation: as a measuring method with final target users being general masses,
it shall be operated easily to the greatest extent. It is not needed to conduct much deep
specialized training for users who can still learn it quickly and operate it easily. And
 58

the user will not make serious mistake during mistake, resulting in unpredictable result.
Take some examples: simplify the operation to such extent as only one start button in
software interface for operating or even free from interface as the software can operate
automatically for a long term in the background. In addition, simplify the fixation way
of phone inside the vehicle for easy operation. For example, consider to fix the phone
onto the support above the centre console (as it is the most common position where the
phone is placed), or the user even does not need to consider the placing of phone. Either
placed in any position inside the vehicle or even put inside the pocket of the user, the
phone can work properly and analyse the accurate results.
3. Accuracy: the system shall be able to conclude accurate results as much as possible
under different conditions. At early experiment stage, in order to simplify test, remove
unnecessary disturbances to the greatest extent. In addition, in order to solve the
primary key problem, design an experiment scheme which has removed minor
disturbances and improved accuracy.
4. Robustness: the system shall have anti-jamming capability. In case of some minor
faults, it shall be provided with some reserved schemes and correction function.
Softwares mentioned in Table 3-6 are tested and show unsatisfying detection rate. It is
hard to find exact reason without access of source code of softwares. One possible
reason is that the process to sensor data is defective, e.g. Ineffective denoise and
baseline drift may affect the accuracy of detection. It reveals that a good data processing
algorithm is essential to obtain high accuracy.
 59

Chapter 4. Experiment design

In Chapter 3 the smartphone is chosen to be fixed inside vehicles as the detection device.
Then in Chapter 4, the detailed experiment plan is discussed, including which road will
be tested, on which vehicle the smartphone will be fixed, with which software to use to
collect road data, and which smartphone will be used.

4.1 Introduction to road for experiment

The common motor lanes can be divided into highway and non-highways. The roads
inside cities are mostly non-highways with many driving limits that may affect
experimental result, such as numerous turns, traffic lights that may cause frequent
acceleration and deceleration, more complex surface of road and more up and downs
that may bring difficulty to verify early experimental results. Therefore, the flat and
straight road sections are selected at early experiment stage, which has low traffic flow,
distinct surface features and is easy to be tested repeatedly.
1. Dundee Tay Road Bridge

Figure 4-1 Dundee Tay Road Bridge

The Dundee Tay Road Bridge was established in 1966 with total length of 2250 meters.
The bridge carries the A92 road across the Firth of Tay from Newport-on-Tay in Fife
to Dundee in Scotland. It is one of the longest road bridges in Europe. The bridge
consists of 42 spans. It is coloured in orange in Figure 4-1. The Figure 4-1 and Figure
4-2 also show that the main body of bridge is straight. There is no traffic light on the
bridge, and from the field survey, it is found out that the sidewalk is paved with cement
 60

boards of length of 2 meters each. The number of cement boards between adjacent
expansion joints is used to safely estimate the distance between adjacent expansion
joints, which is later used as verification of the detection method of this study. So the
expansion joints of the bridge can be considered as the feature of road surface which
makes it easy to verify in the experiments (Zhang, 2005).

Figure 4-2 3D model of Tay Road Bridge

Typical expansion joint on Tay Road Bridge is shown in Figure 4-3.

Figure 4-3 Typical expansion joint on Tay Road Bridge

2. Edinburgh Forth Road Bridge

 61

Figure 4-4 Forth Road Bridge

 62

Figure 4-5 3D model of Forth Road Bridge

The Edinburgh Forth Road Bridge (the blue and orange part in Figure 4-4) was
established in 1964 with total length of 2512 meters (Andrew et al., 2006). The road is
divided into three sections from south to north: S1 (408m), S2 (1006m), S3 (408m) in
suspension type. The daily traffic flow in average is about 65000 vehicles (Carter et al.,
2009). The speed for the vehicles in the bridge is limited to 50 miles/h lower due to
strong wind or other severe weather. Because of the architectural feature of the bridge,
the surface of bridge has arc shape instead of complete flatness (Colford, 2008).
 63

Figure 4-6 Three sections from south to north of Forth Road Bridge: S1 (408m), S2 ( 1006m) and
S3 (408m)

The vehicle was driven at the constant speed of 50miles/h in straight road for several
times. In order to keep constant driving speed, select the time ranging from 1:00 to
3:00a.m, when there is smallest traffic flow for measurement.
 64

There are expansion gaps distributed in the bridge floor. When the vehicle passes by,
there appears obvious bumping. It is deemed that these gaps represent the potholes with
typical sizes on the road. If these expansion gaps can be detected, potholes of size larger
than width expansion gap should also be able to be detected.
According to the general driving experience on the road, the expansion gaps on road
bridges (including Forth Road Bridge and Tay Road Bridge), have the average width
of 2-4cm which may result in obvious bumping of vehicles. Hence, it is considered that
the pothole with such size appearing on the road needs to be detected. And if the width
of pothole is smaller than that value (for example, the width is 1cm), obviously it is too
small which distributes widely on the road. Where it is required to detect such pothole
mandatorily, it will need much more calculation resource but with little sense. As a
result, it is appropriate to set the width of expansion gap on the road bridge as the lower
threshold value (namely, the minimum value) for the width of pothole needing detection.
In further analysis, when the pothole has such narrow width (same with that of the
expansion gap), the tire of the vehicle cannot contact the bottom of pothole. And no
matter how depth the pothole has, it will bring about the same influence and result.
To sum up, set the lower threshold value as 4cm for the width of pothole and under
such width condition, the depth will not be considered.
Hence, the width of expansion gap on the road bridge completely meets the minimum
width value requirement of the pothole which can be treated as the experiment subject
for the research of potholes. Besides, it is acceptable that the detection sensitivity can
be improved without a large increase in calculation quantity.
Features for treating expansion gaps on the road bridge as simulants of potholes:
1. The size value is close to the lower value of pothole. And if the size of pothole similar
to that of expansion gap can be detected, the potholes larger than the gap can be
certainly detected.
2. The road bridge floor is a straight line so that the project can focus on the detection
on the smooth and straight road section with unnecessary disturbances removed;
3. With low traffic flow, it is easy to control the driving speed to prevent the speed-up
and down influencing the results. (Detection time is 1:00 – 3:00a.m.).
4. Proper concentration degree: generally, there is one pothole every 20m with
corresponding interval time being 3s;
5. It brings about convenience for the in-field investigation, photo taking, result
comparison and handling;
 65

6. Same size of all potholes: keep consistency;

7. It is available to conduct experiments for several times within short time to improve
the experiment efficiency;
8. It is flat pavement generally except for the expansion gags, which provides easy
comparison;
9. The additional uneven areas (small-sized cracks) can provide further comparison
details. In conclusion, select the expansion gaps on two sections of bridge as the
research subjects.

4.2 Vehicles to be tested

In order to reduce the influence of gear shifting to the minimum, automatic gearshift
vehicles are selected. Due to limited conditions and considering the universality and
pertinence of application in the future, small and medium-sized domestic vehicles are
selected. The two detection vehicles used in experiments are shown in Table 4-1.
 66

Table 4-1 Two detection vehicles used in experiments

Detection Vehicle 1 Detection Vehicle 2

Nissan Micra K11 Saab 9-3 2.0T
Manufacture year: 1996 Manufacture year: 2002
Automatic transmission: Automatic transmission
continuously variable transmission
(CTV)
Displacement: 1.0L Displacement: 1998cc
Three doors and five seats Four doors and five seats
(hatchback)
Front engine and front drive Front engine and front drive
Dead-weight: about 700kg Dead-weight: 1500kg
Tire: 155/70R13 Tire: 125/85/16
Length: 3700mm, width: 1600mm, Length: 4630mm, width: 1800mm,
height: 1440mm, wheel tread: height: 1440mm, wheel tread:
2300mm 2675mm

4.3 Collection software

The software Sensor Insider Pro is adopted for Android platform smartphones
(Samsung and Nexus 7). Developed by Lucana Tech, Sensor Insider Pro can record all
kinds of data collected by built-in sensors with self-defined sampling frequency and
four set of output format.
In the experiment, the parameters to be measured include acceleration, rotation, GPS
coordinate and GPS speed. In order to meet the accuracy requirement of the experiment,
the sampling frequency needs to be set to the highest level. On Samsung Note the
sampling frequency is set as 150Hz. All the data is stored in MAT format in the form
of (Acc means accelerometer, Gyro means gyroscope, x/y/z means three axes, t means
time stamp):
Acc_x = […….];
Acc_y = […….];
Acc_z = [……];
Acc_t = [……];
Gyro_x = [……];
Gyro_y = [……];
 67

Gyro_z = [……];
Gyro_t = [……];
…..
The advantage of this file format lies in that it can be directly imported into Matlab
without any modification (as Figure 4-7 shows). Note that the expression of name of
parameter such as Location_GPS_Spd()=[……] is illegal in Matlab and result error
message in importing process. The solution is to open .MAT file by txt editor (such as
notepad++ or ultraedit) and delete the brackets in Location_GPS_Spd()=[……], so that
it looks like Location_GPS_Spd=[……].

Figure 4-7 Parameters measured and imported into Matlab

The data is stored in a similar way for GPS. Moreover, GPS stores KML files, which
makes it easier to load and display the route in Google maps.
The software Sensor Data is used for iOS platform (iPhones). It has similar basic
functionality with the Sensor Insider Pro. Both software can access the measurements
of built-in sensors and store the data in smartphone memories. The difference is that
iOS does not have public file system, the data can only be output by tools as iTunes or
iTools. In addition, Sensor Data cannot store data in .MAT format. Instead, the data is
saved in .csv format which can also be imported into Matlab.
 68

4.4 Experiment scheme

Samsung Note (GT-N7000), Apple iPhone4 and Google Nexus 7 are chosen as
detection devices. The differences between detection devices are shown in Table 4-2.
Table 4-2 Differences between detection devices

Seri Brand Model Operation Embedded sensor

al system
1 Samsung Note (GT-N7000) Android 4.0.1 acceleration sensor,
gyroscope, GPS
2 Apple iPhone4 iOS 6.3 acceleration sensor,
gyroscope, GPS
3 Google Nexus 7 Android 4.2.2 acceleration sensor,
gyroscope, GPS
The data is collected between 1-3am with minimum traffic. The smartphones are fixed
inside the locker under the centre console to the most extent, which makes it able to
consider the vibration of phone is closely similar to that of vehicle. (Directions of axes
defined in Figure 4-8) X-axis of the phone points to the direction of the vehicle’s head.
Due to the limited conditions, Mobile is fixed with its LCD screen facing up, and its
front (y direction in Figure 4-8) facing head of vehicle, and z-direction of mobile
forming an angle with horizontal plane so that the driver can see the LCD screen. It is
similar to the placement of mobile when mobile equipped SATNAV (satellite
navigation system) is used. In this way gravity has components on both z and y
directions. And it not only simplifies the experiment but also is featured by some
universality.
 69

Figure 4-8 Axes direction defined by accelerometer and gyroscope

Keep the vehicle pass by the bridge at a constant speed of 40miles/h (on Tay Road
Bridge) or 50miles/h (on Forth Road Bridge) in straight line for several times.
Before measurement, make sure the smartphone has enough storage space left and
sufficient battery life. Restart the smartphone without operating any other software,
then start the measuring software. Start the vehicle and preheat for 1 minute. It also
gives the GPS time to search the satellite, as GPS receiver normally takes tens of
seconds to search the satellite and calculate the position (the actual time depends on the
number of connectable satellites).
In order to verify the measurement accuracy of built-in sensors (mainly GPS to measure
vehicle speed), connect Nissan Micra with OBD II Bluetooth module and adapter to
read the speed data from the sensor built inside the vehicle, and then transmit the data
to Samsung Note through Bluetooth for storage . The speed data from vehicle sensor
and GPS speed data of the smartphone have good quality, which proves that the GPS
speed data of the smartphone meets the requirement of the research.
Here are some data example:
Figure 4-9 shows acceleration data collected from road from Dundee to Edinburgh.
Note that the curve at the top is acceleration of z-axis direction, which is perpendicular
to the ground. Because the gravity acceleration always acts along the z-direction,
therefore there is always a direct current (DC) component with the value close to
10m/S2.
 70

20

15

10

-5

-10
Acceleration
x

-15 Acceleration
y
Acceleration
z
-20
0 2 4 6 8 10 12
Points 10 5

Figure 4-9 1145642 points collected from Dundee to Edinburgh on 24/09/2012 with sample rate of
200Hz. Sample time about 5728 seconds

As can be seen in the Figure 4-9, acceleration data obtained at some points is clipping
distorted (clipped signal is shown in the blue box in Figure 4-9). This is because the
max range of acceleration in mobile or panel is about 20 m/S2, which is about 2g. As
there is always the gravity acceleration along the z-axis, it can be considered as a
numerical 1g DC component. Considering the maximum z-axis acceleration is 2g, the
biggest signal along the z-axis is only 1g. If the acceleration is beyond the range, the
clipping distortion occurs. However the bumping of vehicles are likely to be greater
than 1g. Although the range of built-in accelerometer in mobile or panel may be
changed (for example to exceed the range to 6g) by the software to avoid this kind of
problem, the accuracy (generally within 2g range the accuracy is between 0.04-0.004
m/S2) will decrease.
 71

15

10

Acceleration
x
5 Acceleration
y
Acceleration z

-5

2.5 2.6 2.7 2.8 2.9 3 3.1

Points 10 5

Figure 4-10 Zoom in of previous diagram, containing about 6100 points, sampled in about 30
seconds

14

12

10

6 Acceleration
x
Acceleration
y
4
Acceleration
z
2

-2

-4

2.69 2.695 2.7 2.705 2.71 2.715 2.72 2.725

Points 10 5

Figure 4-11 Further zoom in of previous diagram. About 3100 points measured in 15 seconds

Gyroscope sensor data shows as Figure 4-8:

 72

Voltage

Figure 4-12 1145642 points from Dundee to Edinburgh, measured on 24/09/2012 with sample
frequency of 200Hz. Sample time about 5728 seconds

General precision of built-in gyroscope in phone or panel: it is easy to know from the
Figure 4-12, that the reading of gyroscopes ranges between -0.5 - +0.5 rad/S (The output
value of Gyroscope is measured by voltage. For the same input, various Gyroscope
chips may output different values, but the output values are proportional to the angular
velocity (measured by radius/second). Therefore for the sensor used in this study, the
output data can be considered as being proportional to the angular velocity, which is
measured by radius/second). The gyroscope does not produce clipping distortion like
accelerometer sensor does, and all the waveforms can be properly recorded. In addition,
the precision of gyroscope is normally about 0.0003 per cent, and also better than
acceleration sensor.
 73

Voltage

Figure 4-13 Zoom in of 4-5, containing about 25000 points, or 125 seconds

The kml data of GPS can be shown in Google maps, as Figure 4-10 shows:
 74

Figure 4-14 Dundee to Edinburgh, 24/09/2012. Google maps can only display a small part of
geometry information of kml file, so that the length displayed in this screenshot is 6.91 miles.

Speed data can be obtained by GPS. The positioning information directly measured
with GPS is low in precision, generally error in from a few meters to dozens of meters;
however the speed information is measured using the Doppler effect and is normally
high in precision.
The most direct method to use GPS to measure vehicle speed, is to firstly find current
geographic coordinates, and measure the geographic coordinates after a period of time;
then calculate the distance traveled from the two coordinates, and divide by the time
interval. However, the accuracy of GPS is about 10 meters (for civilian purpose), which
brings large error into calculation. Therefore another method is often used by GPS in
 75

real practice, which is based on Doppler Effect. The signal of clock frequency offset is
detected to obtain the relative speed between GPS receiver and satellite. As the
frequency of the clock signal is very stable, the accuracy of this method is very high,
with the error reduced to be within 1km/h. Therefore, the accuracy of the speed
measured by a smartphone is within 1km/h.
Speed (m/s)

Figure 4-15 Sample of speed data, 5613 points from Dundee to Edinburgh, collected on
24/09/2012

4.5 Summary of chapter

The design of the experiment should follow these conditions:
1. Fix the smartphone inside the vehicle, as discussed in Chapter 3;
2. The vehicle drives straight at constant speed, so the noise can be reduced to
minimum;
3. In order to maintain the stability of the speed, the experiment need to be operate
between 1:00-3:00 in the midnight, which meets the smallest volume of traffic flow.
After the analysis of the characteristics of data collection devices, the data collection
process is presented. And following this process, the real data is collected for further
processing and analysis.
76
 77

Chapter 5. Modelling and simulation

5.1 Simulation of suspension system

In this section, simulation results of the three types of vehicle suspension model in
chapter 2 will be discussed.

5.1.1 2-DOFs suspension system of 1/4 vehicle

A simulation model for the suspension system of 1/4 vehicle in line is established with
the equation 2.1 in MATLAB/SIMULINK (as the following diagram shows):
 78

Figure 5-1 2-DOFs simulation model of suspension system of 1/4 vehicle in

MATLAB/SIMULINK

When the sine wave is used as the road input excitation:

 79

Figure 5-2 Use sine wave as the road input excitation, x-axis represents time (measured by
second), and y-axis represents displacement of tyre in the vertical direction (measured by meter).

The output wave is as in the following diagram:

Figure 5-3 Output wave of 2-DOF vehicle model

 80

The above figure shows that the suspension can absorb the energy of the shock, and
after a period of time, the magnitude of the shock will become stable on a small value.

5.1.2 4-DOFs suspension system of 1/2 vehicle

In this section, the Pitch model is analysed. Figure 5-4 is the simulation model of 4
DOF 1/2 vehicle designed according to Equation 2.2 using Simulink. Figure (a) is the
graphical block diagram of the simulation model, which is formed by a road spectrum
generator, a model parameter input module, and a 4 DOF 1/2 vehicle model and result
output module. The road spectrum generator is the input for the 4 DOF 1/2 vehicle
model, and it represents the feature of the road surface (Guanqiang and Fang, 2007). In
this study the sine wave is used. Other input, such as white noise, can be used to replace
this part to study various types of road surfaces. Figure (b) shows the detail of the grey
block of Figure (a), in which result is shown as waveform by oscilloscope. The result
is also saved in files for further analysis. Figure (c) shows the detail of the yellow block
of Figure (a), which is the implementation of Equation 2-2 by modules such as
differentiator, summator and multiplicator.
81

(a)

(b)
 82

(c)

Figure 5-4 4-DOFs simulation model of suspension system of 1/2 vehicle in

MATLAB/SIMULINK. (a) is the graphical block diagram, (b) shows the detail of the grey block
of Figure (a), (c) shows the detail of the yellow block of Figure (a).

When the sine wave is used as the road input excitation:

 83

Figure 5-5 Use sin wave as the road input excitation

The vertical distance output signal is shown in Figure 5-6.

Figure 5-6 Vertical distance output wave of 4-DOF vehicle model

From Figure 5-6, it is known that the vibration of vehicle in vertical direction quickly
stabilises to a fixed value.
 84

5.1.3 7-DOFs suspension system of the whole vehicle

The system simulation model is established in MATLAB/Simulink:

Figure 5-7 7-DOFs simulation model of suspension system of the whole vehicle in
MATLAB/SIMULINK

The parameter subsystem is established as Figure 5-8 shows:

Figure 5-8 Parameter subsystem of 7-DOFs simulation model of suspension system of the
 85

5.1.4 Simulate suspension using MATLAB/Simulink

Simdriveline is a new expansion of MATLAB Simulink, providing a complete set of
analysis tool for mechanical modelling and simulation of transmission and driving
system. It is an ideal tool for dynamic simulation of vehicle parts.
A sample suspension model designed in simulink using simdriveline is shown in Figure
5-9.

Figure 5-9 A sample suspension model designed in simulink using simdriveline

Figure 5-9 is the 1/2 vehicle model designed with simdriveline. Although this model
has a concise form, it is also based on Figure 2-3 and Equation 2-2. As this model uses
front suspension module and rear suspension module, which are encapsulated in
simdriveline, it is impossible to open these two modules to see the detail. A module
designed in this study based on the same theory is displayed in Figure 5-10.
 86

Figure 5-10 Sub-design of Figure 5-9, the front and rear suspension

According to parameters of general vehicles, default initial values are chosen as follows.

Lf = 0.9; % front hub displacement from body gravity centre (m)

Lr = 1.2; % rear hub displacement from body gravity centre (m)
Mb = 1200; % body mass (kg)
Iyy = 2100; % body moment of inertia about y-axis in (kg)
kf = 28000; % front suspension stiffness in (N/m)
kr = 21000; % rear suspension stiffness in (N/m)
cf = 2500; % front suspension damping in (N sec/m)
cr = 2000; % rear suspension damping in (N sec/m)

Focus on the Road Height input in Figure 5-11. Let it be a step input, as shown in Figure
5-11, which corresponds to vehicle driving over a road surface with a step change in
height. Let the initial height of road is 0, then at time point7, the height changes to
Height=0.01.
 87

Figure 5-11 step input of road, height changes to 0.01 at time point 7

Run the designed suspension model, the vertical displacement of the vehicle body can
be obtained, as Figure 5-12 shows.

Figure 5-12 The vertical displacement of vehicle after it passes the step change of road.
height=0.01m
 88

The result shows that, the minimum value of vertical displacement is 0.12m, the
maximum value is 0.1325m. Thus the vertical displacement with road height = 0.01m
is 0.1325-0.12=0.0125m.
Similarly, the vertical displacement with road height of 0.02, 0.03, 0.05, 0.1 and same
other starting conditions can be obtained, as shown in Table 5-1 (unit is m).
 89

Table 5-1 vertical displacement of vehicle body with various step change of road (unit: m)

height of step minimum maximum displacement of

change of vertical vertical vehicle body h=max-
road displacement of displacement of min
vehicle body vehicle body
0.01 0.12 0.1325 0.0125
0.02 0.12 0.1449 0.0249
0.03 0.12 0.1573 0.0373
0.05 0.12 0.1821 0.0621
0.1 0.12 0.2440 0.1240

0.25
Maximum displacement of vehicle body
Displacement of vehicle body

0.2

0.15

0.1

0.05

0
0 0.02 0.04 0.06 0.08 0.1
Height of step change of road (m)
Figure 5-13 Relationship between height of step change of road, displacement of vehicle body and
maximum displacement of vehicle body

From Table 5-1 and Figure 5-13 show that if we only consider the suspension system,
the vertical displacement of vehicle body and the maximum displacement of vehicle
body are proportional to the height of step change of road.
 90

5.2 Tyre simulation

Figure 5-14 Tyre with center O passes pothole AB with width of w. The velocity of tyre before
passing is v, the velocity after it bounces up from road is v’ which can be decomposed to a
component of vx’ horizontally, and a component of vy’ vertically.

As shown in the figure 5-13, assume the vehicle runs from left to right at the velocity
of v with the tyre radius of R and the pothole width of w, when the rim of the tyre
contacts the cutting edge A and trailing edge B of the pothole at the same time, collision
occurs. That is to say, at this point, both A and B are on the circumference of the tyre.
According to symmetry, it is known that by drawing a vertical line OC from the centre
O to AB, the intersection C is located in the middle of A and B, i.e. AC=CB.
In case of taking no account of the energy loss (Wang et al., 2008), According to
Appendix A, the bouncing height is
v 2 w2
h (5.1)
8 gR 2
Based on the above formula, the gravitational acceleration g is constant, and assume
the tyre radius R remains unchanged, the bouncing height is proportional to the square
of velocity v, as well as to the square of the pothole width w.
Assume the vehicle running at 50mile/h, i.e. 22.35m/s, the pothole (bridge expansion
joint) width w varies with the temperature change, and the width in winter and summer
differs. The mean value 5cm, i.e. 0.05m is taken here; the gravitational acceleration is
 91

taken as g=9.81; the specification of the tyre of Nissan Micra K11 used for this test,
which manufactured in 1996 is 155/70R13, which means:
 155 denotes the tyre width in mm;
 70 denotes flatness ratio, i.e. the ratio of the tyre wall height to the tyre width and
70 representing 70%; generally, the flatness ratio ranges from 30% to 80%;
 R denotes the abbreviation of Radial, which means the tyre is of radiation layer
structure;
 13 denotes the nominal diameter of the wheel rim in inch.
The tyre’s total outer diameter is the sum of the hub’ outer diameter and the two tyres’
wall height, therefore, the outer diameter is
13  0.0254  0.155  0.7  2  0.5472m
Then, the radius
0.5472
R  0.2736m
2
Therefore
22.352  0.052
h  0.0582m
8  9.81 0.27362
The formula above indicates that bouncing height of the tyre is approximately 5.8cm,
when the vehicle contacts the pothole of 5cm in width at 50mile/h, if the loss of energy
is not taken into account.
In fact, however, two principal losses should be taken into account: the energy-
absorbing deformation of the tyre, and that of the vehicle caused by the shock and
suspension, both of which, in essence, convert the mechanical energy into other forms
of energy (such as heat energy) and then released. Therefore, the mechanical energy
before and after the collision is no longer conserved.
Assume that the partial energy loss during the collision is converted into other forms of
energy like heat energy, and define the consumed energy as k, and then
1
mvy 2  mgh  k (5.2)
2
Namely
v 2 w2 k
h 2
 (5.3)
8 gR mg
 92

It is can be seen that the bouncing height is not only proportional to the square of speed
and that of the width of the pothole, but also related to the mass of the vehicle and the
consumed energy.
A further analysis of the process of collision is conducted as below: the collision can
be divided into two steps from the tyre contacts the edge of pothole:
1. As the centre of the tyre approaches to the edge of the pothole, the tyre will deform
due to depress, and in this process, the kinetic energy of the tyre converts into elastic
potential energy. This process comes to its end and enters the next step after the pressure
in question reaches its maximum.
2. The tyre centre starts to leave away from the pothole edge, namely, jump upward.
The tyre begins to resume the deformation, and now the elastic potential energy formed
by depressing in step 1 again converts into kinetic energy. The step ends with the
deformation fully resumed and the separation from the pothole.
Set u ' and v ' as the velocities generated after the collision between ground and tyre
separately, while u and v stand for that prior to the collision, then
u ' v '
Cr  (5.4)
uv
in which Cr is Elastic Loss Coefficient of tyre. Due to the Elastic Hysteresis Loss, part
of energy is lost (in the form of thermal energy) when the tyre deforms. Assume the
ground is fixed, namely, the velocities before and after the collision are both 0, then,
u ' v ' v '
Cr   (5.5)
uv v
It can be see that, without energy loss, the mechanical energy is conserved and the
velocities before and after the collision are equal, i.e.,
u ' v ' v '
Cr   1 (5.6)
u v v
Therefore
v'  v
This is the very matter of the conservation of mechanical energy without energy loss in
question. Under this circumstance, the bouncing height of the tyre is proportional to the
square of the velocity.
Actually, the coefficient of restitution is always smaller than 1, therefore,
u ' v ' v '
Cr   1 (5.7)
u v v
 93

According to Appendix B,
v '2 w 2 C r 2 v 2 w 2
h  (5.8)
8 gR 2 8 gR 2
If the parameters of the vehicle and the width of the pothole are unchanged, then, w, g,
R and Cr , in the formula are constants, therefore, the bouncing height h is still

proportional to the square of velocity v.

The equation 5-8 shows the relationship between vehicle speed, pothole width and
bounce height: when pothole width is constant, the vehicle bounce height is
proportional to square of vehicle speed. When vehicle speed is constant, vehicle bounce
height is proportional to square of width of pothole This shows that if vehicle bounce
height and speed of vehicle are measured, the width of pothole can be calculated with
Equation 5-8 (Note that "pothole" here represents a large enough pothole, on which the
vehicle vibrates. That is to say, if the width of a pothole is too small, this pothole needs
not to be detected, and should not be detected). In other words, to determine whether
there exists a pothole, the detect vehicle bounce height needs to exceeds the threshold,
the threshold setting should not be fixed, instead, it should be changed with the vehicle
speed. The equation 5-8 reveals that, the threshold should be set proportional to the
square of the vehicle speed. When the vehicle speed is slow then the threshold should
be reduced, which can increase the detection accuracy; and when the vehicle speed is
fast, the threshold increases, which can reduce the miscarriage of justice. In the previous
literature, threshold values are fixed, so they got many miscarriages of justice. Equation
5-8 is used to generate speed-dependent dynamic thresholds, so the accuracy of the
theoretical tests is much higher, which is proved in detection result shown in Chapter
7.

5.3 Conclusion
In this chapter, the factors of bounce height of vehicle and their relationship are studied.
The suspension system and the tyre system were modelled and simulated in order to
study the bounce height of vehicle. According to the equation of suspension system in
Chapter2, 1/4, 1/2 and whole vehicle models are established, and then simulated by
using Simulink. The conclusion is: if all other factors are unchanged, the bounce height
of the vehicle has a linear relation with the height of the pothole.
Next, the corresponding formula of vehicle speed and the height of the bounce is
established according to the law of conservation of energy in the simulation of tyre
 94

system. The conclusion is: if all other factors are unchanged, vehicle bounce height has
quadratic relationship with the speed of the vehicle. That leads to the conclusion that
the amplitude threshold is proportional to the square of the speed of the vehicle. This
conclusion will be verified in the experiments in Chapter 7.
 95

Chapter 6. Data processing

6.1 Coordinate correction

There are three coordinate systems that are related to each other, they are mobile phone
coordinate system (Figure 4-4), vehicle coordinate system (Figure 6-1) and global
coordinate system (Figure 6-2), which is based on the starting point of the vehicle. In
order to get the pothole data, mobile phone coordinate system need to be compared with
global coordinate system. Since only the flat section is considered, vehicle coordinate
system and global coordinate system are coincident. Therefore, when rotate the data in
the mobile coordinate system, until it match the vehicle coordinate system, then the
data can be converted into the global coordinate system, so that the position of the
pothole can be obtained. This coordination transformation has nothing to do with the
location information provided by GPS, but to derive the travel distance from
acceleration data gathered by built-in sensor.

Figure 6-1 Vehicle coordinate system

 96

Figure 6-2 Nature coordinate system

The built-in sensors have their own three-dimensional coordinate system. The output
data of sensor that has directions, such as acceleration, is represented in the sensor’s
own coordinate systems, which is called sensor coordinate system. The sensor
coordinate system has X, Y and Z axis that are orthometric. The vehicle, on the other
hand, has its own three-dimensional coordinate system: the X-axis is the direction of
vehicle movement, Y-axis is the right-hand side of vehicle, and Z-axis is vertical to the
chassis and pointing down. There is another (three-dimensional) coordinate system,
which is called natural coordinate system. When vehicle is in uniform linear motion on
a horizontal road, at one time point the vehicle coordinate system is considered as the
natural coordinate system. Thus the X-axis of natural coordinate system is the direction
of vehicle at the point it is set, the Y-axis is the right-hand side of vehicle, and Z-axis
is the direction of gravity.
Note that:
1 The smartphone is fixed on the centre console, facing the rear window or the driver’s
seat and slightly upward sloping. Thus the sensor coordinate system cannot overlap
with vehicle coordinate system;
2 Consider the smartphone and vehicle are rigid connected, the angle between sensor
coordinate system and vehicle have can be considered as fixed;
3 Natural coordinate system never changes after it is set, while the vehicle coordinate
system moves together with vehicle;
4 As in all my experiments, the vehicles are in uniform linear motion, the vehicle
coordinate system is considered overlapped with natural coordinate system.
 97

The pothole is in natural coordinate system, but the data from sensor is in sensor
coordinate system. As the natural coordinate system is considered overlapped with
vehicle coordinate system, the data from sensor should be rotated by the angle between
sensor coordinate system and vehicle coordinate system.
The data can be processed according to the direction of gravity. Because no matter
where is the smartphone is fixed or the angle, the direction of gravity is always the Z-
axis of natural coordinate system. Hence the data from the acceleration sensor are added
together in vector form, which is:
Z 2  x2  y2  z 2 (6.1)
In equation 6-1, Z is the vector of gravity (including its direction and size), on the same
direction of Z-axis of natural coordinate system (vertical to the horizontal plane,
pointing downward); x, y and z are the acceleration on X, Y and Z-axis of the sensor
coordinate system. So:

Z  x2  y 2  z 2 (6.2)

But this transform method requires the vehicle to be in uniform linear motion. Once
there is jolt (can be considered as noise in the field of Signal Process), there will be
error in the value and direction of gravity.
So actually the data is transformed using principal component analysis (PCA). PCA is
a processing method with linear transformation of multiple variables for data analysis
and mathematical model establishment. It is invented by Karl Pearson in 1901. In the
practical research, many relevant variables are chosen to be studied to comprehensively
analyse the problem, in which each variable reflects information in some certain aspects
of the result. The main function of PCA is to select important variables from multiple
ones.
PCA gain the principle component by conducting eigen decomposition to the
covariance matrix. The result of PCA is the eigen vector and eigen value. The
mathematical definition of PCA is: an orthogonalization linear transformation in which
the data is transformed from an old coordinate to a new coordinate system, making the
variable with largest variance of any projection be on the first axis (also called as the
first principle component) and the second variance on the second axis (also called as
the second principle component), and so forth.
Define an n*m matrix, XT is the average value (move to the original point from the
centre of the average value), the row is data sample and the column is data category
 98

(attention: the definition is for XT not X). So the singular value decomposition of X is
= , in which the matrix W with the size of m*m is the eigen vector matrix of
XXT, with the size of m*n is a nonnegative rectangle diagonal matrix, V with the size
of n*n is the eigen vector matrix of XTX. According to Appendix C,
XX T  W  TW T (6.3)

Given a set of points in Euclidean space, the first principal component is corresponding
to the line through equalization point of multi-dimensional space, ensuring the
quadratic sum of the distance between each point and this line is minimal. After
removing the first principle component, get the second one with the same manner, and
so forth. Providing an effective method to reduce the dimension, PCA is often used to
reduce the dimension of dataset and keep the largest contribution of dataset to the
variance. This is achieved by keeping lower-dimensional components and ignoring
higher-order components, thus the lower- dimensional components can keep the major
aspects of the data.
As previously discussed, when the smartphone is fixed in the vehicle, the three axes of
sensor coordinate system will form an angle with that of the vehicle coordinate system.
When the vehicle stands still or is in uniform linear motion, because of the gravity,
there is acceleration along Z-axis of the vehicle coordinate system with value of g, and
in the X and Y axes the acceleration should be zero. However, in practice, the three
axes are all affected by the vibration caused by the vehicle body, which is usually
random noise signal. So on the Z-axis of vehicle coordinate system, the acceleration
value is the sum of gravity g and noise signals. Generally, the acceleration of gravity is
far more than that of the noise signals, while there is only noise signals on the other two
axes. As the three axes are vertical to each other, they are not interrelated. On the other
hand, because components of results of PCA are orthometric (which mean they do not
related to each other), it can be used to do the coordinate correction of acceleration data.
From the above discussion we can see that after PCA processing, the correlation
between 3 data sets was eliminated. In the original data sets collected by the sensor,
since the three axes of the sensor are not completely match with the nature coordinate
system, data of the three axes collected by the sensor acquisition have some correlation.
In other words, if there is a signal in the Z' axis direction of the nature coordinate system,
then it will be shown in the X, Y, Z triaxial data of the sensor. Our aim is to rotate the
coordinates so that in the final result, the signal is only shown on the Z axis, and the
 99

other two axes have a signal amplitude of 0, that is, to eliminate the correlation between
the three axes. PCA's function is through a conversion, transfer a group of coordinates
to another group of coordinate system, but in the new coordinate system, X, Y, Z
components are not related.
Generally, to make sure the minimum mean-squared error of approximate data is found,
zero mean processing shall be done to the data before PCA. But in my study, the
direction of gravity is the first principle component, and the mean value is necessary to
find the direction of gravity. Therefore the zero mean pre-processing shall not be
conducted, otherwise the result will be false.
As the vehicle is assumed to be in uniform linear motion, the vehicle coordinate system
is overlapped with natural coordinate system. The vibration of vehicle should only
happen on Z-axis and is irrelevant to X or Y axes (which are corresponding to the left-
right steering and braking/accelerating respectively), only the first principle (the
acceleration in Z-axis) is needed.
With the above methods, the data is transformed to the natural coordinate system, which
makes subsequent process and analysis easier.

6.2 Convert Acceleration Data into Distance

The data obtained by the acceleration sensor is the acceleration of the smartphone, and
because the smartphone is fixed on the vehicle, the acceleration can be considered as
of vehicle as well. Because the displacement on the vertical direction (along or
opposite the direction of Z-axis) happens on potholes, the position of each vertical
displacement is the position of each pothole. Thus my aim is to obtain the displacement
data of the vehicle, which is ∆Z. The sampling frequency (100 or 200Hz, which
depends on the detection device) is fixed during the sampling process. Theoretically,
speed value can be obtained through a single integration of acceleration data, and the
displacement value can be obtained through a single integration of speed.
In practice these two integrations may be done in the time-domain or frequency-domain
methods.

6.2.1 Time domain integral method

Time-domain integration is a way to obtain the displacement signal directly from
acceleration signal by quadratic integral. Normally trapezoid formula or Simpson
summation formula is used because of their intuitive form. However, the acceleration
 100

signal measured in the experiment actually contains noise and DC component, so the
result will contain trend term if the signal is integrated directly.
Let the acceleration signal be a  f (t )  T , in which T is the measurement error,
then Simpson integral is used (See Appendix D for full details):
t N
v  (ai2  4ai1  ai)
6 i2
(6.4)

In the equation 6.4, the symbol ∆t means sampling interval, and the symbol v means
the signal sequence obtained after a single integral.

6.2.2 Frequency domain integral method

According to Appendix E, Fourier transform of displacement can be obtained after the
second integral:
N 1 n
1 1  i 2 k
S ( n)   2 A   2
H ( k ) an e N
(6.5)
k k  0 (2 k f )

Where
1, f  kf  fu
H (k )   d
 0, else

In equation 6.5, symbol f is frequency resolution, and fd and fu are the upper

and lower limits of cut-off frequency respectively. N is number of points in data, and
k is the frequency of Fourier component.
Velocity and displacement signal can be obtained by processing Fourier inverse
transform to equation 6.4 and 6.5. This shows that, frequency-domain integral directly
employs relationship (phase swap) of sine and cosine integrals in the frequency domain,
which effectively avoids accumulation or amplifying effect on the small error of time-
domain signal in the process of integral, and makes the results more accurate (Ribeiro
et al., 2002). However, note that low-frequency signal has corresponding smaller k ,
and the denominator is larger in the equation 6.4 and 6.5. Considering the low precision
of the acceleration sensor for the low frequency signal, small measurement error will
cause large calculation deviation, so the low frequency signal is an important source of
error of frequency- domain integral.

6.2.3 Discussion
In order to validate the method of the above two integral methods, assume that an
acceleration signal is
 101

a  3  sin(2  30  t )  2  sin(2  8  t )  0.02 (6.6)

The sampling rate is 260Hz and sampling time is 2.25 seconds.
In equation 6-15, constant value 0.02 is used to represent a small DC component. The
curve of 6-15 is shown in Figure 6-3.

-1

-2

-3

-4

-5
0 0.5 1 1.5 2 2.5
Time (second)

Figure 6-3 Image of equation 6-15

Integrate 6-15
A   adt
  (3  sin(2  30  t )  2  sin(2  8  t )  0.02)dt (6.7)
cos(60    t ) cos(16    t ) t
   C
20   8 50
Let C=0, the result of single integration is shown in Figure 6-4.
 102

40

35

30

25

20

15

10

0
0 0.5 1 1.5 2 2.5
Time (second)

Figure 6-4 Integral result of equation 6-15

Figure 6-2 shows that, the constant value in original function results a linear component
in the result of single integration, and the result has a positive offset.
Integrate 6-16 and the second integration is shown in equation 6-17
A   Adt
cos(60    t ) cos(16    t ) t
  (   )dt (6.8)
20   8 50
2
sin(60    t ) sin(16    t ) t
    C
1200   2 128   2 100
Let C  0 , the result of second integration is shown in Figure 6-5.
 103

12000

10000

8000

6000

4000

2000

0
0 0.5 1 1.5 2 2.5
Time (second)

Figure 6-5 Double integral result of equation 6-15

Figure 6-5 shows that, the quadratic term adds a significant offset to the integration
result. This significant offset is caused by the minor DC component in the original
function. Without the DC component the result of single integration is shown in Figure
6-6.
 104

30

25

20

15

10

-5
0 0.5 1 1.5 2 2.5
Time (second)

Figure 6-6 Integral result of equation 6-15 without DC component

The result of second integration is shown in Figure 6-7.

 105

80

60

40

20

-20

-40

-60
0 0.5 1 1.5 2 2.5
Time (second)

Figure 6-7 Second integral result of equation 6-15 without DC component

Both Figure 6-6 and 6-7 show a signal without offset, which is the type of result needed
for this study. This case study shows the importance of DC component removal. Next,
the single and second integration is done to the same signal in time-domain and
frequency-domain. By comparing the result, it is known that which method is better.
As shown in Figure 6-8, the result of single integration in frequency-domain does not
contain offset.
 106

Figure 6-8 Integral result of equation 6-15 using frequency domain method

As shown in Figure 6-9, the result of second integration in frequency-domain does not
contain offset either.

Figure 6-9 Second integral result of equation 6-15 using frequency domain method
 107

Because the process of frequency-domain integration removes the DC component, it

does not introduce offset even in second integration, which is good for the follow up
processes of data.
In order to determine the accuracy of the two integration algorithms, we need to
evaluate the error. In general, there are three kinds of error evaluation indicators (Jiang
et al., 2009): average peak error, average maximum relative error and square sum error.
The average peak error is the mean value of the positive and negative peaks of the
integral displacement time course X(t) relation to the measured displacement time S(t)
positive and negative peaks, respectively,
| [ ( )] [ ( )]| | [ ( )] [ ( )]|
Erp = × [ ( )]
+ [ ( )]
(6.9)

The mean maximum relative error is the mean value of the positive and negative peaks
of the relative error time course [X(t)-S(t)] relation to S(t) positive and negative peaks,
respectively:
| [ ( ) ( )]| | [ ( ) ( )]|
Err = [ ( )]
+ [ ( )]
(6.10)

"Square and Error" describes the energy difference between the integral and the
measured displacement:
∑ [ ( )] ∑ [ ( )]
Ersq = ∑ [ ( )]
(6.11)

Where x(i) and S(i) are the displacement samples of X(t), S(t), and N is the number of
sampling points of the signal.
The reason of the usage of the squard error is that, the peaks of two waveforms may not
always appear at the same time. Thus the overall effect of the waveforms, i.e. the energy
difference of the waveforms, should be examined.
There are 585 sample points in total. For each sample point, the integration is subtracted
by theoretical estimation, the result is error. Error obtained by calculating two integral
methods respectively as shown in the table 6-1:
 108

Table 6-1 Error obtained by calculating two integral methods

The The average quadratic

average maximum sum error
peak error relative error (Ersq) /%
(Erp) /% (Err)/ %
An integral of time domain 23.18 113.46 180.33
An integral of frequency domain 0.12 0.41 1.01
Second integral of time domain 354.8 565.24 673.28
Second integral of frequency 0.01 0.12 0.43
domain
According to Table 6-1, time domain integration introduces offset which becomes more
obvious with the increase of time of integration. On the one hand, the estimation of DC
component contains bias, because the signal is not integer-period sampled. On the
other hand, constant term occurs after the time domain integral, which cause the offset
in the result. In actual experiment, it is hard to conduct integral-period sampling,
because the signal contains many frequency components. Although the error of the DC
component estimation can be reduced by increasing sampling time, tiny error still can
lead to a major offset. It can be seen from the Table 6-1, that result has been distorted
in the second integral with time-domain integration, therefore the calculation is useless.

6.3 De-noising
Influenced by outside interference or artificial operation, the collected data inevitably
contains large amounts of interference signal, and is deviated from the real value. An
important task of digital signal processing is to reduce or eliminate the interference
factors and remove the high-frequency noise in the collected data, so that the processed
data is as close to its true value as possible.
The principles of de-noise are: (1) Smoothness: de-noised signal should have the same
smoothness with the original signal; (2) Similarity: variance of de-noised signal and
original signal should be as small as possible.
A model of a one-dimensional signal containing noise can be expressed as the following
equation:
s(t )  f (t )    e(t ) t  0,..., n  1 (6.12)
 109

In the equation 6-18, f (t ) is the real signal, e (t ) is the noise,  is the strength of
noise, and s (t ) is the signal containing noise. De-noise of signal s (t ) is to eliminate
the noise section in the signal and recover the real signal f (t ) from s (t ) . In practical
applications, the useful signals often manifest as low-frequency signals or stable signals,
while the noise signal usually manifests as high-frequency signal. Therefore, the de-
noise process can be done in accordance with the following methods: First, decompose
into low-frequency part and high-frequency part. The noise is often included in the
high-frequency part. Thus the high-frequency part is to decrease its amplitude. Finally,
the signal is reconstructed.   e(t ) is the noise section need to be eliminated in the
signal and recover the real signal f (t ) from s (t ) .

6.3.1 FFT
Fast Fourier transform (FFT) is a fast algorithm of discrete Fourier transform (DFT).
Fast Fourier Transform (FFT) improves the computation efficiency by using the
periodicity and symmetry of WN factor. FFT algorithm can be basically divided into

two categories: Decimation-in-time, (DIT) and Decimation-in-frequency (DIF). For

full details please refer to Appendix F.
In the de-noise process of the signal, as Fourier analysis is a way to analyse signals in
the frequency domain, any sudden change of the signal on the timeline will affect the
entire spectrum of the signal, so that it cannot show the change of signal at a certain
time point. In some cases, the target frequency exists not only in low frequency part of
the raw data, but also in high frequency part. In these cases, if the original signal has
high frequency in its initial stage, after being transformed by FFT, high-frequency
components is filtered out when the low-pass filtering is added, , which means the
useful signal is lost. This is the disadvantage of FFT as de-noise method.
A case-study is shown here, in which a cosine signal containing noise is de-noised by
using FFT. The cosine signal (without noise) is:
x (t )  cos(2 t )
Represented by MATLAB:
t=0:0.001:2;

x=cos(2*pi*t);

The waveform of this signal is shown in the Figure 6-12:

 110

0.8

0.6

0.4

0.2

-0.2

-0.4

-0.6

-0.8

-1
0 0.5 1 1.5 2
t (second)

Figure 6-10 Original input signal

A random noise is added to the cosine signal:

f=x+0.25*randn(1,2001);
The waveform of signal with noise is shown in the Figure 6-13:
 111

1.5

0.5

-0.5

-1

-1.5

-2
0 0.5 1 1.5 2
t (second)

Figure 6-11 Input signal with random noise

The signal shown in Figure 6-13 is composed by low-frequency cosine signal and high-
frequency random noise. The basic method of signal de-noise is to suppress the high-
frequency section of signal and reserve the low-frequency signal.
The FFT process can be divided into the following steps:
1) Perform FFT computation on signal;
2) Suppress noise section of the signal according to the spectrum;
3) Obtain de-noised signal by implementing Fourier inverse transformation to the
transformed spectrum.
In this signal, cosine signal is f(t), de-noised signal is g(t), the Fourier transform of f(t)
is F ( ) , similarly the Fourier transform of g(t) is G ( ) . Then the process can be
represented as:
G ( )  H ( )  F ( )
H ( ) is the system frequency response of low-pass filter. When used to filter high
frequency, typical H ( ) is in form of a step function in frequency domain. The value
of step function H ( ) is 1 for all frequencies lower than the cutoff frequency, and it
is 0 for all frequency higher than the cutoff frequency. When the Fourier
 112

Transformation of input signal ( ) is multiplied by H ( ) , the high frequency part

of ( ) becomes 0. In this way the high frequency part of ( ) is filtered, and
( ) only contains low frequency part.
The FFT noise reduction processing procedure in MATLAB is as Appendix G. The
result of FFT de-noise is shown in Figure 6-14:

1.5 1.5

1 1

0.5 0.5

0 0

-0.5 -0.5

-1 -1

-1.5 -1.5
0 500 1000 1500 2000 0 500 1000 1500 2000
t (second) t (second)

Figure 6-12 Noise reduction result

Figure 6-14 shows that the noise is significantly reduced, while the useful signal is kept.
However there is still noise at peak and valley of cosine signal.

6.3.2 Wavelet de-noise

6.3.2.1 Application of wavelet

The deduction process of wavelet is in Appendix H.

6.3.2.2 Applying wavelets to reduce noise for signal.

Noise reduction is one of main applications of wavelets analysis in signal processing.
Among orthogonal wavelet, as the selection of orthogonal basis is closer to practical
signal itself than conventional approach, it is possible to separate noise easily through
wavelets transform. The shortcoming of traditional noise-falling method is that it will
 113

increase entropy deduced after signal transformation resulting in non-stationary

property of signal can’t be described.
Steps in the process of one-dimensional signal reduces noise by using wavelets are
illustrated as shown below:
Signal wavelets
decomposition

Signal
preprocessing

Threshold quantization of the high-

frequency coefficient in processing of
wavelet transform

Wavelets
reconstruction

Noise reduction
completion

Figure 6-13 Steps in the process of one-dimensional signal reduces noise by using wavelets

Thus it can be seen that wavelets transform decomposes frequency band of signal into
low-frequency and high-frequency components, and then decomposing low-frequency
component into low-frequency and high-frequency components until reaching the
number of layers needed to be decomposed. Wavelets tree is as shown below:
 114

Figure 6-14 Signal decomposing using wavelet

Using the way of inhibiting detail coefficient with high frequency after the
decomposition to zero to eliminate noise. In the process of wavelets decomposition,
since approximation coefficient deduced from decomposition of each time is smoother
than before, and the high-frequency noise eliminated be kept in detail coefficient. The
goal of reducing noise can be reached as long as inhibiting detail coefficient to zero (it
is also possible to use the method of zooming out detail coefficient on this layer).
Specific steps are as follows:
1). Decomposing noisy signal to eight layers by using dB10 wavelets.
2). Extracting approximation coefficients of 8th layer as signal after noise reduction (if
using the way of zooming out detail coefficient, it needs to reconstruct detail coefficient
after inhibition and approximation coefficient on this layer to get signal after noise-
falling). Here using the way of inhibiting detail coefficient to zero, the result is shown
in Figure 6-17.
The wavemenu toolbox is the tool to process wavelet de-noise in MATLAB. When the
raw signal is imported, parameters in wavemenu such as type of wavelet, order number
of wavelet, and de-noise threshold can be setup, then the filter result can be pre-viewed.
 115

2 1.5

1.5
1

0.5
0.5

0 0

-0.5
-0.5

-1

-1
-1.5

-2 -1.5
0 0.5 1 1.5 2 0 0.5 1 1.5 2
t (second) t (second)

Figure 6-15 Noise reduction using wavelet decomposing

6.3.2.3 Discussion
Comparing Figure 6-14 and Figure 6-17 according to two principles of signal noise
reduction (brought out at the beginning of Section 6.3), it can be seen that the de-noised
curve from the wavelet is smoother. The standard deviation between noise-reduced
signal and original cosine signal using wavelets transform method is 1.2755, while
standard deviation using FFT transformation method is 2.5389. Therefore, from both
two principles of noise reduction, wavelets transform method has obvious advantages
over FFT transformation. Besides, wavelets transform method is more flexible, because
different operations can be done on different layers.

6.4 Conclusion
The methods and steps of data processing is discussed in this chapter. Firstly PCA is
used to correct coordinate. Then time-domain integration and frequency-domain
integration are compared according to the result of two integral of acceleration signal.
The result suggests that the frequency-domain integration is more efficient. At last two
de-noise methods: FFT and wavelet are discussed, in which wavelet has better effect
 116

than FFT. In Chapter 7, the data collected from my experiments will be processed with
PCA, frequency-domain integration and wavelet de-noise.
 117

Chapter 7. Data Processing and Discussion

In this chapter, I will perform analysis and processing to the acquired data of sensor
and discuss the result.

7.1 Data processing

The Tayside Road Bridge was sampled for many times in the experiment. Five sets of
data from the Tay Road Bridge were randomly chosen from all the data to be processed,
analysed and discussed. The data processing procedure is discussed in this section. The
result of Forth Road Bridge will be discussed in Section 7.4.
Data acquisition time: start from 20:59, January 3, 2013
File name: 20130103_205946.m
There are 893,314 points measured by acceleration sensor with the sampling rate of
about 200Hz.
There are 893,318 points measured by gyroscope with the sampling rate of about 200Hz.
There are 4,379 points measured by GPS, with the sampling rate of about 1Hz.
The data from acceleration sensor is [0, 4,391.718]s.
The data from GPS is [0, 4,379]s.
All five sets of data are acquired when the vehicleis running on Tay Road Bridge, in
which the second and fourth sets are acquired when the vehicle is running from north
to south, the other three are acquired when the vehicle is running from south to north.
There are about 30,000 points in each data set, including data of acceleration sensor
and gyroscope, with the sampling rate of about 200Hz.
Due to the quality of road surface, the acquired data contains large noises. So the
processing flow for the data is:
1. Select a proper piece of data generally with 30,000 or 35,000 points;
2. Conduct axis-correction to the data through PCA to make the sensor coordinate
parallel to the natural coordinate system;
3. As the waveform of raw data is similar to dB wavelet, the dB wavelet can be used to
de-noise data. Also from comparison experiments, it is known that dB wavelet is more
effective than other typical wavelet. In summary, dB8 wavelet is chosen for the de-
noise process;
4. The baseline of wavelet in the previous step is stored;
5. Get the baseline-drifted data by reducing the baseline from the de-noised data. This
is for the pothole detection with threshold value;
 118

6. Select a piece of GPS data and expand the data by interpolation so that the the length
of GPS data equals to the length of acceleration and gyroscope data. Use this data as
the x-axis for time (speed) correction;
7. Determine the pothole according to the threshold value.
The whole data process procedure can be divided to initialisation process, common
process and pothole detection process, as shown in Figure 7-1, 7-2 and 7-3.
The initialization is shown in Figure 7-1. Firstly the sample rate of sensor is set as 100
Hz or 200 Hz with a pre-setting threshold. After 5 seconds, the speed of vehicle is
detected: if it is between 50 and 70 miles/hour, the data collection is started; otherwise
the vehicle speed is detected in 5 seconds. The next step is pre-process: as sensors may
output bad points, this step is to remove those points. Next, a set of judgment rules are
applied to detect weather the vehicle is on smooth road surface. If so, the process goes
to a PCA (principle component analysis) to find vertical axis and calculate new
threshold of time domain. If the vehicle is not on a smooth road surface, the process
goes back to the second step, in which data collection restarts. The last step is wavelet.
Because the road and vehicle condition various in each test, an initialization is required
to obtain the value of parameters, which is shown in Figure 7-1. Firstly, data collected
in the first five seconds is used to test vehicle speed and road straightness. If vehicle is
between 50 and 70 miles per hour and road is straight and flat, the data is then processed
by PCA and wavelet de-noise. In de-noise process, the 120% of maximum value of
each level decomposed is used as new threshold.
The next step is the common process shown in Figure 7-2, in which the data is
decomposed by wavelet. The threshold obtained in Figure 7-1 is used here to filter the
data: all values smaller than threshold are set to zero. Then the rule of pothole detection
is applied.
The rule of pothole detection is shown in Figure 7-3. When M number of sequential
data points have value larger than the threshold, followed by N number of sequential
data points have value smaller than the threshold, the last one of M number of sequential
data points is considered as a pothole. Its position is recorded, and the threshold is
updated at the same time. Then the common process repeats until all data is processed.
 119

Pre-set threshold/sample rate

(100Hz/200Hz)

Data collection
(5S?)

Speed range:
50-70 mile/h?

Pre-process:
Remove bad points

Judgement rules:
1. Time domain: all values < threshold;
2. Frequency domain: energy < threshold
N exclude noise from engine;
Smooth road
surface? (Similar to white noise after low pass filtering)

PCA Get the maximum value * 120% =

Choose vertical axis New threshold of time domain

Wavelet
decompose

Get the maximum value of each level * 120%

= new threshold of wavelet decompose
For common process

Figure 7-1 Initialisation process

In the initialisation process (shown in Figure 7-1), the sample rate is set as 100 or 200
Hz (depends on device), then the data is being collected for 5 seconds repeatedly until
the speed of vehicle and condition of road satisfy experimental requirements: firstly the
speed of vehicle is checked, which should be between 50 and 70 miles/hour; and then
 120

the bad points from data are removed, then the data is used to check if the surface of
road is smooth. When both the speed of vehicle and condition of road satisfy the
experimental requirements, the initialisation process uses PCA to do axis-correction.
Then the wavelet analysis is used (decompose then compose) to find the threshold of
wavelet decompose which is then passed to common process, in which the data is
collected and processed to detect the potholes. When all these are finished, the
initialisation quits and the program switches to common process.
In the common process (shown in Figure 7-2), data is being collected and pre-processed
to remove bad points. Then PCA is used for axis-correction, and the threshold of
wavelet decompose passed from initialisation process is used by wavelet analysis for
de-noise. Then the pothole detection algorithm is being operated to find potholes. If
potholes are found, they are recorded with current GPS data. The common process runs
until experiment ends.
 121

Data collection

Pre-process Remove bad points

Axis correction PCA

Choose vertical axis

Wavelet Wavelet type/levels

decompose

Data filtering using pre-setting

threshold

Wavelet compose

Pothole distinguish Threshold adjustment

Mark coordinates with current

GPS data

Figure 7-2 Common process

Pothole detection algorithm (shown in Figure 7-3)is called inside the common process,
and is the core of whole pothole detection program. The pothole detection algorithm
 122

firstly reads and scan a piece of data which is already filtered and de-noised by the
common process. Then the data is compared with preset threshold which may be
dynamically changed. If the data > threshold, then I consider a pothole is found. The
position of this data is recorded. However if most of data is pothole according to the
value of threshold, then it is known that the value of threshold is too low and then
increased. After the value of threshold is changed, the current piece of data will be
processed again.
 123

Pre-set collection
Windows (2S)

Pointer starts from the first data

Scan

N Flat road surface? Judgement rule: value of N data <

threshold continuously

N Pothole? Judgement rules:

1. Next value of M data >
Y threshold continuously
2. Next value of N data <
threshold continuously
Pothole valid

Output position

Most value > threshold Error

Increase threshold
Scan again
Figure 7-3 Pothole detect process
 124

From the five sets of data being processed, one set is chosen as an example, which will
be called Route 1 in the following part of the section the other four sets (Route 2-5) are
attached in Appendix A.
The Route 1 data contains 30,000 points from point 105,001 to 135,000 of the raw data.
The acceleration data is shown in Figure 7-4.

20
Acc1X
Acc1Y
15 Acc1Z

10

-5

-10
0 0.5 1 1.5 2 2.5 3
Points 4
10

Figure 7-4 The acceleration data of Route 1

From the Figure 7-4, we can learn that the red part is Z-axis containing the measurement
of gravity. But as the smartphone is not strictly parallel to the ground, part of the gravity
is contained in the green part which means X-axis.
Axis-correction to the data with PCA:
[coeff,score,latent]=princomp(A);

'Pimcomp' is the function to process PCA in MATLAB. Its input is the original data

that is correlated. The output consists of three parameters, 'coeff', 'score' and 'latent'.

'coeff' is a 3*3 matrix which contains all the eigenvectors of the covariance matrix

corresponding to the original data set. This matrix is also called transformation matrix

or projection matrix.
 125

'score' contains three triaxial components, which are represented by three vectors

arranged by the contribution rate from large to small. As discussed, these three

components have no correlation. 'latent' is a one dimension column vector, in which

each element represents the contribution rate of the corresponding data components. As

there are three components in the data set, 'latent' has three elements, arranged from

large to small. Therefore the the original data set A*coeff = score. By using Primcomp

we can get the 'coeff':

coeff =

0.0395 -0.3348 0.9414

-0.7545 -0.6277 -0.1916

-0.6551 0.7027 0.2774

As shown in Figure 7-5, data of all three axes are de-noised with wavelet. In fact, only
data on Z-axis is necessary to distinguish the pothole, but for clarity data on three axes
are processed.
 126

20
Acc1XDenoise
Acc1YDenoise
Acc1ZDenoise
15

10

-5
0 0.5 1 1.5 2 2.5 3
Points 4
10

Figure 7-5 De-noise data on the three axes with wavelet of Route 1

From Figure 7-5, it is seen that due to the baseline drift, the threshold value is hard to
be determined. Thus it is necessary to eliminate the baseline. Firstly the baseline is
extracted from the result of wavelet decomposition, as shown in Figure 7-6.
 127

10

8 Acc1XBaseline
Acc1YBaseline
Acc1ZBaseline
6

-2

-4
0 0.5 1 1.5 2 2.5 3
Points 4
10

Figure 7-6 Baseline of acceleration data of Route 1

Then subtract the baseline from the de-noised data to get the result, as shown in Figure
7-7.
 128

8
Acc1XAC
Acc1YAC
6 Acc1ZAC

-2

-4

-6
0 0.5 1 1.5 2 2.5 3
Points 4
10

Figure 7-7 AC component of accelerometer data of Route 1

As can be seen from Figure 7-7 that, the red signal (Z-axis) is most sensitive to pothole,
as it contains the largest amplitude, and in the experiment the detection vehicle is
considered as in uniform linear motion and the largest vibration should be from the
pothole. This conforms the assumption that axis Z is approximately vertical to the
ground.
The next step is to calculate and store the position of potholes. The x-axis in Figure 7-
7 means number of data, to obtain the position of each pothole, the x-axis needs to be
transfer to distance. In order to calculate distance from number of data, the speed and
sampling rate is needed. The sampling rate is fixed by smartphone setting. The speed
of the vehicle, not as my assumption, is not exact constant, which explains vibrations
on x and y axes in Figure 7-7. Thus to record the position of potholes, the data from
GPS is used. The specific process is as follows:
Firstly time of beginning point and end point are found.
Acct(105001)=517.445 seconds;

Acct(135000)=665.296 seconds;
 129

It is seen that this piece of data is from No. 517s to No. 665s. A piece of 150 seconds
is selected from No. 517s to No. 666s, containing 150 seconds. The distance change in
this period can be accurately read from the GPS data package.
GPSdist2=Location_GPS_dist(517:666);

Set 0 as the initial value of distance:

GPSdist2=GPSdist2-GPSdist2(1);

As the point frequency of GPS is one every second, there are totally 150 points.
Interpolate the 150 points to 30,000.
xi=0.005:0.005:150;

x=0:149;

Dist2=interp1(x,GPSdist2,xi, 'spline');

Take the distance (0-2500m) as the abscissa:

8
Acc1XAC
6 Acc1YAC
Acc1ZAC

-2

-4

-6
0 500 1000 1500 2000 2500
Distance(m)

Figure 7-8 Use distance as the abscissa in Route 1

As seen from Figure 7-8, the processing result is very good with each pothole clear in
the final result and the flat road surface free from noise.
As a contrast, process the gyroscope data of Route1 in the same way:
 130

Voltage

Figure 7-9 Gyroscope data of Route 1

It is seen that there is no DC component of gravity acceleration in gyroscope data and

the data of three axes all fluctuates around zero.
Conduct PCA to the gyroscope:
[coeff,score,latent]=princomp(G);

coeff =

0.8657 -0.3956 0.3066

-0.2872 0.1090 0.9516

0.4099 0.9119 0.0193

Denoise the gyroscope data with dB8 wavelet:

 131

Voltage

Figure 7-10 Denoise the gyroscope data of Route 1

It is seen that there is still fluctuation in some part, so extract the baseline from result
of wavelet decomposition:
 132

Voltage

Figure 7-11 Baseline of gyroscope data of Route 1

Eliminate the baseline from the de-noised result:

 133

Voltage

Figure 7-12 AC component of gyroscope data of Route 1

Conduct distance correction with Dist2 as the abscissa, the result is as follows:
 134

0.2
Gyo1XAC
0.15 Gyo1YAC
Gyo1ZAC
0.1

0.05

-0.05

-0.1

-0.15

-0.2

-0.25
0 500 1000 1500 2000 2500
Distance(m)

Figure 7-13 Conduct distance correction with Dist2 as the abscissa

It is seen that gyroscope data can represent most of pothole data, but its signal-to-noise
ratio is not better than that of acceleration sensor.
Because Route 1 includes data points No. 105001 - No. 135000, from
Acct(105001)=517.445 seconds
Acct(135000)=665.296 seconds
It is known that Route 1 starts from 517th second of experiment, and ends at 666th
second. Therefore the data from 517th second to 666th second of GPS data is taken for
process:
Thd=Location_GPS_spd(517:666);
The GPS data can be drawn as Figure 7-14, in which x axis is time (measured by second)
and y axis is velocity (measured by meter per second).
 135

Figure 7-14 GPS speed data from 517th second to 666th second. X axis is time (second), y axis is
velocity (m/s)

There are 150 data points in the figure. For convenience of calculation, these 150 data
points are interpolated to 30,000 points. Because there are 30,000 data points for
acceleration data, the speed and acceleration data is now one-to-one correspondence:
Thd2=interp1(x,Thd,xi, 'spline');
Based on the model discussed in Chapter 5, equation 5-8, the threshold of pothole
judgment is proportional with square power of speed. Thus square calculation is applied
to every point in Figure 7-14 GPS speed data from 517th second to 666th second. X
axis is time (second), y axis is velocity (m/s):
Thd2=Thd2.^2;
A new curve is obtained through the square calculation, which is shown in Figure 7-15.
 136

Figure 7-15 Square calculation applied to every point GPS speed data. X axis is points, y axis is
square of velocity

The threshold is
Threshold=EmpCoef*Thd2;
In which EmpCoef is empirical coefficient. With tests, 1/600 is considered as a proper
value of EmpCoef, with which the most potholes are found, at the same time, the non-
pothole areas are not misjudged.
 137

Figure 7-16 Threshold and Acc1zDenoiseAC. X axis is points, y axis is amplitude (m/s²)

In Figure 7-16, it is shown that the first value higher than threshold happens at about
27500th point, which happens before vehicle reaches the surface of bridge. This is
because the speed of vehicle reduced to nearly zero when it passed the first expansion
joint, and with the low speed, the expansion joint resulted a less obvious peak valued
data (but it is still detected). Therefore the second peak valued data from right side
(because the vehicle drives from right to left side) in Figure 7-16 is actually the first
expansion joint on bridge.
From Figure 7-16, it is seen that an extra pothole is detected between the 4th and 5th
expansion joint. Field survey proved that there is a crack on the road surface between
the 4th and 5th expansion joint, which has similar effect of a real pothole. This result can
partly prove the effeciency of the pothole detection method.
The 37th and 39th expansion joints are failed to be detected. The 30th, 31st, 32nd and 33rd
expansion joints are also not detected. However the field survey proved that these four
expansion joints did not exist on bridge. Therefore the detection result for these four
expansion joints are correct.
To sum up, the detect rate is:
Misjudgment rate: 1/41=2.44%
 138

Omission rate: 2/41=4.88%

Accuracy: 38/41=92.68%
Note that data has a very small value at the first expansion joint, at which the speed of
vehicle is nearly zero. If the threshold is not proportional to the square of speed, either
a large threshold value is used, in which case the first expansion joint is omitted; or a
small threshold value is used, in which case, although the first expansion joint is
detected, some noise may be misjudged as potholes. Therefore it is effective to set
threshold as proportional to square of speed, and this increases the effeciency of
detection method under relatively low speed.
Through above data processing, most potholes can be detected precisely. In total, 42
potholes can be detected. According to the sequence from north to south, the potholes
are as follows:
 139

Table 7-1 Tay Road Bridge pothole result

Pothole No. Location(Uni Value of Amplitude

t 0.001m)
1 479428 3.493479389617304
2 503359 12.282863975315701
3 562333 3.414440299183515
4 601293 10.426649739276137
5 647251 2.504766783280768
6 702528 13.162327754843496
7 758365 1.223170039320437
8 814388 23.958198894239505
9 869185 7.528773067211463
10 922705 13.943086050762023
11 977700 0.873244979196262
12 1033416 1.229766842508431
13 1085459 26.066322240274541
14 1140805 11.044496681966052
15 1192367 4.805444052843307
16 1247661 19.422785688228405
17 1301204 33.911657251870253
18 1361839 17.904420663242227
19 1411051 6.143552087759477
20 1471183 10.213872255771888
21 1525120 3.697307425589718
22 1579967 17.816857735391114
23 1632241 18.858027414966717
24 1682090 0.043926201928237
25 1741562 53.848259283426117
26 1791860 13.443920524897299
27 1848963 7.365587601658930
28 1904895 1.572325588118663
29 1953478 9.567219474089635
30 N/A
 140

31 N/A
32 N/A
33 N/A
34 2244573 11.239200705316158
35
36 2353446 3.733065333508728
37 2408206 4.101902478160083
38
39
40 2568936 1.244674347299704
41 2621007 0.311282845814263
42 2669333 3.138815035002274

7.2 Comparison between actual measurement of Tay Road Bridge pothole

position and data processing result
In order to verify whether data processing result is right, writer conducts actual measure
to Dundee Tay Road Bridge, measuring location of every pothole and taking pictures
for width of every pothole. Measurement result is as follows.
 141

Table 7-2 Pothole ID and the interval distance between potholes

Pothole ID Interval Pothole ID Interval Pothole ID Interval

distance distance distance
(m) (m) (m)
D1 25.0 D15 54.0 D29 54.0
D2 26.0 D16 56.0 D30 ?
D3 32.0 D17 54.0 D31 148.0
D4 38.0 D18 56.0 D32 ?
D5 46.0 D19 54.0 D33 146.0
D6 56.0 D20 54.0 D34 56.0
D7 57.0 D21 56.0 D35 54.0
D8 54.0 D22 54.0 D36 54.0
D9 56.0 D23 54.0 D37 56.0
D10 54.0 D24 56.0 D38 54.0
D11 56.0 D25 54.0 D39 54.0
D12 54.0 D26 54.0 D40 56.0
D13 56.0 D27 56.0 D41 56.0
D14 54.0 D28 56.0 D42

Figure 7-17 Normal expansion joint style of No. 1-29 and 34-42
 142

Figure 7-18 Abnormal expansion joint style of No. 30, 31, 32, 33

According to actual measurement result, there are 42 expansion gaps totally. What
should be noted here is that vehicles drive different lanes on the journeys from south to
north and from north to south, thus there is difference of road surface condition, and it
can’t be regarded route1/3/5 as reverse order of route2/4. But location of pothole almost
should be overlapped.
Among these 42 potholes, the last one (D42) routinely can’t be detected because the
last pothole locates in the end of bridge, and then roundabout. Vehicles have to slow
down here to stop momentarily, and can’t pass until confirm no vehicles passing on the
right side of roundabout so that vehicles running speed has already been close to zero,
resulting in the data of the last pothole has lost its value. Among the rest of 41 potholes,
D30, D31, D32 and D33 has the lowest detectable rate. As we can see from actual
photographs that these four potholes are very special. Actually they have been made
asphalt perfusion treatment so that there almost is no difference between their surface
and well-paved road surface, so observer can’t find them unless he is very attentive. In
fact, these three small wave crests can be seen among acceleration data, which
evidences exactly the sensitivity of the algorithm is very high; they can’t be seen when
observing waveform in time domain. But it is possible to misjudge other well-paved
road surface as pothole if detection threshold is set lowly, thus the three expansion gaps
shouldn’t be counted that pothole has already been detected. Actually, these three
 143

expansion gaps with very small height shouldn’t be judged as pothole, otherwise this
algorithm will lost its meaning.
Except these four potholes, this algorithm has very high detectable rate for the rest of
potholes, thus achieving purpose of design.
Compare the results with the 3D model of the bridge, we can find out that they match
very well as expected. In Figure 7.19, the red lines are results from pothole detection,
the black lines are the real position of potholes. It is shown that, the red lines appear at
the same position of black lines, or appear near black lines, which means the potholes
are detected, also the position of detected potholes are close to their real position.

Figure 7-19 Tay Road Bridge result vs 3D model

7.3 Forth Road Bridge Data Processing result

The same method can be used with Forth Road Bridge, as shown in the following figure.
 144

Figure 7-20 Forth Road Bridge result vs 3D model

The blue line is the accelerometer's data, and the red line is the gyroscope's data. As
can be seen, most of these potholes can be identified out (more than 90 of these 100
potholes in total), but the S2 segment has several missing, which may be due to the
expansion joint of Forth Road Bridge is relatively shallow and hard to be recognized
well. on the other hand may be due to the arch style of the bridge which is not satisfied
with the condition defined in chapter 5.

7.4 Real pothole detect test

A road in Camperdown Country park was chosen as the test road. There are some
potholes on the road surface. The test date is 27/06/2016, and the test device is Samsung
S4/GT-9505 (OS: Android 5.0.1) smart mobilephone. The software is Sensor Insider
3.1.2. Use a Vauxhall Astra 2001 1.6L automatic transmission car as the test vehicle.
 145

Figure 7-21 Road for test in Camperdown Country park (shown as the red route)

The testing road is shown as the red route in Figure 7-21. The car runs from the north
side to the east side, and then return back to the start point.
Actual measurement result shows there are 11 potholes on the route. The distances
between each pothole and the start point shows as Table 7-3.
 146

Table 7-3 Distances between each pothole and the start point

ID Distance to the start point (m)

1 770
2 813
3 875
4 1080
5 1103
6 1140
7 1180
8 1780
9 1877
10 1930
11 1960

The measurement repeated 4 times.

From the algorithm process we can see that, the final processed signal is the Z axis data
after coordinate adjustment, so the X axis and Y axis data after coordinate adjustment
do not need to be process further more. So in the data processing, after the three axes
are corrected, only the z-axis data needs to be processed further more.
Some routes have two peaks at the beginning and end of the route, which are the
vibrations caused by the start of the vehicle and the gear shift, and are removed when
processing the data.
Those much closed data will be combined into one pothole.
Route1
 147

20
AccelerationX
AccelerationY
15 AccelerationZ

10

-5

-10

-15
0 0.5 1 1.5 2 2.5 3 3.5 4
Points 4
10
Figure 7-22 Three acceleration data of Route 1

The Route 1 data contains 36,700 points, 367 seconds. The maximum speed is 9.25m/s,
the average speed is 5.40m/s.
Firstly process the axis-correction using PCA, then de-noise using wavelet.
 148

12

11

10

5
De-noiseZ

4
0 0.5 1 1.5 2 2.5 3 3.5 4
Points 4
10
Figure 7-23 Wavelet de-noised result of Z axis of Route 1

Then the baseline of Z axis is extracted from the result of wavelet decomposition, as
shown in Figure 7-24. So we get the Z axis without baseline drift as shown in Figure
7-25. Let the threadhold as:
Thd=speed.^2/100;
The coefficient is different from the expansion joint detection. That’s because the test
cars are different.
And using distance as the x axis, get the final result as shown in Figure 7-26.
 149

Figure 7-24 Baseline of Z axis of Route 1

Figure 7-25 Z axis of Route without baseline

 150

-1

-2

-3

-4 AccZDenoise
Threadhold
-5
0 500 1000 1500 2000
Distance (m)
Figure 7-26 De-noised Z signal with threadhold. The x axis is distance

So we get the position of potholes.

 151

Table 7-4 Detection result of Route 1

ID Distance to the start point (m)

1 762.9
2 813
3 871
4 1072
5 Didn’t check out
6 1143
7 1182
8 1779
9 1877
10 1928
11 1960

There is 1 pothole missed, the accuracy is 10/11=91%.

For full details of other 3 routes processing, please refer to Appendix .
To sum up, the detection result of all 4 routes shown as Table 7-5. We can see that there
are 6 times missed, the total accuracy is 38/44=86.4%.
 152

Table 7-5 The detection result of all 4 routes

ID Distance to the start point (m)

Route 1 Route 2 Route 3 Route 4
1 762.9 758.6 762 755
2 813 Didn’t check out 811 Didn’t check out
3 871 869 873 863
4 1072 1078 1087 1086
5 Didn’t check out 1102 1103 1103
6 1143 1149 1152 1152
7 1182 Didn’t check out 1192 Didn’t check out
8 1779 Didn’t check out 1787 Didn’t check out
9 1877 1882 1889 1887
10 1928 1930 1936 1923
11 1960 1958 1966 1935

7.5 Discussion
Seen from the above results of data processing, the original data is all with loud noise
covering part or all of useful data. It is mainly because that when the vehicle is running,
the sensor in the smartphone can not only detect signals of vibration from the vehicle
caused by pothole but also noise from the road and noise caused by driver of the vehicle
including engine and gearbox. The mix signals make it hard to detect the pothole signals
in time domain. After wavelet denoising, the quality of signals becomes very good but
with huge fluctuation in low frequency. The baseline drift may be caused by
comprehensive reasons in many aspects in which the major possibility is the component
of acceleration in the front-rear direction on axis Z of acceleration sensor during
acceleration or deceleration. As three axes of the sensor in smartphone are not exactly
parallel to that of the earth, it is hard to totally avoid the gentle baseline variation. This
kind of baseline drift will make the threshold value used for detecting the pothole hard
to be determined, so eliminating the baseline is very essential after denoising. In this
experiment, the reusing of baseline after wavelet decomposition can not only achieve
good result but also take advantage of the result in last step, which improves the
processing efficiency. As the speed of vehicle is not a constant value in the experiment,
 153

the accurate distance cannot be obtained by multiplying starting speed with experiment
duration. The proper method is integral speed over time. Because GPS measures and
stores the speed of vehicle up to once per second, the GPS data package is used in this
study in order get accurate distance value from vehicle speed. The accurate distance
value will help in obtaining accurate position of detected pothole. So here we use the
speed provided by GPS data package. In the chapter for sensor, we have discussed it in
detail and come to the conclusion: As the speed value directly adopts the calculation
result of Doppler Effect, it is reliable with high accuracy and stability. Therefore the
experimental result we get is accurate.
 154

Chapter 8. Conclusion and Future Work

8.1 Conclusion and Prospects

8.1.1 Conclusion
In the Chapter 1, five questions are brought out:
1. In which way the traditional methods can be improved? Is there a new method that
can exceed the traditional methods?
2. What equipments are required?
3. Which physical measurements should be kept?
4. What algorithm is required to analyse the data and obtain the result?
5. The new method should be proved to meet our goal.
With previous discussions in this thesis, all five questions can be answered:
1. In the Chapter 3 a new pothole detect method is carried out (for detail please refer
to the following paragraph), comparing to the traditional methods, it is more
efficient and has lower cost. Thus this new method is valuable and has good
prospects;
2. As described in the Chapter 3, in this mew method a smart-phone with built-in
acceleration sensor is required to be fixed in a vehicle. Especially, when a large
number of vehicles have smart-phones with built-in acceleration sensors installed,
the efficiency and accuracy of detection will be better than the traditional methods;
3. The new method will measure the acceleration of the vehicle while the vehicle is
moving along the road;
4. The acceleration data is analysed by axis transformation, de-noise processing based
on wavelet transform, please refer to the Chapter 3-6 for detailed description.
5. The smart-phone can read the acceleration data from its sensor do the analysis. In
this way the pothole is detected. The whole process can be done by installing the
app on smart-phone which is now very popular. In the Chapter 6, the new method
is tested on the Tay Road Bridge. The result proved that, the new method can
efficiently detect the potholes on the road. Therefore the new method does meet our
goals.
In this paper, a dynamic threshold based on the instantaneous speed is proposed for the
first time, which can improve the accuracy of detection, and experiments show high
accuracy. The main problems encountered at this stage, is the need for a large number
of experiments, through the acquisition and processing of different vehicles and
 155

different traffic data, to obtain a large number of experimental data, to summarize the
judgment threshold conditions under various conditions, in order to improve the
detection of universality. In addition, it will be helpful to fuse the acceleration sensor
data and the gyroscope data together to improve the stability and reliability of the
pothole detection.

8.1.2 Future work and considerations

The following progresses are worthy of note.
1. Due to time constraints, app used herein is from the third party, it can only collect
sensor data and stored in the phone's functions, we need to import data into the
phone after post-processing of the data is pending completion of the experiment the
computer to complete in the laboratory. So in the future a new app should be
designed and implemented in order to achieve a full-fledged app all the features we
need, these features include: collecting sensor data, process data and send the results
to the server through wireless network (such as 3G/4G/Wi-Fi).
2. Design and implement server-side software functions, which should include:
receiving mobile client app sent to deal with complete data and store in database,
and compare with the road network data and uploaded data with other mobile phone
to ensure the credibility of the data, and to notify the user by some easy ways.
3. Process the data in order to obtain a result of this step is completed either on the
phone side, or can be placed on the server side, both have their advantages and
disadvantages, from the care of the user's point of view, it should work to minimize
the mobile phone amount, so the large amount of data processing work on the server
side may be reasonable, but it should also take into account the load capacity of the
server.
4. The threshold used in pothole detection should not be fixed, it should be related
with the vehicle model. Therefore, the vehicle model should either can be input
manually by the user, or mobile phone should be able to obtain these parameters
through ODB-II port or some other vehicle communication interface automatically.
As discussed in Chapter 7, comparing the detection result with real pothole data of the
Tay Road Bridge, it shows that the pothole detection on smartphone has good accuracy.
Comparing with traditional detection methods in Chapter 2, this new method has
advantage of high efficiency, high detection speed and low cost. In addition it is purely
 156

automatic for the smartphone user. This method is suitable for promoting for large-
scale application with proper operating design.

8.2 Design of Operating Mode

The design has two goals to achieve: 1.Covering road segment as much as possible
(wherever there is road segment opened for vehicles, there is also data); 2. Refreshing
data quickly (whenever there is new information of pothole, there should be reflected
in the result in first time); the ideal situation would be that all the vehicles driving on
the road is providing data for our system which means all vehicles is installed
smartphone collecting and sending data. Of course, this is only an extreme ideal
condition, the precondition must be there are enough vehicles as “inspection vehicle”
if wanting the system run in high efficiency and get favorable effect. Thus, one issue
that appears when considering the operating mode is: how to attract customers to use
this software and make them willing to provide data for us?
First, what can be sure is that this software will not bring too much negative effects and
burdens. Today, almost all OS of smartphone has the ability of running multitasks in
parallel, therefore such software program without any running interface (start-up
interface and control interface are certainly possible, but simplified control can be made
completely) but only can be run in the background hardly causes any interference to the
operation of the user, moreover for single driver, manipulating mobile phone in the
process of driving originally should have been forbidden. The only thing may be
affected is that the program will consume more power when running in the background,
and invoke all sensor, control transmission of 3G module, receive power from data,
increasing power consumption of mobile phone and shorting stand-by time. As for
extent of affection, different result is from the different mobile phone. Further analysis
is as follows:
1. CPU: nowadays, the CPU of the vast majority of smartphone, based on ARM kernel,
has automatic power dissipation control functions and optimization function, and power
dissipation of CPU can be adjusted while operating on the basis of load. This program,
part of light load (sampling is only 100-150times per second, and there hardly is other
treatments), has little effect on power dissipation; CPU of many new smartphone has
two kernels, of which one is used for handling high performance computing and anther
is used for bearing few calculations at very low power consumption. Since this
 157

framework is able to further reduce power consumption, so it is very suitable for this
program.
2. Transducer: today, almost power consumption of all acceleration sensor, gyroscope
sensor is very low, which is mainly decided by their design proposal and processing
technic.
3. Wireless communication module: this module is comparatively current consuming
when working, and the current routinely is 6mA/12V. While since the data transmission
speed of 3G/4G is so fast that we can store the collected data in memory as temporary
storage and sends it out when accumulating to certain volume. With data volume of one
hour, for example, saving 60X60X150 acceleration sensors, 60X60X150 gyroscopes,
60X60 GPS/speeds and data (4338000 data points totally) will takes space of 16MB
totally on the basis of each data takes 4 bytes; it will takes 16000X8/200=640S (about
11 minutes) to complete uploading work on the basis of each data takes 4 bytes, with
upload speed of 200kb, which needs to consume little power. If applying data
compression technic or using 4G network to transmit in the future, the power
consumption can be reduced further.
4. GPS receiver: generally, GPS is comparatively current consuming, while the power
consumption of new GPS has already been improved.
But it is noticed that many users would make SATNAV by using smartphone in the
long journey, which requires users to activate GPS module installed in smartphone itself,
thus this program will not increase extra power consumption, and only needs GPS
module to transmit measured data to STANAV software and itself.
In a word, when using smartphone, according to statistics, about 30%-40% power
consumption can be displayed on screen, and about 30% power is consumed on
connection of internet application, and the rest 1/3 is used for other software and
processes.
Besides, most of the vehicles is equipped with vehicle-mounted charger (elicited by
cigarette lighter or USB interface) so as to charge for mobile phone at any moments,
and drivers also pay more attention when making SATNAV which is application with
high power consumption by using mobile phone, so this is not a problem at all.
While just no harm is not enough. What need to do is to encourage users to install and
use the software and make users benefit from use. The concrete implementation plan,
of course, also need to be further planned, but since the system can achieve high-
precision pothole maps, the driver should be informed of such information, from which
 158

driver can benefit. When software, for example, perceives that the vehicle will soon
cross a known area ahead where may have pothole, which can be broadcasted with the
voice prompt. Software also can prompt the driver to pay attention to the road
conditions and further broadcast the size or severity of the possible pothole so that the
driver can take avoidance or deceleration measures timely. As long as the data is
sufficient and valid and the rate of false alarm rate is low enough, I believe there will
be a lot of users would use it. Meanwhile as the number of users increases, it will, in
turn, promote the collected data further more detailed and timely, so the result will be
more accurate. This will form a virtuous circulation.
Software can also add other functions to attract users, for example, voice broadcast
warning information of an accident in front of the road, or the weather alerts, the wet
and slippery road.
There is a type of important users who own moderate-load vehicles or heavy-load
vehicles, and the most typical representative is the operating vehicle, including taxi,
freight car, bus, etc. Running time of these vehicles on the road is much longer than
that of general private cars, thus they can provide more data. Cooperation agreement
can be reached with these vehicles’ operating company to install this system on the
vehicles, thus the efficiency of collecting data will be much higher. Certainly, the
running time of taxi on the highway is less, and other vehicles mostly belonging to
medium or large vehicles whose frequency responses are different from those of small
cars. Proprietary software can be developed for specific models, and specific model
parameters will be written fixedly to software in the future. Because vehicle models can
match with collecting equipment, therefore, it is feasible to do so. Meanwhile, this
method is provided with advantages of low costs but significant benefits.
This paper focus on the academic category such as technology and the algorithm of the
project, while with respect to business model and operation mode the paper just
conducted the preliminary design and discussion. Specific idea on this aspect can be
cooperatively studied and implemented with relevant experts in the future.
So we can assume some future development trend:

1. Software will replace hardware;

2. Generality will replace specificity;
3. Service will replace purchase;
4. The cloud mode will replace local mode;
 159

5. Big data integration will replace stand-alone device;

6. Free will replace payment (to the final customers);
7. Intelligent will replace non-intelligence.

8.3 Future works on real pothole detection methods

In this thesis, firstly the feasibility of pothole detection using smartphone is discussed.
Then expansion joint is chosen as simplified pothole. The vehicle runs through
expansion joint, and data is gathered by smartphone at the same time. Then various
vehicle models are discussed, based on which data is analysed. The data analysis gives
initial result. There are more work to be done in future, including:
1. Many parameters used in signal processing, including threshold of each order of
wavelet filter and threshold of pothole judgment, are related to parameter of individual
vehicle (such as suspension, tyre, etc.), which are obviously different for different
vehicles. The algorithm should be able to calculate vehicle parameters by analysing
sensor data, and then use these parameters to modify thresholds;
2. In this study, the smartphone is considered fixed on vehicle. However the smartphone
may be anywhere in the vehicle, such as inside pocket, in real life. In addition, the
smartphone may vibrate in different frequency from the vehicle if it is not properly
fixed. To avoid the error, a better axis correction method needs to be considered;
3. Roads in real life are not all straight and flat. The accuracy of axes transformation is
easily affected by turnings and slopes. In order to enable the detection method to deal
with turnings and slopes, two improvements are required. One improvement is to
continuously detect the direction of gravity to know whether the downward direction
of vehicle is changed during driving. The other one is to calculate rotation of vehicle
by integrate data gathered by gyroscope (which is angular velocity), so that change of
the road direction including turnings and slopes can be mapped to three axes.
4. Some widely used wavelet basis and filter methods are used in this study due to limit
of time. Although they are proved to be effective, a better result may be achieved with
other wavelets. Also other filter methods may lead to higher signal-noise ratio or a
decreased amount of calculation. More experiments are needed to get a balance between
accuracy and calculation speed;
5. The data used in this study is gathered by third party applications, and processed by
MATLAB. In future, an application should be designed to implement the data
collection and processing functionality. The design and implementation of this
 160

application should be carefully planned, including prephase investigation, requirement

analysis, sub-system division, implementation, version control etc. Apart from that, in
order to test and optimise the application, the value of parameters used in signal
processing, including threshold of each order of wavelet filter and threshold of pothole
judgment should be able to be modified during the running of application. Also the
application should be able to display intermediate data to developers;
6. The data collection and processing application is part of whole pothole detection
system. The result should be uploaded to a server, in which the data from all smartphone
is stored and analysed, so that the wrong data can be filtered to provide more accurate
result.
 161

Reference
1. Aleksander Hać, Iljoong Youn. Optimal Semi-Active Suspension with Preview Based on a Quarter
Car Model[J]. Journal of Vibration & Acoustics, 1992, 114(1):433 - 438.
2. Acosta, J. Adolfo, J. Ludwig Figueroa, and Robert L. Mullen. "Low-cost video image processing
system for evaluating pavement surface distress." Transportation research record 1348 (1992).
3. Andrew, Andrew AS, and B. R. Colford. "Forth Road Bridge-maintenance challenges."
In Advances in Cable-Supported Bridges: Selected Papers, 5th International Cable-Supported
Bridge Operator's Conference, New York City, 28-29 August, 2006, p. 57. CRC Press, 2006.
4. Bakker, Egbert, Lars Nyborg, and Hans B. Pacejka. Tyre modelling for use in vehicle dynamics
studies. No. 870421. SAE Technical Paper, 1987.
5. Barone, Martin R. Impact vibrations of rolling tires. No. 770612. SAE Technical Paper, 1977.
6. Bhoraskar, Ravi, Nagamanoj Vankadhara, Bhaskaran Raman, and Purushottam Kulkarni.
"Wolverine: Traffic and road condition estimation using smartphone sensors." In Communication
Systems and Networks (COMSNETS), 2012 Fourth International Conference on, pp. 1-6. IEEE,
2012.
7. Bujari, Armir, Bogdan Licar, and Claudio E. Palazzi. "Movement pattern recognition through
smartphone's accelerometer." In Consumer communications and networking conference (CCNC),
2012 IEEE, pp. 502-506. IEEE, 2012.
8. Cadenas, Alejandro, José María González, and Oscar M. Solá. "Context-aware processing of
mobile device sensor information: Car accident detection hosted service." In Consumer
Communications and Networking Conference, 2009. CCNC 2009. 6th IEEE, pp. 1-5. IEEE, 2009.
9. Carter, Matt, Steve Kite, Naeem Hussain, Alan Seywright, Mike Glover, and Billy Minto. "Forth
Replacement Crossing: Scheme design of the bridge." InIABSE Symposium Report, vol. 96, no. 6,
pp. 107-116. International Association for Bridge and Structural Engineering, 2009.
10. Cheng, Heng-Da, and Mario Miyojim. "Automatic pavement distress detection
system." Information Sciences 108, no. 1 (1998): 219-240.
11. Chen, Jia, et al. "Fatigue detection based on facial images processed by difference
algorithm." Biomedical Engineering (BioMed), 2017 13th IASTED International Conference on.
IEEE, 2017.
12. Chen Long, Zhang Xiaoliang, Nie Jiamei and Wang Ruochen. "Performance Analysis of Two-stage
Series-connected Inerter-spring-damper Suspension Based on Half-car Model." Journal of
Mechanical Engineering, no. 6 (2012):102-108.
13. Colford, Barry R. "Forth Road Bridge—maintenance and remedial works."Proceedings of the ICE-
Bridge Engineering 161, no. 3 (2008): 125-132.
14. Dupont, Edmond M., Carl A. Moore, Emmanuel G. Collins Jr, and Eric Coyle. "Frequency
response method for terrain classification in autonomous ground vehicles." Autonomous Robots 24,
no. 4 (2008): 337-347.
15. ElMadany, Mohamed M. "Control and evaluation of slow-active suspensions with preview for a
full car." Mathematical Problems in Engineering 2012 (2012).
16. Eriksson, Jakob, Lewis Girod, Bret Hull, Ryan Newton, Samuel Madden, and Hari Balakrishnan.
"The pothole patrol: using a mobile sensor network for road surface monitoring." In Proceedings of
the 6th international conference on Mobile systems, applications, and services, pp. 29-39. ACM,
2008.
17. Fan, ZHENG, YU Jiang, PEI Yi-jian, and Xiao Min. "The Co-simulation of mobile robot
surmounting obstacles based on ADAMS-MATLAB."Journal of Yunnan University (Natural
Sciences Edition) (2007): S2.
18. Fukuhara, Toshihiko, Keiji Terada, Makoto Nagao, Atsushi Kasahara, and Shigeki Ichihashi.
"Automatic pavement-distress-survey system." Journal of Transportation Engineering 116, no. 3
(1990): 280-286.
19. Furuichi, Takahisa, and Hideo Sakai. Dynamic cornering properties of tires. No. 780169. SAE
Technical Paper, 1978.
20. Gao, Wei, Nong Zhang, and H. P. Du. "A half-car model for dynamic analysis of vehicles with
random parameters." In 5th Australasian Congress on Applied Mechanics (ACAM 2007), vol. 1, pp.
595-600. Engineers Australia, 2007.
21. GAO, Zhen-hai, Nan-ning ZHENG, and Hong Cheng. "Soft sensor of vehicle state based on
vehicle dynamics and kalman filter." Acta Simulata Systematica Sinica 1 (2004): 007.
22. Georgopoulos, A., A. Loizos, and A. Flouda. "Digital image processing as a tool for pavement
distress evaluation." ISPRS Journal of Photogrammetry and Remote Sensing 50, no. 1 (1995): 23-
33.
 162

23. GUAN, Xin, Su-min ZHANG, Zhen-hai GAO, and Jun ZHAN. "Vehicle State Estimation Using
Automobile Dynamic Model with 7DOFs." Science Technology and Engineering 16 (2010): 017.
24. Guangqiang, Wu, and Fang Yuan. "T ime-domain simulation and analysis of vehicle ride
comfort." Automobile Technology 2 (2007): 8-11.
25. Hesami, Reyhaneh, and Kerry J. McManus. "Signal processing approach to road roughness analysis
and measurement." In TENCON 2009-2009 IEEE Region 10 Conference, pp. 1-6. IEEE, 2009.
26. Hou, Xiangshen, Hua Wang, Qi Wang, and Zheren Wang. "Development of automated inspection
system for highway surface distress." In Fundamental Problems of Optoelectronics and
Microelectronics III, pp. 65951Y-65951Y. International Society for Optics and Photonics, 2007.
27. Huan-yao, D. A. I. "Application and Study of Matlab on Time-frequency Analysis of Signal
[J]." Modern Machinery 5 (2005): 000.
28. Hunt, H. E. M. "Modelling of road vehicles for calculation of traffic-induced ground vibration as a
random process." Journal of Sound and Vibration 144, no. 1 (1991): 41-51.
29. Hunt, H. E. M. "Stochastic modelling of traffic-induced ground vibration." Journal of Sound and
Vibration 144, no. 1 (1991): 53-70.
30. Huston, Dryver R., Noel V. Pelczarski, Brian Esser, and Kenneth R. Maser. "Damage detection in
roadways with ground penetrating radar." In 8th International Conference on Ground Penetrating
Radar, pp. 91-94. International Society for Optics and Photonics, 2000.
31. Jiang Liangwei, Yao Lingkan, Wu Wei. "Study on Calculation of Dynamic Displacement from
Time Integration of Acceleration in Shaking Table Model Tests of Side Slope." Journal of Disaster
Prevention and Mitigation Engineering, 29 no. 3 (2009): 261-266.
32. Jiali, Wang, Jiang Li, and Liu Hong. "Auto static calibration of multi-axis force sensor based on
triaxial accelerometer." Chinese Journal of Scientific Instrument 29, no. 2 (2008): 432.
33. LI, Wen-liang, Wei ZHOU, Zhi-ping GUO, and Kan ZHAO. "Safety Dynamic Simulation Study on
Car Impact against the High Curb." Tractor & Farm Transporter 2 (2008): 010.
34. Lin, Jin, and Yayu Liu. "Potholes detection based on SVM in the pavement distress image."
In Distributed Computing and Applications to Business Engineering and Science (DCABES), 2010
Ninth International Symposium on, pp. 544-547. IEEE, 2010.
35. Lin, Paul P., Maosheng Ye, and Kuo-Ming Lee. "Intelligent observer-based road surface condition
detection and identification." In Systems, Man and Cybernetics, 2008. SMC 2008. IEEE
International Conference on, pp. 2465-2470. IEEE, 2008.
36. LU, Zheng, Hai-lin YAO, and Zhi HU. "Dynamic response analysis of rough pavement under
vehicle-road system coupled vibration." Chinese Journal of Geotechnical Engineering (2013): S1.
37. Mathur, Suhas, Tong Jin, Nikhil Kasturirangan, Janani Chandrasekaran, Wenzhi Xue, Marco
Gruteser, and Wade Trappe. "Parknet: drive-by sensing of road-side parking statistics."
In Proceedings of the 8th international conference on Mobile systems, applications, and services,
pp. 123-136. ACM, 2010.
38. Mednis, Artis, Girts Strazdins, Reinholds Zviedris, Georgijs Kanonirs, and Leo Selavo. "Real time
pothole detection using android smartphones with accelerometers." In Distributed Computing in
Sensor Systems and Workshops (DCOSS), 2011 International Conference on, pp. 1-6. IEEE, 2011.
39. Mohajeri, M. Jerry H., and Patrick J. Manning. "ARIA (TRADEMARK): AN OPERATING
SYSTEM OF PAVEMENT DISTRESS DIAGNOSIS BY IMAGE
PROCESSING." Transportation Research Record 1311 (1991).
40. Mohan, Prashanth, Venkata N. Padmanabhan, and Ramachandran Ramjee. "Nericell: rich
monitoring of road and traffic conditions using mobile smartphones." In Proceedings of the 6th
ACM conference on Embedded network sensor systems, pp. 323-336. ACM, 2008.
41. Monti, Max. "Large-area laser scanner with holographic detector optics for real-time recognition of
cracks in road surfaces." Optical Engineering 34, no. 7 (1995): 2017-2023.
42. Nejad, Fereidoon Moghadas, and Hamzeh Zakeri. "A comparison of multi-resolution methods for
detection and isolation of pavement distress." Expert Systems with Applications 38, no. 3 (2011):
2857-2872.
43. Pacejka, Hans B. "Tyre factors and vehicle handling." International journal of vehicle design 1, no.
1 (1979): 1-23.
44. Perttunen, Mikko, Oleksiy Mazhelis, Fengyu Cong, Mikko Kauppila, Teemu Leppänen, Jouni
Kantola, Jussi Collin et al. "Distributed road surface condition monitoring using mobile phones."
In Ubiquitous Intelligence and Computing, pp. 64-78. Springer Berlin Heidelberg, 2011.
45. Promwongsa, Nattakorn, Pramookh Chaisatsilp, Supawat Supakwong, Chalermpol Saiprasert,
Thunyasit Pholprasit, and Passakon Prathombutr. "Automatic accelerometer reorientation for
driving event detection using smartphone." In 13th ITS Asia Pacific Forum, Auckland, New
Zealand. 2014.
 163

46. Ravi, Nishkam, Nikhil Dandekar, Preetham Mysore, and Michael L. Littman. "Activity recognition
from accelerometer data." In AAAI, vol. 5, pp. 1541-1546. 2005.
47. RIBEIRO, Josh GT, Jaime TP DE CASTRO, and Josh LF FREIRE. "Filtering in frequency domain
to avoid time aliasing." In SPIE proceedings series. Society of Photo-Optical Instrumentation
Engineers, 2002.
48. Sayers, Michael W. "The little book of profiling: basic information about measuring and
interpreting road profiles." (1998).
49. Scavuzzo, Richard W., Timothy R. Richards, and L. T. Charek. "Tire vibration modes and effects
on vehicle ride quality." Tire Science and Technology 21, no. 1 (1993): 23-39.
50. Strazdins, Girts, Artis Mednis, Georgijs Kanonirs, Reinholds Zviedris, and Leo Selavo. "Towards
vehicular sensor networks with android smartphones for road surface monitoring." In 2nd
international workshop on networks of cooperating objects, Chicago, USA. 2011.
51. Tai, Yu-chin, Cheng-wei Chan, and Jane Yung-jen Hsu. "Automatic road anomaly detection using
smart mobile device." In conference on technologies and applications of artificial intelligence,
Hsinchu, Taiwan. 2010.
52. Takayama, Masahiro, and Koichi Yamagishi. "Simulation model of tire vibration." Tire Science
and Technology 11, no. 1 (1983): 38-49.
53. Tao, Yin, et al. "Study of a new amblyopia diagnostic and therapeutic method along with the
system implementation." Biomedical Engineering (BioMed), 2017 13th IASTED International
Conference on. IEEE, 2017.
54. Wang, Kelvin CP, and Omar Smadi. "Automated imaging technologies for pavement distress
surveys." Transportation Research E-Circular E-C156 (2011).
55. WANG, Qian-ting, Qi-zhong YI, and Yong-nan LIN. "Relation Between Off-road Terrain
Unevenness and Wheel Jumping Estimation [J]." Journal of System Simulation 11 (2008): 061.
56. Wei, Liu, T. F. Fwa, and Zhao Zhe. "Pavement roughness analysis using wavelet theory." In In
Proceedings from the 6th International Conference on Managing Pavements. Queensland,
Australia. 2002.
57. Whiffin, A. C., and D. R. Leonard. A survey of traffic induced vibrations. No. Lr 418. 1971.
58. Wu, yanping, Dongyan Wu, Xu Xiem and Xu Zhang. "Response and vibration control of steel box-
girder bridge when vehicles passing through the expansion joint [J]." Noise and Vibration Control.
2 (2013): 021.
59. Xiaoya, Chen Weizhen Wang Bingwen Hu. "Acceleration signal processing by aumerical
integration [J]." Journal of Huazhong University of Science and Technology (Nature Science
Edition) 1 (2010): 000.
60. XIE, Xu, Dong-yan WU, Jian-feng WANG, Shi-qi ZHANG, and Yong-jie ZHOU. "Dynamical
behavior of steel box girder bridges due to vehicle-induced vibration at expansion joint
[J]." Journal of Zhejiang University (Engineering Science) 10 (2009): 029.
61. Xu, Zhi, Kun Bai, and Sencun Zhu. "Taplogger: Inferring user inputs on smartphone touchscreens
using on-board motion sensors." In Proceedings of the fifth ACM conference on Security and
Privacy in Wireless and Mobile Networks, pp. 113-124. ACM, 2012.
62. Yan, A. Z., and H. Y. Zhou. "Mathematical model of the step road and analysis on the response of
two DOF vehicle model." Journal of Henan Polytechnic University(Natural Science) 30, no. 1
(2011): 113-117.
63. Yang Xiaofeng, Yang, Li Peng, Zhang Xin, Bian Zhengzhong, and Wang Bo. "De-Noising of the
doppler fetal heart rate signal with wavelet threshold filtering based on spatial correlation."
In Bioinformatics and Biomedical Engineering, 2007. ICBBE 2007. The 1st International
Conference on, pp. 928-931. IEEE, 2007.
64. ZHANG, Hong-liang, H. U. Chang-shun, and G. A. O. Jiang-ping f Education. "Parameters'
Influence on Determination of Allowable Differential Settlement between Bridge and Approach
Embankment." Journal of Highway and Transportation Research and Development 3 (2005): 012.
65. Zhang, Hong-Liang. "Determination of allowable differential settlement between bridge abutment
and approach embankment with five-degree-of-freedom vehicle model." International Journal of
Pavement Research and Technology 3, no. 6 (2010): 311-319.
 164

Appendix A: Detailed deduction process of bounce height h without considering

energy loss in tyre simulation
In Section 5.2, after the collision, the velocity has changed its direction form heading
to the right horizontally to v  , with the same magnitude, i.e. v  v . At this point, the

direction of v  is tangent to the circumference. That is to say, v  has a component of

vx horizontally, and a component of vy vertically, vertically upward. Since vx and

vy are perpendicular; the included angle  formed by v  with x axis is equal to

 AOB , i.e. angle  . Therefore, the two triangles shown in the figure are similar. It
can be concluded that
R v'
 (A.1)
CB v y '

Since C is located in the middle of AB,

AB w
CB   (A.2)
2 2
Put in the formulas above and the following can be obtained
v'w
vy '  (A.3)
2R
Since only magnitude of vy is required and v  v , therefore

vw
vy  (A.4)
2R
According to the law of conservation of energy, the gross energy before motion shall
be equal to that after motion, that is to say, the sum of the kinetic energy and potential
energy shall be conservative before and after motion:
1 2 1
mv1  mgh1  mv22  mgh2 (A.5)
2 2
In case of taking no account of the energy loss, before the tyre bounces vertically, it has
an upward initial velocity equalling to vy and at the same time, the potential energy is

assumed as 0. After the tyre bounces, the kinetic energy converts into potential energy;
the vertical velocity at the highest point is 0, namely the kinetic energy is also 0, so that
the kinetic energy fully converts into potential energy. Therefore,
1
mvy 2  mgh (A.6)
2
So the bouncing height is
 165

v y '2
h (A.7)
2g
Therefore
v 2 w2
h (A.8)
8 gR 2
 166

Appendix B: Detailed deduction process of bounce height h with considering

energy loss in tyre simulation

In Section 5.2, when considering energy loss, therefore

v '  Cr  v  v (B.1)

That is to say, the initial velocity is smaller than that of the incidence. Assume the
1 2 1
kinetic energy of incidence is mv , after rebounding, it is mv '2 , and both the
2 2
potential energy of incidence and bouncing are both 0, therefore, the kinetic energy is
equal to the mechanical energy, and therefore, the energy lost is
1 2 1 1
k mv  mv '2  m(1  Cr 2 )v 2 (B.2)
2 2 2
After collision, along with the rising height of tyre, the vertical kinetic energy
completely converts into potential energy after rebounding, therefore
v '2 w 2 C r 2 v 2 w 2
h  (B.3)
8 gR 2 8 gR 2
 167

Appendix C: Deduction process of PCA

In Section 6.1, let

Y T  X TW
 V  T W TW (C.1)
VT

When m<n-1, generally V is not uniquely defined but Y is. W is an orthogonal matrix,
YT is the transposition of XT, and the first row of which is made up of the first principle
component, the second row is made up of the second principle component, and so forth.
To obtain an effective method to reduce data dimension, we could map X to a low
dimensional space only using the former L pieces of vectors.
Y  WLT X   L V T where  L
(C.2)
 I Lm  with I Lm the L  m rectangular identity matrix

W, as a single vector matrix of X, is also equivalent to the eigen vectors C=XXT of

covariance matrix,
XX T  W  TW T (C.3)
 168

Appendix D: Time domain integral method

In Section 6.2.1, the velocity is obtained after the first single integration of acceleration
data
v(t )   a(t )dt   ( f (t )  T )dt   f (t )dt  Tt  X (D.1)

The displacement is obtained in the second integral:

s(t )   v(t )dt   [  f (t )dt ]dt  0.5Tt 2  X t   (D.2)

The derivation of equations 6-6 and 6-7 shows that error term increases as time of
integral increases, which cause large error in integral result. Therefore the acceleration
signal needs to be de-noised and the DC component needs to be removed before
acceleration signal is integrated in the time-domain. Generally, the DC component of
the signal is estimated by mathematical expectation (mean value), i.e.
X
1 i
a
N
a
i0
(D.3)

After the DC component is removed, the acceleration signal turns to:

a  ai  a (D.4)

Then Simpson integral is used:

t N
v  (ai2  4ai1  ai)
6 i2
(D.5)
 169

Appendix E: Frequency domain integral method

In the process of frequency-domain integral, the signal is processed by Fourier

transform before the integral. According to the integral feature of Fourier transform
t 1
F  f (t ) dt  F [ f (t )] (E.1)
 ik
In which F means Fourier transform - Fourier transform should be processed when
obtain speed signal and displacement signal from the integral of acceleration signal and
then perform Fourier inverse transformation and take the real parts of results, thus can
get time-domain speed signal and displacement signal. Taking Fourier transform of
acceleration signal will get
N 1 n
 i 2 k
A(k )   an e N
(E.2)
n0

Fourier transform of speed can be obtained after the first integral:

n
A N 1 1  j 2 k
V (n)   H ( k ) an e N
(E.3)
jk k  0 j 2 k f
Fourier transform of displacement can be obtained after the second integral:
N 1 n
1 1  i 2 k
S ( n)  
k2
A  
k  0 (2 k f )
2
H ( k ) a n e N
(E.5)

Where
1, f  kf  fu
H (k )   d (E.6)
 0, else

In equation E.5, symbol f is frequency resolution, and fd and fu are the upper

and lower limits of cut-off frequency respectively. N is number of points in data, and
k is the frequency of Fourier component.
 170

Appendix F: DFT and FFT

DFT of N point sequence x (n ) can be expressed as:

N 1
X (k )   x(n)WN k (F.1)
n 0

In which
2
j
N
WN  e

The coefficient WNnk  is periodic

WNnk  WNk (  n )  WN( k  N ) n (F.2)

n
k 
Thus some items in DFT operation can be merged. Using WN 2
 WNkn (symmetry)
and periodicity, DFT of long sequences can be decomposed into DFT of short
sequences.
 171

Appendix G: MATLAB source code of FFT noise reduction processing procedure

. The FFT noise reduction processing procedure in MATLAB is as following:

t=0:0.001:2;
x=cos(2*pi*t);
y=x+0.25*randn(1,length(t));
subplot(1,2,1);
plot(y);
axis([0,2000,-1.5,1.5]);
Y=fft(y);Y=abs(Y);
index2=50:1950;
Y(index2)=zeros(size(index2));
X=ifft(Y);X=real(X);
subplot(1,2,2);
plot(X);
axis([0,2000,-1.5,1.5]);
 172

Appendix H: Deduction process of wavelet

Definition of wavelets transform: assume  (t )  L( R)  L2 ( R) , ˆ ( ) means its

Fourier transform, if ˆ ( ) satisfies the following conditions (Wei et al., 2002):

ˆ ( )2
C   d   (H.1)
K 
Then  (t ) is called basic wavelet. Through dimension scaling and translation,  (t )
generates the following function family:
1
 t b
 ab  a 2  a  R, a  0, b  R (H.2)
a
It is called continuous wavelets generated by  (t ) , of which a is called scale
parameter, and b is translation parameter.
Under the discrete condition, wavelets sequence is:
 j ,k  t   2 j 2  2 j t  k  j, k  z (H.3)

According to the allowable conditions, when   0 , in order to ensure integrand as

effective value, there must be ˆ (0)  0 , so equivalent condition obtained from
equation above is

ˆ (0)    (t )dt 0 (H.4)


This equation shows that there is no direct component included in  (t ) , in which only
alternating component, namely unsteady , so it is called “wave”.
In order to make  (t ) has locality (namely attenuating quickly to zero outside of
limited interval), it has to be added one attenuation condition
c
 (t )    0, c  0 (H.5)
(1  t )1

The implication is: when t   , attenuation speed of  (t ) is faster than that of

1
, attenuation condition urges wavelets must have locality. This locality is called
t

“small” that “wavelets” derives its name from it. Wavelet transform is defined as:
1 t b 
w f (a, b)   f (t ) a ,b (t )dt  a 2
 f (t )  dt (H.6)
R R
 a 
Reverse transform is:
 173

1 1 t b 
f t     a W  a, b   a  dadb
2 f (H.7)
C RR

Among them, a is as scale factor, b reflects displacement.

 174

Appendix I: Tay Road Bridge Data processing: Route 2 to 5

A, Data of Route 2
This piece of data is with 30,000 points from point 145,001 to 175,000. Firstly, process
the acceleration data:
g(m/s²)

Denoise the data with dB8 wavelet:

 175

g(m/s²)

It is seen that the placement attitude of the smartphone is different from that in Route1.
Extract the baseline from the data:
g(m/s²)
 176

Eliminate the baseline by subtracting of last two results:

g(m/s²)

Conduct time correction with reference of GPS data:

Acct(145001)=714.559;

Acct(135000)=862.437;

The distance change in the 150 seconds selected from No. 714s to No. 863s is as follows:
GPSdist3=Location_GPS_dist(714:863);

Set 0 as the initial value of distance:

GPSdist3=GPSdist3-GPSdist3(1);

Interpolate the 150 points to 30,000:

xi=0.005:0.005:150;

x=0:149;

Dist3=interp1(x,GPSdist3,xi, 'spline');

Take the distance of 0-2500m as the abscissa:

 177

-2

-4

Acc2XAC
-6
Acc2YAC
Acc2ZAC
-8
0 500 1000 1500 2000 2500
Distance(m)

Process the gyroscope data as above:

Voltage

Denoise with dB8 wavelet:

 178

Voltage

Extract the baseline from the result of wavelet decomposition in last step:
Voltage

Eliminate the baseline by subtracting of the last two results:

 179

Voltage

Take Dist2 as the abscissa:

0.4
Gyo2XAC
Gyo2YAC
0.3 Gyo2ZAC

0.2

0.1

-0.1

-0.2

-0.3
0 500 1000 1500 2000 2500
Distance(m)

B, Data of Route 3
 180

The piece of data contains 30,000 points from point 175,001 to 205,000. Firstly, process
the acceleration data:
g(m/s²)

Denoise with dB8 wavelet:

 181

g(m/s²)

Extract the baseline from the result of wavelet decomposition:

g(m/s²)

Eliminate the baseline by subtracting of last two results:

 182

g(m/s²)

Time correction:
Acct(175001)=862.442;

Acct(205000)=1010.351;

The distance change in the 150 seconds from No. 862s to No. 1011s:
GPSdist4=Location_GPS_dist(863:1011);

Set 0 as the initial value of distance:

GPSdist4=GPSdist4-GPSdist4(1);

Interpolate the 150 points to 30,000:

xi=0.005:0.005:150;

x=0:149;

Dist4=interp1(x,GPSdist4,xi, 'spline');

Take the distance of 0-2500m as the abscissa:

 183

g(m/s²)

Process the gyroscope data as above:

Voltage
 184

Denoise with dB8 wavelet:

Voltage

Extract the baseline from the result of wavelet decomposition:

Voltage
 185

Eliminate the baseline:

Voltage

Take Dist4 as the abscissa:

 186

Voltage

C, Data of Route 4
As the vehicle runs slower, in this piece there are 40,000 points from point 210,001 to
250,000.
Firstly process the acceleration data:
 187

g(m/s²)

Denoise with wavelet:

g(m/s²)

Extract the baseline:

 188

g(m/s²)

Eliminate the baseline:

g(m/s²)

Time correction:
 189

Acct(210001)=1035.024;

Acct(250000)=1232.24;

Select 200 seconds from No. 1034s to No. 1233s:

The distance change in the 200 seconds is as follows:
GPSdist5=Location_GPS_dist(1034:1233);

Set 0 as the initial value of distance:

GPSdist5=GPSdist5-GPSdist5(1);

Interpolate the 200 points to 40,000:

xi=0.005:0.005:200;

x=0:199;

Dist5=interp1(x,GPSdist5,xi, 'spline');

Take the distance of 0-3000m as the abscissa:

g(m/s²)

Then process the gyroscope data:

 190

Voltage

Denoise with wavelet:

Voltage

Extract the baseline:

 191

Voltage

Eliminate the baseline:

Voltage

Take Dist5 as the abscissa:

 192

Voltage

D, Data of Route 5
This piece of data contains 40,000 points from point 250,001 to 290,000.
Firstly process the acceleration data:
 193

g(m/s²)

Denoise with wavelet:

g(m/s²)

Extract the baseline:

 194

g(m/s²)

Eliminate the baseline:

g(m/s²)

Time correction:
 195

Acct(250001)=1232.245;

Acct(290000)=1429.52;

Select 200 seconds from No. 1231s to No. 1430s:

The distance change in the 200 seconds is as follows:
GPSdist6=Location_GPS_dist(1231:1430);

Set 0 as the initial value of distance:

GPSdist6=GPSdist6-GPSdist6(1);

Interpolate the 200 points to 40,000:

xi=0.005:0.005:200;

x=0:199;

Dist6=interp1(x,GPSdist6,xi, 'spline');

Take the distance of 0-2500m as the abscissa:

-2

-4
Acc6XAC
Acc6YAC
Acc6ZAC
-6
0 500 1000 1500 2000 2500
Distance(m)

Then process the gyroscope data:

 196

Voltage

Denoise with wavelet:

Voltage

Extract the baseline:

 197

Voltage

Eliminate the baseline:

Voltage

Take Dist6 as the abscissa:

 198

0.2
Gyo6XAC
Gyo6YAC
0.15
Gyo6ZAC

0.1

0.05

-0.05

-0.1

-0.15
0 500 1000 1500 2000 2500
Distance(m)
 199

Appendix J: Real road data process: Route 2 to Route 4

Route 2

20

15

10

-5

AccelerationX
-10
AccelerationY
AccelerationZ

-15
0 0.5 1 1.5 2 2.5 3 3.5 4
Points 4
10
Figure J-1 Three acceleration data of Route 2

The Route 2 data contains 38,000 points, 380 seconds. The maximum speed is 7.75m/s,
the average speed is 5.20m/s. Firstly process the axis-correction using PCA, then de-
noise using wavelet.
 200

12
De-noiseZ
11.5

11

10.5

10

9.5

8.5

7.5

7
0 0.5 1 1.5 2 2.5 3 3.5 4
Points 4
10
Figure J-2 Wavelet de-noised result of Z axis of Route 2

Then the baseline of Z axis is extracted from the result of wavelet decomposition, as
shown in Figure J-3. So we get the Z axis without baseline drift as shown in Figure J-
4. Let the threadhold as:
Thd=speed.^2/100;
And using distance as the x axis, get the final result as shown in Figure J-5.
 201

Figure J-3 Baseline of Z axis of Route 2

2.5
AC Z
2

1.5

0.5

-0.5

-1

-1.5

-2

-2.5
0 0.5 1 1.5 2 2.5 3 3.5 4
Points 4
10
Figure J-4 Z axis of Route 2 without baseline
 202

2.5
ACZDenoise
2 Threadhold

1.5

0.5

-0.5

-1

-1.5

-2

-2.5
0 500 1000 1500 2000
Distance (m)
Figure J-5 De-noised Z signal with threadhold. The x axis is distance

So we get the position of potholes.

 203

Table J-1 Detection result of Route 2

ID Distance to the start point (m)

1 758.6
2 Didn’t check out
3 869
4 1078
5 1102
6 1149
7 Didn’t check out
8 Didn’t check out
9 1882
10 1930
11 1958

There are 3 potholes missed, the accuracy is 8/11=72.7%.

 204

Route 3

20

15

10

-5

-10

AccelerationX
-15 AccelerationY
AccelerationZ

-20
0 0.5 1 1.5 2 2.5 3
Points 4
10
Figure J-6 Three acceleration data of Route 3

The Route 3 data contains 26,000 points, 260 seconds. The maximum speed is 15.75m/s,
the average speed is 7.71m/s. Firstly process the axis-correction using PCA, then de-
noise using wavelet.
 205

16
DenoiseZ
14

12

10

-2
0 0.5 1 1.5 2 2.5 3
Poiints 4
10
Figure J-7 Wavelet de-noised result of Z axis of Route 3

Then the baseline of Z axis is extracted from the result of wavelet decomposition, as
shown in Figure J-8. So we get the Z axis without baseline drift as shown in Figure J-
9. Let the threadhold as:
Thd=speed.^2/100;
And using distance as the x axis, get the final result as shown in Figure J-10.
 206

Figure J-8 Baseline of Z axis of Route 3

8
AC Z
6

-2

-4

-6

-8

-10
0 0.5 1 1.5 2 2.5 3
Points 4
10
Figure J-9 Z axis of Route 3 without baseline
 207

8
AccZDenoise
6 Threadhold

-2

-4

-6

-8

-10
0 500 1000 1500 2000
Distance (m)
Figure J-10 De-noised Z signal with threadhold. The x axis is distance

So we get the position of potholes.

 208

Table J-2 Detection result of Route 2

ID Distance to the start point (m)

1 762
2 811
3 873
4 1087
5 1103
6 1152
7 1192
8 1787
9 1889
10 1936
11 1966

All the 11 potholes are detected out, the accuracy is 100%.

 209

Route 4

20

15

10

-5

-10

AccelerationX
-15 AccelerationY
AccelerationZ
-20
0 0.5 1 1.5 2 2.5 3
Points 4
10
Figure J-11 Three acceleration data of Route 4

The Route 4 data contains 26,300 points, 263 seconds. The maximum speed is 14.00m/s,
the average speed is 7.59m/s. Firstly process the axis-correction using PCA, then de-
noise using wavelet.
 210

16
DenoiseZ

14

12

10

2
0 0.5 1 1.5 2 2.5 3
Points 4
10
Figure J-12 Wavelet de-noised result of Z axis of Route 4

Then the baseline of Z axis is extracted from the result of wavelet decomposition, as
shown in Figure J-13. So we get the Z axis without baseline drift as shown in Figure J-
14. Let the threadhold as:
Thd=speed.^2/100;
And using distance as the x axis, get the final result as shown in Figure J-15.
 211

Figure J-13 Baseline of Z axis of Route 4

6
AC Z

-2

-4

-6

-8
0 0.5 1 1.5 2 2.5 3
Points 4
10
Figure J-14 Z axis of Route 4 without baseline
 212

6
AccZDenoise
Threadhold
4

-2

-4

-6

-8
0 500 1000 1500 2000
Distance (m)
Figure J-15 De-noised Z signal with threadhold. The x axis is distance

So we get the position of potholes.

 213

Table J-3 Detection result of Route 4

ID Distance to the start point (m)

1 755
2 Didn’t check out
3 863
4 1086
5 1103
6 1152
7 Didn’t check out
8 Didn’t check out
9 1887
10 1923
11 1935

There are 3 potholes missed, the accuracy is 8/11=72.7%.

 214

Appendix K: Publications

1. Chen, Jia, Yin Tao, Dalong Zhang, Xiaoli Liu, Zhen Fang, Qinwu Zhou, and Bo Zhang.
"Fatigue detection based on facial images processed by difference algorithm."
In Biomedical Engineering (BioMed), 2017 13th IASTED International Conference on, pp.
208-211. IEEE, 2017.
2. Tao, Yin, Jia Chen, Xiaoli Liu, Zhen Fang, Dalong Zhang, Qinwu Zhou, and Bo Zhang.
"Study of a new amblyopia diagnostic and therapeutic method along with the system
implementation." In Biomedical Engineering (BioMed), 2017 13th IASTED International
Conference on, pp. 121-124. IEEE, 2017.
3. Dalong Zhang, Wei Wei, Qinwu Zhou, Bo Zhang, “Road Pothole Detection Using Built-in
Accelerator in Smartphone.” In Transactions on Internet and Information Journal Systems
(TIIS). (Submitted)