POLITECNICO DI TORINO

Corso di Laurea Magistrale in Ingegneria Elettronica

Tesi di Laurea Magistrale

UWB Tracking System for Patient

Monitoring in Home Environment

Relatori:
Prof. Danilo Demarchi
Dr. Paolo Motto Ros
Dr. Stefano Sapienza
Candidato:
Maurizio Capra

Aprile 2018
Acknowledgments

I
Table of contents

Acknowledgments I

Summary VI

1 Introduction 1
1.1 Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 ToF Ranging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Positioning process . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.1 Geometric approach . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.2 Statistical approach . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 UWB technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4.1 Features and characteristics . . . . . . . . . . . . . . . . . . . 12
1.4.2 Impulse Radio (IR) . . . . . . . . . . . . . . . . . . . . . . . . 13
1.4.3 Multipath . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4.4 UWB regultion . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4.5 Commercial UWB systems . . . . . . . . . . . . . . . . . . . . 16
1.5 SLAM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2 Hardware description 20
2.1 Home Robot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.1.1 Hardware . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.1.2 Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 Decawave RTLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2.1 Hardware . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2.2 Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3 Study of the TREK1000 system and errors 37

3.1 Static error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.1.1 Static error respect to anchor’s position in the space . . . . . . 37
3.1.2 Oscillation distribution respect to anchor’s position in the
space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.2 Dynamic error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.2.1 Pendulum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

4 Positioning Algorithm 63
4.1 Trilateration algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.1.1 Closed-form Localization from Range-difference measurements 64

II
 4.2 Corrective function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3 Double system: 8 anchors . . . . . . . . . . . . . . . . . . . . . . . . 69
4.4 Best anchors filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.5 Offset correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.6 NLOS correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.7 Algorithm overall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

5 Test and Performance 75

5.1 Corrective function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.2 Double system: 8 anchors . . . . . . . . . . . . . . . . . . . . . . . . 82
5.3 Offset correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.4 NLOS correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

6 Anchors Calibration 88
6.1 Step 1: mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.2 Step 2: locate minimum ranges . . . . . . . . . . . . . . . . . . . . . 89
6.3 Step 3: reverse trilateration . . . . . . . . . . . . . . . . . . . . . . . 90

7 Conclusion and future work 95

Bibliography 97

III
List of figures

1.1 Time of Flight using IR light [20]. . . . . . . . . . . . . . . . . . . . . 4

1.2 ToA and TWR approaches [2]. . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Spheres intersection [2]. . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 a) ideal interception, b)interception with noisy ranges. [17] . . . . . . 7
1.5 RSS and ToA used for the range estimation. Target node’s position
obtained by intersection [3]. . . . . . . . . . . . . . . . . . . . . . . . 8
1.6 TDOA used for the range estimation. Target node’s position obtained
by intersection of hyperbola [3]. . . . . . . . . . . . . . . . . . . . . 9
1.7 AoA used for the range estimation. Target node’s position obtained
by intersection [3]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.8 Typical bandwidth of a UWB signal [1]. . . . . . . . . . . . . . . . . 11
1.9 Channel-Bandwidth relationship [1]. . . . . . . . . . . . . . . . . . . . 12
1.10 IR UWB signal [1] . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.11 Mutipath example [18]. . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.12 FCC regulation for emitted signal power [3]. . . . . . . . . . . . . . . 15
1.13 Allocated channels in different countries [5]. . . . . . . . . . . . . . . 16
1.14 Zebra Dart UWB kit [13]. . . . . . . . . . . . . . . . . . . . . . . . . 17
1.15 Pozyx board [19]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.16 DW1000 with antenna [7]. . . . . . . . . . . . . . . . . . . . . . . . . 18
2.1 Home Robot Hardware [5] . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Mimo monitor. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 Logitec webcam . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.5 Software architecture of the home robot [5]. . . . . . . . . . . . . . . 24
2.6 Navigation use case [8]. . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.7 EVB1000 front and riar view [7]. . . . . . . . . . . . . . . . . . . . . 26
2.8 Component side (left) and display side (right) of the EVB1000 [6]. . . 27
2.9 DIP switch for device configuration [7]. . . . . . . . . . . . . . . . . . 28
2.10 High level scheme of the DW1000 IC [6]. . . . . . . . . . . . . . . . . 28
2.11 Message standard format [6]. . . . . . . . . . . . . . . . . . . . . . . . 30
2.12 TWR message types [6]. . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.13 TWR algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.14 Three spheres intersection. . . . . . . . . . . . . . . . . . . . . . . . . 33
2.15 Log file [7]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.1 Anchor configurations and grid. . . . . . . . . . . . . . . . . . . . . . 38
3.2 2D plane and 3 anchors (first) configuration measurements. . . . . . . 39

IV
3.3 3D surface error for constellation (a) figure 3.1. (a) top view, (b) side
view. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.4 2D plane and 3 anchors (second) configuration measurements. . . . . 42
3.6 2D plane and 3 anchors (third) configuration measurements. . . . . . 45
3.8 2D plane and 4 anchors configuration measurements. . . . . . . . . . 47
3.10 2D plane division. The centre is represented by the average position
of the measurement cluster. Blue lines represent limit boundaries to
discriminate between different zones. . . . . . . . . . . . . . . . . . . 50
3.11 Amplitude oscillation for different release angles. . . . . . . . . . . . . 62
4.1 3D scheme of the geometric distances between anchors (sensors) and
tag (source). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2 Static measurements obtained using 4 anchors and the closed-form
algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3 Corrective function. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.4 Algorithm overall. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.1 Static measurements obtained using 4 anchors and the closed-form
algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.4 Static measurements corrected using the corrective function. . . . . . 81
5.5 Static measurements with 8 anchors witout any correction. . . . . . . 83
5.7 Moving obstacle, no NLOS correction applied. . . . . . . . . . . . . . 87
5.8 Moving obstacle after applying the NLOS filter. . . . . . . . . . . . . 87
6.1 Calibration procedure, L movement. . . . . . . . . . . . . . . . . . . . 90
6.2 Calibration measuremet, L movement. . . . . . . . . . . . . . . . . . 94

V
Summary

The purpose of this thesis work is the development and testing of a position algo-
rithm, based on Time of Flight (TOF) of UWB (Ultra WideBand) radio signals, for
autonomous indoor navigation. The system is composed by two types of transceivers:
anchors and tags. The former are fixed in known positions and work as a reference
while the latter are the units to track . Analysing the TOF of the UWB pulses
among transceivers, the algorithm is able to derive the distances (called ranges) of
the tags respect to all the anchors and consequently, applying a trilateration tech-
nique, the position of the tags in 2D. The tracking code has been translated and
embedded in a java library and implemented on a Home Robot monitoring system.
The final purpose of the platform is providing a rapid intervention and assistance to
elder people in case of health issues. The system will work in synergy with wearable
sensors which continuously record personalized health parameters depending on the
subjects pathology.
Since in this context false positives (i.e. health status reported as good when it
is not) are not acceptable, the systems sensitivity is extremely high. However this
inevitably leads to lot of false negative(health status reported as critic when it is
not) that translates in unnecessary stress for the subject and the clinicians. The
robot works as a filter for the alarms. The basic idea is to have a tag placed on the
patient and another on the robot, with several anchors spread around the flat ( one
per room).
In presence of potentially harmful symptoms , the robot activates and, guided by
the UWB tracking algorithm, travels to the patient establishing in the meanwhile a
web call with an operator that, thanks to a camera and a monitor, will verify the
subjects conditions. The content of this thesis describes step by step all the design
process of the UWB tracking platform together with the selection of the hardware.

The first chapter introduces some basics concept about TOF and the trilateration
process. Furthermore UWB technology and the reason why it is suitable for indoor
application are explained in detail.

VI
In the second chapter the hardware of the system and the robot are introduced in
detail. The UWB tracking platform is the Decawave Real Time Localization System
(RTLS), TREK 1000 based on the DW1000 module that creates the radio pulses.
This kit comes with a pre embedded tracking algorithm configured for 3 anchors
and 1 tag(since each kit contains 4 devices). In order to improve the performance
exploiting redundant information others two TREK 1000 has been used, having a
total of 8 anchors and 2 tags. The main hardware component of the Home Robot is
represented by the iRobot Create Development kit, this robotic platform is based on
robotic vacuum cleaners produced by iRobot. This robot has two wheels driven by
motors and a third one for support with encoders that measure the distance travelled
and the orientation. The control software is in a computer that communicates with
the iRobot through a 7 pin MiniDin connector. The platform is also equipped with a
webcam and a monitor that allow it to establish a web connection with an operator.
Furthermore the operator can manually move the iRobot thanks to a remote control.

The study of the TREK 1000 system and of the errors that affect it are presented
in the third chapter. Here the tracking speed and the sample frequency are investi-
gated with the antenna performance.

The fourth chapter includes a detailed description of the position algorithm and
all the steps performed during its development. First of all is analysed the new tri-
lateration algorithm implemented instead of the Decawaves one that allows to rely
on the ranges of only three anchors. The new algorithm considers n anchors(with
n equal or greater than 3), exploiting the redundant information to increase the
resolution of the tracking. In the second part of the fourth chapter are explained
all the corrective functions and the filters applied. The corrective functions allow to
mitigate the error due to non linear effects that could affect the UWB system, while
the filters are exploited to increase the robustness discarding range measurement
that are not considered consistent with respect to the current position of the Robot.

The fifth chapter is dedicated to the static and dynamic measurements and test
performed to verify the performance of the algorithm. Here measurements obtained

VII
placing the tag on test points forming a grid in the 2D plan are presented in graphic
and numeric way. It is shown how the error changes with different boundary condi-
tions and according to the corrective functions and filters applied.

The sixth chapter describes how the algorithm is embedded in the Robot application,
and it introduces the calibration process and the role played by the Simultaneous
Localization And Mapping (SLAM) procedure to obtain the floor plan of the house.
The calibration process is fundamental in case the anchors position is unknow, so
prior to perform the trilateration, they must be found. This operation performed
with the help of the SLAM can lead to obtain anchor positions with high accuracy
respect to the flat dimensions.

Finally the last chapter presents the conclusions, with the discussion of the results
obtained, and possible future improvements for the Home Robot application.

VIII
Chapter 1

Introduction

World’s population is aging and the number of people over the age of 65 is increasing.
It’s important for elder people to feel free and independent in their home environ-
ment. On the other hand, with aging several health issues and deseases can occur,
then is also importat to provide them with a reliable mean to support and prevent
tragic events during daily activities.

The idea of the Home Robot monitoring system is to track the health status of
the elder people that want to live in an independent way in their habitations. The
Home Robot is a tool used by the caregivers to track the patient’s status thanks
to a sensor and a webcam placed on the robot. The sensor is placed on the pa-
tient in order to supervise his status according to his patology. The robot remains
in a stand-by condition connected to its charge base, and in case of helth issues
communicated by the sensor, it navigates in autonomous way throught the home
environment establishing a connection with a caregiver using the webcam.
The caregiver has then the duty to decide if it is the case to call the 911, in fact
since the system must be very sensitive, some false negative alarms (health status
reported as critic when it is not) could occur.

The aim of the thesis is to support the Home Robot with a reliable indoor nav-
igation system and a position application able to locate the bot and the patient in
any home environment.
The system is composed by two kind of device: anchor and tag. The anchors are
devices positioned in fixed points while tags are devices free to move in the area
delimited by the anchors. Positioning one tag on the bot and one on the patient is
possible to track their positions and to direct the former toward the latter.

1
 1 – Introduction

The localization problem is composed by two phases: ranging and localization. In

the next paragraph those two steps will be introduced.

1.1 Technology
There is increasing demand of accurate localization system for many applications
such as robot control, goods and people tracking, indoor navigation and support
during critic events like earthquake or avalanche for people rescue. Nowadays cur-
rent technology offers a lot of possible implementation for positioning system: GPS,
infrared, ultrasound and RF (radiofrequency) techniques.

Global Positioning System (GPS) is a satellites based technology that provides po-
sition and time information. This kind of systems are very accurate and are used
in a lot of devices such as mobile phone, tablet and on-board car console. However
GPS cannot provide position with high resolution in dense urban and indoor envi-
ronment casue signal coming from satelites is reflected by the buildings and walls
and in many cases is too weak to penetrate them.

Infrared and ultrasound techniques are among the most used in indoor environ-
ment thanks to their semplicity and low cost. However in order to work properly
they require the Lign Of Sight(LOS) condition between anchor nodes and tags, so
this means that in scenario populated by many obstacles or walls these technologies
have a limited usage. This is due to the fact that light and sound are not able to
penatrate objects and walls, on the countrary RF signals are able to easily propagate
in home environment.

The four most used ranging techniques based on RF are: Angle of Arrival (AoA),
Received SIgnal Strength (RSS), Time of Arrival (ToA) also called Time Of Flight
(TOF) and Time Difference Of Arrival (TDOA).
In AoA method, anchors exploit the angle of arrival of the RF signal coming from
the tag to determine the position of the latter. In general this approach requires
directional antennas or antenna arrays, and many times it is used together with

2
 1 – Introduction

ToA in order to obtain a better accuracy. RSS methods are based on the fact that
an electromagnetic signal decays with the distance between the transmitter and the
receiver, but they are very susceptible to multipath and No Lign Of Sight (NLOS),
this means that they are not very reliable.
Time based technique (ToA) requires clock synchronization among nodes, generally
achieved by means of wired connections. When synchronization is not possible it is
called asynchronous ToA, also known as two way ranging.
When synchronization is achieved only between anchors and no information are
available about tag’s clock is possible to use TDOA.

1.2 ToF Ranging

In Time of Flight ranging the distance between two devices (anchor and tag) is
obtained measuring the propagation time of a signal that is travelling at a known
speed. The principle is very simple, knowing the travelled time τ and the wave
speed c, the distance d between to object is:

d=τ ∗c (1.1)

One of the simplest way to exploit this principle is using a source of light, tipically
infrared, and a sensor capable to identify the presence of this kind of light. Emitting
a IR ray in a certain direction and recording the time of emission t1, this will travel
till the object covered with reflectors and, once reflected, backward to the sensor
that can assert the time of reception t2. Tof will be calculated as:

t2 − t1
τ= (1.2)
2
In fact the difference between t2 and t1 represent the round trip time, i.e. the
time spent by the signal travelling from the emitter to the reflector and from the
object to the light sensor (that is tipically placed near the emitter). This is twice
the ToF.
Since IR light cannot travel through walls and objects (NLOS), RF signals are
preferred in harsh indoor environment full of possible obstacles such as can be an
home environment. The main difference using RF is that the tag is not represented

3
 1 – Introduction

Figure 1.1: Time of Flight using IR light [20].

by a reflector, but is another device able to receive and send a message from/to the
anchor.
The speed of an EM signals is comparable to the speed of light (300000 km/s), this
means that the measure of τ must be very accurate, in fact even a small uncertainty
multiplied by the speed of light could lead to big error in the range estimation.
In order to reach the maximum possible accuracy, ToA requires to have all the an-
chors synchronized among them. In this case, a single signal( or better a message)
is sufficient for ranging [2].
The receiver, once it gets the message, is able to estimate the range by subtracting
the timestamp of the transmission, recorded by the transmitter, from its own receive
timestamp. This approach is very simple, but since the anchors must be synchro-
nized, and in general this process is done by wiring all the anchors together, is very
expensive.

In the absence of global synchronization, the ranging process is called two way
ranging. Though it doesn’t require the synchronization, at least two messages are
required and it’s very sensitive to imperfections of the reference crystal of the nodes.

Setting treplyB and treplyA to constant values, when node A receives the response
message from B, it can calculate the ToF by subtracting the transmitted timestamp
from the reived one, as well as treplyB :

4
 1 – Introduction

Figure 1.2: ToA and TWR approaches [2].

troundA − treplyB
τ = tp = (1.3)
2
Where troundA is:

troundA = 2tp + treplyB (1.4)

the roundtrip time respect to node A.

A more robust approach consists in including one more message as depicted in figure
1.2, in this way the propagation time (ToF) is less affected by the errors introduced
by the crystals. This approach is called symmetrical double sided two way ranging
(SDS-TWR), and the ToA is:

troundA − treplyA + troundB − treplyB

τ = tp = (1.5)
4
The system used in this thesis exploits the symmetrical double sided two way
ranging approach since its anchors are not synchronized. As will be explained later,
the fact that the anchors are not wired is a point of strength for this application,

5
 1 – Introduction

since it make it more portable and easier to use in an home environment.

1.3 Positioning process

Position, generally speaking, is calculated using measurements of distances or angles
referred to fixed points (anchors) whose positions are known. A certain number of
fixed nodes are necessary to find a position of a tag that is free to move in the area
delimited by the formers. In order to estimate the position in 3D, at least four
anchors are needed, while in 2D just three.
In fact the problem in 2D consists of finding the intersection between three circum-
ferences, while in 3D consists of finding the interception of four spheres.

Figure 1.3: Spheres intersection [2].

The ranges obtained during the ranging process are always affected by errors,
like any other measurements. This means that during the positioning process the
intersection will not be perfect, or better the intersection will not result in a specific
point, but rather in a delimited area. In fact considering the ranges as measured
distances with random errors, their intersection represents an area in which the tag
is situated. The errors that affect the range measurements result in an uncertainty
in the tag’s position. In RSS and ToA these errors results in an uncertainty in the
circle shapes, while in AoA the uncertainty affects the shape of a line [1].

6
 1 – Introduction

Figure 1.4: a) ideal interception, b)interception with noisy ranges. [17]

There are two main positioning method: geometric approach and statistical ap-
proach.

1.3.1 Geometric approach

In the geometric approach the tag’s position is determined by the intersection of un-
certainty regions. In case of RSS or ToA, as described above, the distances between
the tag and at least three anchors are necessary to compute the position. In AoA
just two anchors are needed.

Positioning using ToF Time based positiong methods rely on the range mea-
surements obtained using the ToA. Once obtained the distances between the tag
and the anchors, by using a geometric approach, is possible to find the unknown
position. By collecting at least three range measurements from three fixed nodes in
known positions, is possible to locate the tag by interception.
As depicted in figure 1.5 the 2D location is obtained using three ranges d1 ,d2 ,d3
referred to three different source nodes. By solving these equations jointly is possible
to derive the position of the tag:

p
di = (xi − x)2 + (yi − y)2 i = 1,2,3 (1.6)

where x and y are the tag’s coordinates in 2D, [xi , yi ] is the position of the i-th
anchor and di is the i-th range.

7
 1 – Introduction

Figure 1.5: RSS and ToA used for the range estimation. Target node’s position
obtained by intersection [3].

When synchronization is achieved only between anchors and no information are

available about tag’s clock, is possible to use TDOA. In this approach, the param-
eter is the difference between two time of arrival. Multiplying this factor by the
speed of light, the position of the target is obtained with uncertainty on the shape
of a hyperbola as shown in figure 1.6 [3].
ToA is measured at two anchor nodes, τ1 and τ2 . Since anchors and tag are not
synchronized, an offset on τ1 and τ2 exist, and since the anchors are synchronized,
this offset is equal on both the nodes.

τT DOA = τ1 − τ2 (1.7)

τT DOA , that is the estimation of TDOA, is offset free thanks to the difference
operation [9], [10], [3].

8
 1 – Introduction

Figure 1.6: TDOA used for the range estimation. Target node’s position obtained
by intersection of hyperbola [3].

Positioning using RSS In RSS ranges are estimated measuring the signal strength
(the power). Knowing that the signal power decreases with the distance (path loss),
is possible to understand the distance from a source that transmits with a known
power.

d
P (d) = P0 − 10 · n · log10 ( ) (1.8)
d0
where P(d) is the received power at distance d, n is the path loss exponent, d0
is the reference distance. Once obtained the range measurements, the positioning
process is the same as seen for the ToA.

Positioning using AoA Using the AoA approach only two source nodes are
required to estimate the position of the tag [1].
As depicted in figure 1.7 the tag’s position is found by intersection of these two

9
 1 – Introduction

Figure 1.7: AoA used for the range estimation. Target node’s position obtained by
intersection [3].

lines. Each angle of arrival gives an equation like:

y − yi
tanψ = i = 1,2 (1.9)
x − xi

1.3.2 Statistical approach

Since geometric approaches are not able to solve the problem in presence of noisy
ranges, a statistical approach is preferred. Literature is plenty of source localization
techniques based on additive measurement error model.
A primary distinctions can be between iterative and closed-form algorithms. By
definition an iterative algorithm is an algorithm that needs to be performed several
times in order to reach a result that satisfies certain parameters, while the closed-
form one can compute the result istantly. In general for real time applications,
closed-form approach is preferred since is faster.
Other important distinctions are likelihood-based versus least-squared. All these
approaches consist of formulating an error or cost function and trying to minimize
it.

10
 1 – Introduction

1.4 UWB technology

Ultra Wide-Band is a type of RF technology characterized by a large bandwidth
and low energy level transmission protocol.
By definition stated by the Federal Communication Commission (FCC) an Ultra
Wide-Band system has a bandwidth larger than 500 MHz and the frequency in
which the system has the maximum power density fc must be greater than 2.5 GHz.
In case fc is lower than 2.5 GHz, Bf rac must be larger that 0.2, where:

B
Bf rac = (1.10)
fc
where B is the bandwidth of the system [1].
Being fH and fL the frequencies at which the power spectral density is 10 db below
the one in fc :

fH + fL
fc = (1.11)
2

2(fH − fL )
Bf rac = (1.12)
fH + fL

Figure 1.8: Typical bandwidth of a UWB signal [1].

11
 1 – Introduction

For any RF signal, Shannon’s theory characterizes the capacity of the signal
channel. It states that a direct relationship between capaciry and bandwidth exists
as well as an inverse relationship between bandwidth and power consumption:

C = B · log2 (1 + SN R) (1.13)

where C is the chanel capacity, B the bandwidth and SNR the signal to noise
ratio. For a specific capacity is possible to consume less power using a larger band-
width, this is one of the most important feature of the UWB: low power.

Figure 1.9: Channel-Bandwidth relationship [1].

1.4.1 Features and characteristics

UWB has many important characteristic that makes this technology suitable for
ranging and consequently for positioning systems. The most important feature is
the large bandwidth in comparison with the other RF systems that are based on
narrow-band.
Thanks to the reverse relationship existing between frequency and time domain, the

12
 1 – Introduction

direct result of the large bandwidth is the short life time of a UWB signal. This
means high time resolution that is an important characteristic for positioning since
makes the signal robust against multipath.
High bandwidth means also high speed comunication. Furthermore, UWB allows to
use low carrier frequencies, where signals can more easily travel through objects.
High time resolution and short wavelength result in robust technology against mul-
tipath and fading.
Last but not least, since UWB can be transmitted in base band allows the hardware
to be simpler and cheaper.
Conventional RF systems transmit information by modulating frequency, power or
phase tipically of sinusoidal wave, while UWB transmits data by generating radio
energy at specific time intervals, this meakes it suitable for time position or time
modulation.

1.4.2 Impulse Radio (IR)

One of the most used method for UWB transmission is the Impulse Ratio (IR). In
this method information of the simbol is uttered by position and/or polarity of the
signal. The UWB pulse can be generated from a Gaussian pulse:
√
2 − 2πt22
pulse(t) = ± e α (1.14)
α
where

α2 = 4πθ2 (1.15)

where α is the pulse factor, while θ is the variance.

As stated in equation 1.14 α is responsible of the pulse width, greater its value, the
narrower the width. Pulses of the duration of the order of fractions of nanoseconds
are UWB signals.
UWB signal can be modulated using Time-Hopping(TH) and Direct Sequence(DS)
modulation. Here an example of TH transmission:
Two consecutive signals form a symbol. IR signal occupies one of the chip-
intervals(Tc ) inside a frame (Tf ). In the figure depicted in 1.10 the stream represents

13
 1 – Introduction

Figure 1.10: IR UWB signal [1]

a binary data.
Impulse UWB is a good technology for location sensor networks, thanks to its high
quality communication. The fine time resolution of some sub-nanosecond pulses
allows an accuracy of a few centimeters in distance measurements.

1.4.3 Multipath
By definition multipath is the propagation phenomenon that results from a radio
signal reaching the receiving antennna through two or more paths. Typical causes
of multipath are reflection and refraction of the ionosphere, or more in general in
an indoor environment, walls and objects. Signals affected by multipath result
in constructive or destructive interference. The latter causes fading, that is the
variation or attenuation of the signal.
Signal reflected by multiple objects results in several signals that reach the receiver in
different moments causing interference that makes the receiver unable to distinguish
the original signal that travels in LOS.

Thanks to the large bandwidth that results in high time resolution(short pulses),
UWB is a robust technology respect to this problem. This enables the receiver to
identify multipath reflections from the original signal. This property makes the
UWB suitable for indoor environments full of obstacles and objects.

1.4.4 UWB regultion

UWB technology is license free, so anyone can implement an UWB system without
the necessity of a patent. However since UWB cover a wide band of frequency a

14
 1 – Introduction

Figure 1.11: Mutipath example [18].

regulation must exist to avoid collision and interference among different communi-
cation methods.
Federal Communication Commission (FCC) is one of the most important organiza-
tions that provides regulations. This was the first organization to design rules about
UWB in 2002.

The IEEE 802.15.4-2011 states that the standard output level for UWB transmis-
dBm
sion(Equivalent Isotropically Radiated Power EIRP) is -41 M Hz
, and the spectrum
from 3.1 to 10.6 GHz is divided in 14 channels of 500 MHz each.

Figure 1.12: FCC regulation for emitted signal power [3].

Figure 1.12 demonstrates the FCC’s regulation about the transmitted UWB
power.

15
 1 – Introduction

Many countries, as result of the regulations, have allocated frequency spectrum for
UWB use. As stated above, the UWB spectrum is divided in 14 channels, as it is
possible to see from figure 1.13 many countries don’t allow the use of several chan-
nels since they are already used by other applications such as could be satellites
communication.

Figure 1.13: Allocated channels in different countries [5].

DAA stands for Detect and avoid, and it is a technique to avoid interference
between transmitter and wireless environment. Thanks to this UWB can use the
designed channel without interfeering with other technologies that work on the same
frequencies.

1.4.5 Commercial UWB systems

Nowadays many commercial UWB systems are present on the market. They dif-
fer in accuracy, cost, dimension and maximum covered range. Some examples are
presented in the next paragraphs.

16
 1 – Introduction

Zebra Zebra Dart UWB [13] (figure 1.14) is designed for real time position in
2D and 3D applications . This system is compliant with the international standard
IEEE 802.15.4.f.
Its features include high perormance in environments affected by strong multipath,
battery life up to 7 years at 1Hz blink rate, range capability up to 200m, pro-
grammable frequency sample of the tag up to 200Hz, waterproof antenna and an
accuracy around 30cm.
Zebra tags have a diameter of anly 4 cm and thanks to their long battery life they
can last for 7 years. The major limitation is the cost that is 12,000 USD.

Figure 1.14: Zebra Dart UWB kit [13].

Pozyx Pozyx system, figure 1.15, provides positioning and motion information.
The system is composed by 5 devices, 4 anchors and 1 tag. 4 anchors are required
for 3D positioning, 3 for 2D. This board is Arduino compliant and it’s equipped with
accelerometer, gyroscope and magnetometer that are used to obtain the orientation
of the device. All these sensors are affected by bias and error, however together
these kind of errors can be mitigated.
The frequency sample of the system is up to 140Hz with an accuracy of 10cm. The
major limitation is the maximum range that is typically around 30m. The cost is
aound 740 USD.

Decawave Decawave is an Irish company that designed the DW1000 IC (1.16)

[7], an integrated circuit designed for indoor real time location. It is IEEE 802.15.4-
2011 compliant. It has an accuracy of 20cm even with moving tag up to 5m/s.

17
 1 – Introduction

Figure 1.15: Pozyx board [19].

Communication range up to 290m and high fading environment robustness.

Its low power consumption makes it suitable for monitoring aplication. TREK1000
is Decawave kit composed by 4 configurable EVB1000 that mount the DW1000.
The cost for the entire kit is 900 USD.

Figure 1.16: DW1000 with antenna [7].

According to the low cost and good performance such as high accuracy and the
programmable kit, Decawave system was the designated one for this thesis work.

18
 1 – Introduction

1.5 SLAM
Mobile robots must be able to position theirself inside working environment. In ideal
case GPS can be used to obtain absolute position of the bot, but in many cases and
application, the GPS is not suitable such as an indoor environment. Sometimes also
a problem related with cost, size and weight could limit the use of that technology.
In some cases, knowing details about the working area, a robot can locate itself using
exteroceptive sensors like laser scanner or camera for relative position measurements
to known landmarks.

The problem of Simultaneous Localization And Mapping (SLAM) [4] is of high

interest in many applications, ranging from planetary exploration to transportation
and environmental monitoring. SLAM problem consists of constructing and updat-
ing a map of a certain environment and at the same time tracking the position of
an object within it.
SLAM uses several types of sensors such as: laser rangefinders, LiDAR, sonar sen-
sors, cameras. According to sensors many different algorithms have been exploited
based on Kalman filters or particle filters. They elaborate a probability function for
the position of the robot and the shape of the map.

In 2D cases the kinematics of the robot is given by a mixture of rotation and move
forward commands. All this command are implemented with some additional motor
noise. An alternative approach consists of reading odometry data from the robot’s
encoders of the wheels after each command.

So thanks to SLAM a robot can navigate in an indoor enviroment, constructing

a map of this, and track its position. Giving this a bot is able to navigate in an
unknown surroundings.

19
Chapter 2

Hardware description

This section describes the hardware used starting from the robot and its feature
finishing to the UWB Real Time Location System (RTLS).

2.1 Home Robot

Figure 2.1: Home Robot Hardware [5] .

The Home Robot is composed by three main parts: iRobot, monitor and webcam.
The iRobot [14] represents the most important element, the one that is in charge
of the movement and it is also the physical base that support all the rest. The
monitor allows the patient to interact with the axternal oparator connected by
videoconference. The webcam reprents the ”eyes” of the caregiver.
All these elements require an external processing unit such as a computer in order
to interact with each other. The PC has the duty of coordinate all the processes by
means of a dedicated java software.

20
 2 – Hardware description

2.1.1 Hardware
iRobot Create

iRobot Create is a development kit produced by Roomba that allows the user to
program it without knowing the low level code. iRobot Create’s Open Interface (OI)
provides to the user with a set of drive commands. It is possible to attach external
electronics parts to it like light display or ranging sensors.

(a) iRobot Create top view [14]. (b) iRobot Create bottom view [14].

This platform is based on typical vacuum cleaners. It is equipped with two dif-
ferential driven wheels. Thanks to optical encoders placed on the wheels is possible
to track the motion and perform accurate movements. The OI provides the travelled
distance within 1 mm of resolution, as well as the turned angle within 1 degree of
accuracy.
Each wheel can rotate independently up to 500 mm s
. Thanks to simple formulas is
possible to obtain the geometric center speed and the angular velocity:

ωl + ωr m
v= r [ ] (2.1)
2 s

ωl − ωr rad
ω= r [ ] (2.2)
2d s
where v is the geometric speed, ωl and ωr are the angular speed of the left and
right wheel respectively, r the wheel radius and d the distance between one wheel
and the geometric center.

21
 2 – Hardware description

iRobot Create is equipped with two communication port: cargo bay connector and
a serial connector. Connecting the last one with the serial port of a PC is possible to
start a serial terminal program capable of sending commands. This communication
is set to work at 57600 baud, 8 data bits and 1 stop bit.

MIMO monitor um 720s

Figure 2.3: Mimo monitor.

Mimo Um-720s is a compact foldable 800x480 monitor that can be tilted till 90
degrees. Video, touch and power are provided via USB connection. This 7-inch
resitive LCD display is compatible with Windows, Linux and Mac OS.

Webcam Logitec Orbit

Quickcam Orbit is a webcam by Logitec mounted on the bot in order to provide the
user with sight of the current environment around it.
It’s a 2 MegaPixel camera equipped with autofocus and USB 2.0 connection. The
latter allow the remote control thanks to which is possible to rotate the camera of
102 degrees.

Figure 2.4: Logitec webcam

22
 2 – Hardware description

2.1.2 Software
The PC used for this application is equipped with Linux distribution, Ubuntu 17.10.
This provides the serial communication with the robot, and it has also all the li-
braries needed to control the webcam and the external monitor.

The software is in charge of managing the movements of the robot transmitting

commands through the serial port and receiving the feedback form the sensors. An-
other important task is the management of the videoconference and the remote
control. In fact the caregiver has the ability of controlling the robot, in particular
iRobot movements and webcam orientation, thanks to an easy interface.

The framework used for the software interface is Node.js. This is based on JavaScript
language that is one of the most used programming language nowadays. Node.js is
a JavaScript interpreter with low level APIs that allows the user to access processes
and files. Thanks to its event-driven approach, its asynchronous architecture never
blocks the execution.

Unlike the traditional programming where the execution is sequential and in case
of blocking function the entire execution is stopped, Node uses callback functions
that are invoked when the result from a blocking function is available. Thanks to
this the execution is always in a running mode. In the Home Robot application
this particular approach is preferred since allows the interface to monitor multiple
features at the same time, avoiding the interruption of the execution that will cause
loss of information coming from onboard sensors.

Software architecture

In order to exploit the event-driven asynchronous approach the software is divided

in two parts: the Controller that is in charge of control the robot hardware and
the Player that manages the web application used by the caregiver to establish a
communication channel with the patient and the romote control [5].
The Player is designed to run on an external server, in this way it could work as a
server-client framework allowing multiple robots to connect with externl operators.

23
 2 – Hardware description

Figure 2.5: Software architecture of the home robot [5].

In this specific application both Player and Controller run on the PC.
The communication from Player to Controller is called publish-subscibe and allows
the former to communicate with one robot at time, group of them or all of them.
The technique used to communicate from the Controller to the player is instead a
producer-consumer link. The producer is the robot that starts sending messages on
the communication channel, while the Player is the consumer that reads information
present on the channel that acts like a FIFO buffer memory.

2.2 Decawave RTLS

The Wireless Sensors Network (WSN) represented by the Decawave RTLS is used for
two main reasons in this application: localize the robot and the patient in the indoor
environment and transmit the alarm messages from the wearable sensor placed on
the patient to the caregiver by means of the software present on the PC connected
to the robot.

24
 2 – Hardware description

Decawave RTLS is composed by anchors and tags that are represented by config-
urable devices. For this application TREK1000 development kit has been chosen.
Its features are presented in the following sections.

2.2.1 Hardware
TREK1000

TREK1000 is composed by 4 EVB1000, evaluation

boards for the DW1000 IC, stands and USB cables.
TREK stands for Two Way Ranging(TWR) RTLS IC
Evaluation Kit[7]. This kit provides a solution to the
RTLS problem using UWB technology. UWB based
RTLS use time based measurements like ToA to reach
high accuracy position. TREK1000 exploits its software
application and UWB ToA measurements for 2D and 3D
location. Since for 3D positioning 4 anchors are needed
and only 4 devices are provide (3 anchors and 1 tag),
Decawave software takes as solution of the sphere inter-
Figure 2.6: Navigation
section the one below the anchor’s plan. Furthermore
use case [8].
exists the possibility to expand the kit using more EVB
for a maximum of 4 anchors and 8 tags. The fourth an-
chor doesn’t increase the accuracy, but according to Decawave algorithm, is used to
identify if the tag is above or belove the anchor’s plan.

It works in three different use cases: tracking use case, geo-fancing use case and
navigation use case. In the first one the tag is located respect to fixed anchors,
the second one determines when tags leave or enter specific areas near an anchor
positioned in the center (suitable for child monitoring, security bubble and personal
safety). Navigation use case is the same of the tracking one, but the processing unit
(PC) is connected to the tag instead of the anchor, and is the configuration used
for the final application. The robot is equipped with one tag connected to the PC,
another tag is placed on the patient and the anchors are placed around the home
environment.

25
 2 – Hardware description

Anchors placement Decawave gives some hints to the user about the position
of the fixed anchors using the TREK1000 in order to obtain the maximum possible
accuracy[6]. The three anchors must be mounted at the same high, at height around
2-3 meters, above people’s head to avoid interference. Anchors mounted to form a
triangle, in fact if placed on a straight line the uncertainty intersection area of the
three circle would be too wide resulting in poor accuracy. Antennas more than 15
cm far from the nearest wall to avoid interference.

EVB1000

Figure 2.7: EVB1000 front and riar view [7].

EVB1000 is an 70mmx70mm evaluation board equipped with DW1000 IC. Ev-

erything is coordinated thanks to a STM32F105 ARM Cortex M3 processor that
communicates with DW1000 IC by means of a SPI interface. In order to be pro-
grammed, the processor is connected to a twenty pin JTAG header. Two power
connectrors can supply the system, a micro USB and a DC 3.3V power input.
This board allows the development of RTLS applications and WSN useful in many
different fields like healthcare, warehousing, logistic and automation.
As depicted in figure 2.8 on the display side the Printed Circuit Board (PCB) is
provided with an LCD display and an eight(8) pin DIP switch. Thanks tho the latter

26
 2 – Hardware description

Figure 2.8: Component side (left) and display side (right) of the EVB1000 [6].

is possible to configure the device. Each DIP switch selects a certain configuration
and in the up position is considered ON, while in the down position is OFF:

• The first one must always be on;

• the second allows the selection of the data rate, 6.8Mbps [ON] or 110Kbps
[OFF];

• the third selects the UWB channel, channel 5 [ON] or chanel 2 [OFF];

• The fourth establish if the device is an anchor [ON] or a tag [OFF];

• DIP from 5 to 7 represent the number of anchor[0-2] or the number of tag[0-7]

in binary code;

• The last one is reserved.

As explained above the board is capable to work on two different channels:

channel 2 at 3.6GHz and channel 5 at 6.5GHz. Each country as its own regulation
about the allocated spectrum. In USA both channels 2 and 5 are allowed, while in
Europe only channel 5 is permitted.
On the top of the device there is a SMA connector for the antenna. The antenna

27
 2 – Hardware description

Figure 2.9: DIP switch for device configuration [7].

used is omni-directional with 1 dBi of gain for channel 2 and 3 dBi gain for channel
5.

DW1000

Figure 2.10: High level scheme of the DW1000 IC [6].

DW1000 is Decawave UWB transceiver IC. It integrates RF circuitry, power

management, antenna and clock temporization ciruitry all on board. The one used
on EVB1000 uses an external antenna in order to exploit the benefits of a bigger
one. Since it is IEEE 802.15.4-2011 compliant can be used in many countries ex-
ploiting its 4 RF bands from 3.5 GHz to 6.5 GHz. The 23 mm x 13 mm x 2.9 mm 24
pin package host also a SPI interface to allow the communication with an external
processor.
Thanks to its integration, no external RF circuit are required, so this makes its

28
 2 – Hardware description

usage simpler. Furthermore its low power consumption with its low cost make this
module one of the best choise for the design of a RTLS system.
The on board 38.4 MHz crystal has been trimmed during the production process to
reduce the initial error to 2 ppm. This is important for the ranging accuracy where
anchors and tags are not synchronized among them. The board hosts an OTP(One
Time Programmable) memory that the user can exploit to save the antenna cali-
bration information.

2.2.2 Software
In this section the software relative to TREK1000 system will be analyzed. This
comprends the WSN messages exchange, Two Way Ranging process and the GUI
interface for PC designed by Decawave.

WSN

The Wireless Sensors Network provides mean through which anchors and tags ex-
change messages that are foundamental for the localization process. In this way
Tag0 connected to the robot can receive the position of Tag1 placed on the patient.
The communication protocol is organized in a Time-Division Multiplexing (TDM)
approach. The superframe is represented by a TDM cycle, which is composed by
a fixed number of time-slots of equal duration. For each time slot only one tag is
allowed to transmit using UWB channel, while the other is sleeping. The Anchor0
is the TDM master, and it is on charge of maintaining tags into their own time-slots.

The message format is the IEEE 802.15.4 standard figure 2.11[6]. Since TREK
application always uses data frames with 2 octet for source and destination address,
the two frame control slots are constant.
The sequence number of the octet is incremented modulo 256 every time a frame
is sent. The source and destination values depend on EVB1000 board’s configuration
of the DIP switches, anchor or tag mode and number. The 2 byte FCS are a
CRC frame check sequence that is generate by the DW1000 IC and attached to the
transmitted message.
The content of the ranging message depends on the type of the message sent. Three

29
 2 – Hardware description

Figure 2.11: Message standard format [6].

types exist according to the sequence exchange for the TWR and are presented in
the next paragraphs.

Figure 2.12: TWR message types [6].

Poll message Poll message is sent by the tag and has the purpose to initiate range
measurement. The function code is 0x81, this octet identifies the Poll message. The
range number depends on the sequence and each time is incremented.

Response Message Anchor is responsible of sending the Response meassage after

the Poll message. It is composed by 8 octets. The first is the function code 0x70,
the second and the third together form the sleeping correction, a parameter that
adjusts the tag’s sleep duration so the tag’s activity can be synchronized with the
time slots and avoid, in this way, to interfere with other tags. Then 4 octets, 32

30
 2 – Hardware description

bits, that represents the ToF value measured during the previous message exchange.
The last octet is the range number.

Final message Final message is sent by the tag as conclusion of the message ex-
change. It is 44 octets long. The first is the function code 0x82, the second the range
number, then 5 octets for the Poll TX time that is the timestamp, so the precise
time the Poll frame was trasmitted by the tag. Then 4 equal fields composed by
5 octets, each one representing the Resp RX time, the timestamp for the response
time from anchor 0, 1, 2, 3. Final TX time is a 5 octets field that contains the
timestamp of the final message. The last frame is 8 bit and specifies which response
time is valid.

The final application is based on two foundamental steps: the measure of the
distances between anchors and tags, and the positioning algorithm that determines
the 2D location of the Home Robot and the patient in the environment. The first
step is performed thanks to the TWR process analyzed in the next section.

TWR

Figure 2.13: TWR algorithm.

31
 2 – Hardware description

The Tag, that is in charge of starting the TWR process, attemps to range to
four anchors and then, once sent the final message, go to deep sleep mode to save
energy. After a superframe time, it will attemp again.
In figure 2.13 is depicted the TWR algorithm scheme, the tag sends a Poll message
which is received by three (or four) anchors. Then anchors reply sequentially with
packets RespA, RespB and RespC (RespD, not present in the picture). At the end
the tag sends the Final message received by all the anchors. This allows to locate
the tag after sending only 2 messages and receiving 3. The low number of exchanged
messages represents a saving in terms of battery power and air-time.

Recording the timestamps of the messages is possible to calculate the ToF be-
tween the tag and the anchors as shown in picture. Times called Treply are known
since are constant written in the code used to synchronize the response process.
Treply1A < Treply1B < Treply1C in this way the tag knows the exact sequence of the
response messages. These time periods consider also an important issue that is rep-
resented by the time during which the message in not in air but is handled by the
device. In fact there is a small time that elapses from when the processor sends
the message to when the message is in air. The same happens when a message is
received, there is a delay from when the message reaches the antenna to when it is
processed by the processor. So these time periods are tuned in order to cosider this
delay.

Positiong algorithm

Decawave’s positioning algorithm is a trilateration algorithm. it exploits 3 anchors

to determine the 2D position and the fourth (if present) to choose the right solution
on the z axis, otherwise the solution is always the one below the anchor’s plan. The
input of the algorithm are the 3D position of the anchors and the measured ranges
obtained by means of the TWR process.
The trilateration process represents the steps necessary to find the solution of the
intersection of three spheres. These spheres are located on the plane z = 0 and their
radius are represented by the ranges. Trilateration will lead to two different solution
with the same x and y value, but one with z value below the anchor’s plane and the

32
 2 – Hardware description

other above. The range coming from the fourth anchor could be used to chose the
right solution.

The tag position is found by formulating this geometric problem as intersection

of spheres and then solving for the unknown x, y and z.

Figure 2.14: Three spheres intersection.

According to figure 2.14 the equation of the three spheres are:

r12 = x2 + y 2 + z 2 (2.3)

r22 = (x − d)2 + y 2 + z 2 (2.4)

r32 = (x − i)2 + (y − j)2 + z 2 (2.5)

Where d, i and j are the distances from the center of the first sphere located in
(0;0) as depicted in figure 2.14. Spheres intersection is located in (x, y, z), obtained
solving the three equations simultaneously. First of all equation 2.3 and 2.4 to find

33
 2 – Hardware description

x:

r12 − r22 = x2 − (x − d)2

r12 − r22 = 2dx − d2 (2.6)
r2 − r22 + d2
x= 1
2d

Now x is known since the formula is composed of known value. Next step is to
find y using equations 2.3 and 2.5, writing the first as z2 = r21 - x2 - y2 :

r32 = (x − i)2 + (y − j)2 + r12 − x2 − y 2

r12 − r32 − x2 + (x − i)2 + j 2
y= (2.7)
2j
r − r3 + i2 + j 2
2 2
i
y= 1 − x
2j j

Knowing x also y is known. Last step consists in finding z value. Once again
from equation 2.3:
q
z = ± r12 − x2 − y 2 (2.8)

As explained before this equation admits two possible solutions, one above the
anchor’s plane and one below.

The values of i, j and d are found considering the centers of the spheres as vec-
tors from the origin and building a cartesian plane consistent with their directions.
The x axis is in the direction from P1 to P2:

P2 − P1
êx = d = ||P 2 − P 1|| (2.9)
||P 2 − P 1||
i is the x component of the distance between P1 and P3, so geometrically speaking
is the scalar product between the distance vector and the x unit directional vector:

i = êx · (P 3 − P 1) (2.10)

34
 2 – Hardware description

The unit directional vector for the y axis is:

P − P 1 − iêx
êy = (2.11)
||P − P 1 − iêx ||
j represents the value of the distance between P3 and P1 along the y direction,
so the scalar product between the distance vector and the y unit directional vector:

j = êy · (P 3 − P 1) (2.12)

Once obtained the unit directional vectors of x and y, the z one is simply the
cross product between these two.

êz = êx × êy (2.13)

Finally the intersection point is calculated as:

P~T = P 1 + xêx + yêy ± zêz (2.14)

This result is expressed respect to the origin and using coordinate obtained
respect to the anchors position.

Log File

Using the Pc application provided by Decawave is possible to save Log file produced
by this application. This file is a text one in which all the information about the
anchors position, range measurements and timestamps are written.
Figure 2.15 shows how a log file is made. In the first raw several information
about the system configuratior are written. Then raw by raw range values appear
with range sequence number and the anchor and tag ID representing to which devices
the range belongs to.

35
 2 – Hardware description

Figure 2.15: Log file [7].

36
Chapter 3

Study of the TREK1000 system

and errors

In order to verify the performance of the system and to be sure of the feasibility of
the final application using Decawave TREK1000 development kit several tests have
been made.
The first test is a static one in order to understand which anchors configuration
would lead to the best accuracy. This means placing the anchors in different fixed
constellations and compare measurements to figure out when the smallest error is
reached.
Then since the range measurements can be seen as the real value plus a random
one that represent the error, from measure to measure the position of the tag varies
randomly of a certain value. This variation lead to an oscillation of the tag position
around the meadian value. This oscillation is shown in the next sections.
The last test is a dynamic one and allows to verify the ability of the system to track
a moving object.

Note: all measurements have been obtained by means of the log file from decawave
and then sucbsequently eleborated using designed ad hoc MatLab scripts.

3.1 Static error

3.1.1 Static error respect to anchor’s position in the space

In this section static measurement errors will be analyzed comparing different an-
chor constellations. In order to do so, a grid of 6 meters x 4 meters is built on the
floor. The anchors are positioned at the vertices of this rectangular grid using 1,17

37
 3 – Study of the TREK1000 system and errors

m tall tripods, while the tag is moved(on the floor) from measure to measure inside
and on the side of this rectangle and on the vertices of the small square of 1m x 1m
size forming the rectangle.

Four different anchor configurations have been investigated in order to figure out
which is the best solution, analyzing measurements dispersion and error distribution.

(a) (b)

(c) (d)

Figure 3.1: Anchor configurations and grid.

In figure 3.1 the four configurtion are showed, where the black diamond repre-
sent the anchors, while the green dots are the test points where the tag is placed.
First 3 configurations exploit only 3 anchors positioned forming a triangle, while the
last one exploits 4 anchors. Anchor 0 always in position (0,0), anchor 1 always in

38
 3 – Study of the TREK1000 system and errors

position (6,0), anchor 2 changes from (0,4) to (3,4) or (6,4), anchor 3 only in the
last configuration in position (6,4) on a tripod 1,64 m tall. In fact, as explained
by Decawave, the fourth anchor must be on a different plane respect to the other
anchors to be able to choose the right solution among the two obtained thanks to
the spheres intersection

Note: in the following pictures anchors are represented by black diamonds, test
points are green dots, each blue dot is a measurement. Each cluster of measurements
belonging to the same test point has a red line that represents the linear interpolation
of this cluster and helps the observer to understand in which direction it extends.
Furthermore the area among the anchors is highlighted.

3 anchors, first configuration results

Figure 3.2: 2D plane and 3 anchors (first) configuration measurements.

In figure 3.2 is shown the first anchor constellation, where anchors form a rectangle
triangle. Clusters of points are more accurate in the core of the highlighted area,
so their average point is closer to the ideal point(green dots). This has a geometric
explanation, in the middle of the area covered by the anchors the intersection area

39
 3 – Study of the TREK1000 system and errors

is tight, while on the border the uncertainty area obtained by intersection is spread.
This lead to an higher uncertainty on the border and in the outside region.

The following table shows the error on the x and y axis and the absolute error
calculate as the euclidean distance between the average point of the cluster and the
ideal point. At the end maximum absolute error, Root Mean Square Error (RMSE)
and standard variation (σ) of the absolute error are presented.

x[m] y[m] X err [m] Y err [m] Abs err [m]

0 1 -0,1854 -0,0806 0,2022
0 2 -0,2034 -0,0258 0,2050
0 3 -0,1788 0,0503 0,1857
1 0 -0,0757 -0,0690 0,1024
1 1 -0,1473 -0,0511 0,1559
1 2 -0,1146 0,0037 0,1146
1 3 -0,1478 0,0887 0,1724
1 4 -0,1026 0,1574 0,1879
2 0 -0,0404 -0,0299 0,0503
2 1 -0,1060 -0,0288 0,1098
2 2 -0,0762 0,0113 0,0771
2 3 -0,0951 0,0557 0,1102
2 4 -0,0723 0,1583 0,1740
3 0 0,0239 0,0425 0,0488
3 1 -0,0517 -0,0116 0,0530
3 2 -0,0145 0,0491 0,0512
3 3 -0,0005 0,1111 0,1111
3 4 -0,0242 0,1583 0,1601
4 0 0,1023 0,0673 0,1224
4 1 0,0260 0,0047 0,0264
4 2 0,0429 0,0811 0,0918
4 3 0,0207 0,1424 0,1439
4 4 0,0309 0,1842 0,1868
5 0 0,1140 0,0860 0,1428

40
 3 – Study of the TREK1000 system and errors

5 1 0,1087 0,1115 0,1557

5 2 0,1064 0,1314 0,1691
5 3 0,0697 0,1586 0,1732
5 4 0,0635 0,1692 0,1808
6 1 0,1034 0,0425 0,1118
6 2 0,1050 0,1560 0,1881
6 3 0,0774 0,0772 0,1093
6 4 0,0775 0,0942 0,1220
max err[m] 0,2050 σ[m] 0,0506
RMSE [m] 0,0844

Table 3.1: Static errors of the anchor constellation (a) figure 3.1.

The sign of the errors derives from the the fact that is calculated as measured
position - test point position.
Following pictures are 3D representation of absolute error as difference between test
point and average cluster position. The surface is an interpolation of test points
with absolute error as height in the z axis direction.

4 0.25 0.25

3.5
0.2 0.5 0.2
3

0.4
2.5
0.15 0.15

2 0.3

0.1 0.1
1.5
0.2

1
0.05 0.1 0.05
4
0.5
2
0
0 0 0 1 0
2 3 4 0
0 1 2 3 4 5 6 5 6

(a) (b)

Figure 3.3: 3D surface error for constellation (a) figure 3.1. (a) top view, (b) side
view.

Figure 3.3 shows the surface error, where x-y plane is the area covered by the

41
 3 – Study of the TREK1000 system and errors

experiment, while in z direction the absolue value of the error is represented. The
legend goes from 0 to 0.25 meters and colors change according to the gravity of the
error.
Zone with lowest error, highlighted in blue, is the zone among the anchors. Error is
greater outside this area.

3 anchors, second configuration results

Figure 3.4: 2D plane and 3 anchors (second) configuration measurements.

In figure 3.4 is shown the second anchor constellation, where anchors form a rect-
angle triangle. Even in this case clusters of points are more accurate in the core of
the highlighted area, so their average point is closer to the ideal point(green dots).
Respect to the first configuration the clusters and error distributions are similar,
the only difference is the orientation of the clusters. While in the first case they are
oblique with a positive slope, in this case is the opposite.

The following table shows the error on the x and y axis and the absolute error
calculate as the euclidean distance between the average point of the cluster and the
ideal point.

42
 3 – Study of the TREK1000 system and errors

x[m] y[m] X err [m] Y err [m] Abs err [m]

0 1 -0,1618 0,0764 0,1789
0 2 -0,1637 0,1501 0,2221
0 3 -0,1253 0,0309 0,1291
0 4 -0,0992 0,0196 0,1011
1 0 -0,0391 -0,0556 0,0680
1 1 -0,1138 0,0881 0,1439
1 2 -0,1004 0,0587 0,1163
1 3 -0,1402 0,1330 0,1932
1 4 -0,0961 0,1129 0,1483
2 0 -0,0425 0,0627 0,0757
2 1 -0,0543 -0,0404 0,0677
2 2 -0,0532 0,0599 0,0801
2 3 -0,0905 0,1392 0,1660
2 4 -0,0671 0,1656 0,1786
3 0 0,0198 -0,0488 0,0526
3 1 -0,0635 0,0326 0,0714
3 2 -0,0249 0,0530 0,0585
3 3 -0,0116 0,0920 0,0927
3 4 -0,0394 0,1964 0,2003
4 0 0,1105 -0,0071 0,1107
4 1 0,0467 -0,0413 0,0623
4 2 0,0412 0,0315 0,0518
4 3 -0,0011 0,1382 0,1382
4 4 0,0230 0,1662 0,1678
5 0 0,1135 0,0080 0,1138
5 1 0,0944 0,0325 0,0998
5 2 0,1413 0,0699 0,1576
5 3 0,0712 0,1360 0,1535

43
 3 – Study of the TREK1000 system and errors

5 4 0,0498 0,1933 0,1996

6 1 0,1246 0,0481 0,1335
6 2 0,1005 0,0272 0,1041
6 3 0,0852 0,0907 0,1244
max err[m] 0,2221 σ[m] 0,0487
RMSE [m] 0,0893

Table 3.2: Static errors of the anchor constellation (b) figure 3.1.

Error values are similar to the previos case. Both maximum error an RMSE have
the same order of magnitude of the first arrangement.

4 0.25 0.25

3.5
0.2 0.5 0.2
3

0.4
2.5
0.15 0.15

2 0.3

0.1 0.1
1.5 0.2

1 0.1
0.05 0.05
4
0.5
0 2
0 1
0 0 2 3 0
4 5 0
0 1 2 3 4 5 6 6

(a) (b)

Figure 3.5: 3D surface error for constellation (b) figure 3.1.(a) top view, (b) side
view.

Again the blue zone is the one among the anchors. The behavior of this constel-
lation is similar to the first one considering the error distribution.

44
 3 – Study of the TREK1000 system and errors

3 anchors, third configuration results

Figure 3.6: 2D plane and 3 anchors (third) configuration measurements.

In figure 3.6 is shown the third anchor constellation, where anchors form a isosceles
triangle. Respect to the first two cases this time the configuration is very symmet-
rical and the cluster, especially in the center zone, are tight and round.

The following table shows the error on the x and y axis and the absolute error.

x[m] y[m] X err [m] Y err [m] Abs err [m]

0 1 -0,1441 0,0558 0,1545
0 2 -0,1483 0,0837 0,1703
0 3 -0,1084 0,0877 0,1394
0 4 -0,0887 0,0920 0,1278
1 0 -0,0509 0,0529 0,0734
1 1 -0,1409 0,0998 0,1726
1 2 -0,1105 0,0669 0,1292
1 3 -0,1467 0,1387 0,2019
1 4 -0,0756 0,1669 0,1832
2 0 -0,0441 0,1248 0,1323

45
 3 – Study of the TREK1000 system and errors

2 1 -0,1129 0,1367 0,1773

2 2 -0,0539 0,0873 0,1026
2 3 -0,0968 0,1312 0,1631
2 4 -0,0835 0,1699 0,1893
3 0 0,0387 0,1296 0,1353
3 1 -0,0543 0,1234 0,1349
3 2 -0,0196 0,0865 0,0887
3 3 -0,0118 0,1145 0,1151
4 0 0,1022 0,1037 0,1456
4 1 0,0231 0,0678 0,0716
4 2 0,0045 0,0758 0,0759
4 3 -0,0244 0,1051 0,1079
4 4 0,0237 0,1740 0,1756
5 0 0,1124 0,0945 0,1468
5 1 0,0894 0,0722 0,1149
5 2 0,1229 0,1242 0,1747
5 3 0,0521 -0,0224 0,0567
5 4 0,0413 0,1416 0,1475
6 1 0,1150 0,0991 0,1518
6 2 0,1110 0,0902 0,1430
6 3 0,0732 0,0816 0,1097
6 4 0,0664 0,0054 0,0667
max err[m] 0,2019 σ[m] 0,0389
RMSE [m] 0,0884

Table 3.3: Static errors of the anchor constellation (c) figure 3.1.

The shape of the blue zone is quite triangular. Looking at the errors is possible
to notice a certain symmetrical behavior respect to the 2D plane. this means that
with a symmetric constellation the behavior of the error in the space is predictable.

46
 3 – Study of the TREK1000 system and errors

4 0.25 0.25

3.5
0.2 0.2
3 0.5

2.5 0.4
0.15 0.15

2 0.3

0.1 0.2 0.1

1.5

1 0.1
4
0.05 0.05
3
0.5 0
0 2
2 1
0 0 4 0
0 1 2 3 4 5 6 6 0

(a) (b)

Figure 3.7: 3D surface error for constellation (c) figure 3.1. (a) top view, (b) side
view.

4 anchors configuration results

Figure 3.8: 2D plane and 4 anchors configuration measurements.

In figure 3.8 is shown the fourth anchor constellation, where anchors form a rectan-
gle. Respect to the first three cases this time the configuration exploits 4 anchors in
a symmetrical position. The fourth anchor placed in the top right corner is 1.64m
tall, on a different plane rispect to the other anchors. This time clusters of points are

47
 3 – Study of the TREK1000 system and errors

round and compact and the error reaches its minimum in the middle of the 2D plane

The following table shows the error on the x and y axis and the absolute error.

x[m] y[m] X err [m] Y err [m] Abs err [m]

0 1 -0,2097 -0,0544 0,2167
0 2 -0,1906 -0,0381 0,1943
0 3 -0,2282 -0,0173 0,2289
1 0 -0,1163 -0,0285 0,1198
1 1 -0,1996 -0,0369 0,2030
1 2 -0,1969 -0,0297 0,1991
1 3 -0,2100 0,0053 0,2101
1 4 -0,1856 0,0335 0,1887
2 0 -0,0769 -0,0160 0,0785
2 1 -0,1079 0,0226 0,1103
2 2 -0,1422 -0,0422 0,1484
2 3 -0,1706 -0,0413 0,1755
2 4 -0,1832 0,0016 0,1832
3 0 -0,0066 0,0128 0,0144
3 1 -0,0659 0,0030 0,0659
3 2 -0,0809 0,0007 0,0809
3 3 -0,0602 -0,0063 0,0606
3 4 -0,0512 0,0515 0,0727
4 0 0,0709 -0,0222 0,0743
4 1 -0,0176 0,0125 0,0216
4 2 0,0211 0,0334 0,0395
4 3 -0,0176 0,0400 0,0437
4 4 0,0204 0,0181 0,0273
5 0 0,0847 0,0521 0,0994
5 1 0,0411 -0,0074 0,0418
5 2 0,0873 0,0251 0,0909
5 3 0,0466 0,0518 0,0696

48
 3 – Study of the TREK1000 system and errors

5 4 0,0517 0,0279 0,0588

6 1 0,1009 0,0194 0,1028
6 2 0,1538 0,0523 0,1625
6 3 0,1141 0,0236 0,1165
max err[m] 0,2289 σ[m] 0,0660
RMSE [m] 0,0894

Table 3.4: Static errors of the anchor constellation d figure 3.1.

4 0.25 0.25

3.5
0.5
0.2 0.2
3
0.4
2.5
0.15 0.15
0.3
2

0.1 0.2 0.1

1.5

1 0.1
0.05 0.05

0.5 4
0
0 2
0 0 2 0
4 0
0 1 2 3 4 5 6 6

(a) (b)

Figure 3.9: 3D surface error for constellation (d) figure 3.1. (a) top view, (b) side
view.

With this configuration almost all the area among the anchors is blue, this means
low error. This is probably the best configuration considering the compact clusters
and the low error zone size.

Changing configuration doesn’t change the maximum or the RMSE, but using sym-
metrical constellation clusters are more compact and the error has a symmetric and
predictable behavior on the 2D plane. This important fact will be exploited in the
fourth chapter.
In the following section the oscillatory behavior of clusters is analyzed.

49
 3 – Study of the TREK1000 system and errors

3.1.2 Oscillation distribution respect to anchor’s position in

the space
The aim of this experiment is to analyse the oscillation of the tag position during
static measures. In order to do so, the set up is the same of the previous experiment,
a grid of 6 meters x 4 meters is built on the floor. The anchors are positioned at the
vertices of this rectangular grid using 1,17 m tall tripods, while the tag is moved(on
the floor) from measure to measure inside and on the side of this rectangle and on
the vertices of the small square of 1m x 1m size forming the rectangle.

Four different anchor configurations have been investigated in order to figure out
which is the best solution, analyzing oscillation distributions.
In order to figure out which kind of oscillation affects the tag, for each point of
the grid, the 2D plane is divided in 9 zones. For each cluster of points obtained
during a measurement in a test point, the average value is calculated, and placed
in the center of the plane division. Then each measure of the cluster is compared
respect to the average to determine if it is moved and in which direction.

Figure 3.10: 2D plane division. The centre is represented by the average position
of the measurement cluster. Blue lines represent limit boundaries to discriminate
between different zones.

As it is possible to see from the picture above, some zones are greater than others,
because the aim of this test is to investigate significant oscillations, so the small ones
around the average point are not considered. Blue lines delimit zone boundaries, in

50
 3 – Study of the TREK1000 system and errors

this case these lines are 1 cm far from the axis.

There are 5 kinds of oscillation:

• vertical;

• horizontal;

• positive slope;

• negative slope;

• neutral.

Vertical Vertical oscillation is when there is a vertical movement among subse-

quent measurements. According to the plane division a vertical oscillation occurs
when the tag’s position moves :

• from 2 to 8 or from 8 to 2;

• from 2 to 5 or from 5 to 2;

• from 8 to 5 or from 5 to 8;

• from 1 to 7 or from 7 to 1;

• from 3 to 9 or from 9 to 3.

Horizontal Horizontal oscillation is when there is an horizontal movement among

subsequent measurements. According to the plane division a vertical oscillation
occurs when the tag’s position moves :

• from 4 to 6 or from 6 to 4;

• from 6 to 5 or from 5 to 6;

• from 4 to 5 or from 5 to 4;

• from 1 to 3 or from 3 to 1;

• from 7 to 9 or from 9 to 7.

51
 3 – Study of the TREK1000 system and errors

Positive slope Positive slope oscillation is when there is an oblique movement

among subsequent measurements with a positive cosine value of the angle obtained
respect to the horizontal axis. According to the plane division a vertical oscillation
occurs when the tag’s position moves :

• from 3 to 7 or from 7 to 3;

• from 3 to 5 or from 5 to 3;

• from 7 to 5 or from 5 to 7.

Negative slope Negative slope oscillation is when there is an oblique movement

among subsequent measurements with a negative cosine value of the angle obtained
respect to the horizontal axis. According to the plane division a vertical oscillation
occurs when the tag’s position moves :

• from 1 to 9 or from 9 to 1;

• from 1 to 5 or from 5 to 1;

• from 9 to 5 or from 5 to 9.

Neutral When there are no movement of the tag’s position, it is considered a

neutral movement, so is not counted as an oscillation.

In order to evaluate the oscillations, each measurement has been scheduled in

tables. Each row represent a point of the grid, and each type of oscillation is ex-
pressed in percentage.

3 anchors, first configuration results

From picture 3.2 is possible to see how clusters tend to be dense and compact in
the highlighted zone, while outside (outside the zone covered by anchors) cluster

52
 3 – Study of the TREK1000 system and errors

tend to elongate. The other key point is that cluster are closer to test points in
the orange area. It means that the system accuracy is better in this zone. Cluster
interpolations have a positive slope and tend to flat in the selected zone.

Oscillations are analysed in the following table:

Position Vertical [%] Horizontal [%] Positive slope [%] Negative slope [%] Neutral [%]
(0,0) anchor 0 anchor 0 anchor 0 anchor 0 anchor 0
(0,1) 1,72 3,45 1,03 0,00 93,80
(0,2) 1,35 7,07 1,68 0,34 89,56
(0,3) 1,63 4,58 2,61 0,00 91,18
(0,4) anchor 1 anchor 1 anchor 1 anchor 1 anchor1
(1,0) 4,78 3,75 1,02 0,34 90,11
(1,1) 9,21 2,96 0,98 0,66 86,19
(1,2) 3,70 3,03 1,01 0,00 92,26
(1,3) 3,40 2,72 1,02 0,00 92,86
(1,4) 2,69 2,02 8,75 0,00 86,54
(2,0) 8,85 1,97 1,96 0,00 87,22
(2,1) 5,21 1,63 6,19 0,00 86,97
(2,2) 5,30 2,65 12,58 0,66 78,81
(2,3) 4,12 1,37 7,56 0,00 86,95
(2,4) 4,62 3,30 8,58 0,00 83,50
(3,0) 8,45 1,35 3,04 0,00 87,16
(3,1) 4,15 4,79 7,98 0,00 83,08
(3,2) 2,04 0,34 3,74 0,00 93,88
(3,3) 4,30 6,62 3,31 0,33 85,44
(3,4) 3,09 3,09 9,96 0,00 83,86
(4,0) 16,15 2,17 1,55 0,00 80,13
(4,1) 7,86 1,89 3,14 0,00 87,11
(4,2) 4,45 1,71 5,48 0,00 88,36
(4,3) 6,71 4,03 1,34 0,00 87,92
(4,4) 4,27 1,97 14,75 0,00 79,01
(5,0) 13,07 0,98 1,63 0,00 84,32

53
 3 – Study of the TREK1000 system and errors

(5,1) 7,00 0,96 6,05 0,00 85,99

(5,2) 9,32 1,61 3,21 0,00 85,86
(5,3) 5,02 0,67 3,34 0,00 90,97
(5,4) 12,86 2,89 3,54 0,00 80,71
(6,0) anchor 2 anchor 2 anchor 2 anchor 2 anchor 2
(6,1) 4,39 2,02 9,80 0,34 83,45
(6,2) 4,39 2,03 9,80 0,33 83,45
(6,3) 7,24 0,99 8,88 0,00 82,89
(6,4) 8,61 5,63 5,30 0,66 79,80

Table 3.5: Oscillation table referred to constellation (a) in figure 3.1.

In the table are reported the percentage of oscillations, for each row the highest
percentage (excluded the neutral) is highlighted in red. Results show that oscilla-
tions in this configuration are most vertical or in the positive slope direction.

3 anchors, second configuration results

Once again, from figure 3.4, clusters are more focused in the highlighted zone than
respect to the rest. In this case cluster interpolations have a negative slope, opposite
behaviour respect to the first configuration, and again they tend to flat among the
anchors.

Position Vertical [%] Horizontal [%] Positive slope [%] Negative slope [%] Neutral [%]
(0,0) anchor 0 anchor 0 anchor 0 anchor 0 anchor 0
(0,1) 11,92 0,66 0,00 2,32 85,10
(0,2) 6,98 0,33 0,00 11,30 81,39
(0,3) 5,24 2,10 0,00 7,69 84,97
(0,4) 6,80 2,59 0,00 13,92 76,69
(1,0) 0,92 0,31 0,00 16,00 82,77
(1,1) 9,06 2,35 0,00 7,05 81,54
(1,2) 5,61 0,99 0,33 9,24 83,83
(1,3) 7,57 0,99 0,00 4,28 87,16
(1,4) 8,53 8,53 0,00 4,78 78,16

54
 3 – Study of the TREK1000 system and errors

(2,0) 6,43 0,96 0,00 5,15 87,46

(2,1) 7,28 1,98 0,33 11,26 79,15
(2,2) 3,69 4,03 0,34 2,35 89,59
(2,3) 7,90 3,78 0,34 8,59 79,39
(2,4) 7,67 5,75 1,28 6,39 78,91
(3,0) 8,33 3,67 1,00 1,00 86,00
(3,1) 8,55 0,00 0,00 0,33 91,12
(3,2) 5,28 5,61 0,66 7,26 81,19
(3,3) 3,95 2,96 0,00 11,51 81,58
(3,4) 2,88 0,64 0,00 11,22 85,26
(4,0) 5,79 0,32 0,00 0,96 92,93
(4,1) 5,48 1,94 0,00 3,23 89,35
(4,2) 5,56 2,94 0,00 1,96 89,54
(4,3) 3,99 0,66 0,00 4,32 91,03
(4,4) 5,03 5,03 1,34 2,68 85,92
(5,0) 5,23 5,88 0,98 1,96 85,95
(5,1) 14,77 1,68 1,34 0,34 81,87
(5,2) 6,00 2,67 0,33 1,00 90,00
(5,3) 2,94 2,29 0,65 1,31 92,81
(5,4) 3,98 0,92 0,00 5,50 89,60
(6,0) anchor 1 anchor 1 anchor 1 anchor 1 anchor 1
(6,1) 6,05 5,10 0,64 0,96 87,25
(6,2) 4,22 7,47 0,32 0,65 87,34
(6,3) 2,96 4,28 0,33 0,33 92,10
(6,4) anchor 2 anchor 2 anchor 2 anchor 2 anchor 2

Table 3.6: Oscillation table referred to constellation (b) in figure 3.1.

Oscillations in this configuration are directed in the vertical and negative slope
direction.

55
 3 – Study of the TREK1000 system and errors

3 anchors, third configuration results

From figure 3.6 is possible to see that clusters are more focused in the highlighted
zone than respect to the rest. In this case cluster interpolations have an horizontal
behaviour in the orange zone, while on the left they have a negative slope, the
opposite of the right side. In respect to the first two configurations cluster are more
focused and compact.

Position Vertical [%] Horizontal [%] Positive slope [%] Negative slope [%] Neutral [%]
(0,0) anchor 0 anchor 0 anchor 0 anchor 0 anchor 0
(0,1) 5,76 3,05 0,34 2,71 88,14
(0,2) 2,21 0,63 0,00 6,94 90,22
(0,3) 4,43 6,65 0,00 0,95 87,97
(0,4) 5,76 7,80 1,36 2,71 82,37
(1,0) 1,64 3,28 0,00 8,20 86,88
(1,1) 4,01 2,68 0,33 2,01 90,97
(1,2) 1,36 3,06 0,00 7,14 88,44
(1,3) 1,72 2,06 0,00 2,06 94,16
(1,4) 4,89 7,49 0,98 1,30 85,34
(2,0) 10,56 1,98 1,32 1,32 84,82
(2,1) 6,33 5,70 0,95 1,58 85,44
(2,2) 4,29 7,30 0,61 2,45 85,35
(2,3) 3,49 4,76 1,27 0,00 90,48
(2,4) 1,92 7,69 2,24 0,64 87,51
(3,0) 11,22 4,76 1,02 1,02 81,98
(3,1) 4,97 6,95 0,00 0,99 87,09
(3,2) 6,60 4,40 2,20 0,94 85,86
(3,3) 5,59 5,26 0,66 0,99 87,50
(3,4) anchor 2 anchor 2 anchor 2 anchor 2 anchor 2
(4,0) 7,12 0,34 0,68 1,02 90,84
(4,1) 8,25 1,65 2,64 0,00 87,46
(4,2) 4,00 3,00 1,33 0,00 91,67
(4,3) 2,96 4,61 2,30 0,33 89,80

56
 3 – Study of the TREK1000 system and errors

(4,4) 3,50 5,73 2,54 0,00 88,23

(5,0) 3,32 2,66 1,33 0,00 92,69
(5,1) 12,75 3,59 1,63 0,33 81,70
(5,2) 3,10 3,73 1,24 0,00 91,93
(5,3) 4,00 4,33 1,67 0,00 90,00
(5,4) 5,36 6,62 1,26 0,63 86,13
(6,0) anchor 1 anchor 1 anchor 1 anchor 1 anchor 1
(6,1) 7,24 1,97 2,96 0,66 87,17
(6,2) 6,78 3,73 3,39 0,34 85,76
(6,3) 3,63 3,30 2,31 0,33 90,43
(6,4) 14,18 5,32 3,55 2,48 74,47

Table 3.7: Oscillation table referred to constellation (c) in figure 3.1.

Oscillations have an horizontal and vertical behaviour, this explain why cluster
are so focused and they didn’t extend in a oblique way.

4 anchors configuration results

In this configuration, as in the third, clusters are focused and close to test points.
Cluster interpolation are most in the horizontal direction.

Position Vertical [%] Horizontal [%] Positive slope [%] Negative slope [%] Neutral [%]
(0,0) anchor 0 anchor 0 anchor 0 anchor 0 anchor 0
(0,1) 7,00 6,67 0,00 1,00 85,33
(0,2) 7,02 4,68 0,67 0,00 87,63
(0,3) 6,57 5,54 0,00 3,81 84,08
(0,4) anchor 2 anchor 2 anchor 2 anchor 2 anchor 2
(1,0) 3,91 3,26 1,95 0,00 90,88
(1,1) 4,89 1,95 0,00 0,00 93,16
(1,2) 9,76 2,69 0,34 0,00 87,21
(1,3) 3,17 3,17 0,00 2,46 91,20
(1,4) 4,32 3,32 0,33 0,00 92,03

57
 3 – Study of the TREK1000 system and errors

(2,0) 9,84 1,59 0,00 0,32 88,25

(2,1) 4,71 1,01 0,00 0,00 94,28
(2,2) 7,64 2,66 1,33 0,33 88,04
(2,3) 6,50 0,68 9,25 0,00 83,57
(2,4) 5,14 2,05 5,14 0,00 87,67
(3,0) 8,67 2,48 2,78 0,00 86,07
(3,1) 6,02 3,01 0,67 0,00 90,30
(3,2) 6,67 2,00 0,00 0,00 91,33
(3,3) 6,25 2,78 1,39 0,00 89,58
(3,4) 5,41 1,35 1,35 0,34 91,55
(4,0) 5,08 2,03 0,00 0,34 92,55
(4,1) 8,79 0,65 0,32 0,00 90,24
(4,2) 5,25 1,64 0,00 0,00 93,11
(4,3) 5,61 4,62 4,29 1,32 84,16
(4,4) 4,97 1,98 0,33 0,99 91,73
(5,0) 8,11 1,48 1,10 0,00 89,31
(5,1) 5,48 0,00 0,97 0,00 93,55
(5,2) 8,31 1,99 2,32 0,00 87,38
(5,3) 8,97 1,66 0,00 2,33 87,04
(5,4) 6,91 0,66 1,97 0,33 90,13
(6,0) anchor 1 anchor 1 anchor 1 anchor 1 anchor 1
(6,1) 5,90 1,96 3,61 0,33 88,20
(6,2) 5,30 1,99 2,98 0,00 89,73
(6,3) 9,33 4,33 1,33 0,00 85,01
(6,4) anchor 3 anchor 3 anchor 3 anchor 3 anchor 3

Table 3.8: Oscillation table referred to constellation (d) in figure 3.1.

In this configuration oscillation are almost completely in the vertical direction.

This allows cluster to be compact.

All the configurations are characterized by a blue low error area located in the

58
 3 – Study of the TREK1000 system and errors

middle of the grid. This means that for the first two configurations this area is
asymmetric respect to the anchors distribution.
Among the 4 configurations tested in this experiment, the last two are better respect
to the rest from the cluster concentration point of view. 3 anchors, third configura-
tion has focused cluster, and his low error zone is placed at the centre of the area
covered by anchors, while in the first two configurations this area is asymmetric.
Furthermore oblique oscillations are drastically reduced.
4 anchors configuration is most likely the most precise according to its low error zone
size and to the fact that the oscillation are only in the vertical direction allowing
clusters to be more focused.

3.2 Dynamic error

The study of the dynamic error is related to the ability of the TREK1000 develpment
kit to track a moving tag. Since the final application has to track a robot and a
patient is foundamental to prove that this system is able to track till a certain speed.
The test concern the use of a pendulum with variable length. The tag attached to
the end of the pendulum is left free to oscillate and the speed of the movement is
varied changing the angle of oscillation.
Since the Home Robot can travel at maximum speed of 0.5 ms , a very low speed,
the limit is represented by the patien’s speed. A persor walks with an average speed
that varies from 1 ms to 1.5 ms , so this is the upper limit that the system must reach.

3.2.1 Pendulum
The movement of the pendulum is periodic and deterministic. The choice to use a
pendulum is explained in the fact that this kind of movement emulates the one done
by a person that is walking.
The set up for this experiment is composed by the 4 EVB1000 of the development
kit and the pendulum represent by a tiny rope 2.24 m long attached to the roof.
The anchors are positioned in:

• anchor 0 (0.0m; 0.0m, 0.0m);

59
 3 – Study of the TREK1000 system and errors

• anchor 1 (1.80m; 0.0m, 0.0m);

• anchor 2 (0.0m; 3.0m, 0.0m);

• tag in resting position (0.9m; 1.50m, 0.76m)

The following chapter presents some measurements done, pendulum oscillation

in time and the Root Mean Square Error (RMSE) of the oscillation peaks. Four
measurements are performed keeping constant the oscillation period and varying the
amplitude of the oscillation, in this way the average speed changes.

The damping factor is evaluated using an exponential interpolation of the oscil-

lation peaks as time constant of the function:

−t
d · exp( τ )
(3.1)

where τ is the damping factor.

The table is composed of:

• L represents the lenght of the pendulum;

• T is the period;

• d is the oscillation amplitude;

• v is the average speed estimated like T · 1/2 · 1/d;

• θ the angle between the pendulum and the vertical axis at the moment of the
release of the pendulum;
sinθ
• θ
index of the validity of the pendulum model;

• τ damping time constant.

L [m] T [s] d [m] v [ ms ] θ [◦ ] sinθ

θ
τ [s]
2.24 3 0.75 1 18.51 0.983 54.1
2.24 3 1 1.333 24.06 0.971 43.8
2.24 3 1.10 1.467 26.15 0.966 51.2
2.24 3 1.20 1.60 28.18 0.960 41.7
60
 3 – Study of the TREK1000 system and errors

Model limits The limits of this measurements are due to the model and the pen-
dulum movement.

In order to the model to be valid, the ratio sinθθ

must be as much as possible equal
to one, this limits the ascillation angles and so the amplitude that means low speed.
The other limit is represented by the movement done by the pendulum and on the
fact that the speed of it depends on the angle. The speed is maximum in the middle
of the oscillation, when the pendulum converts all the potential energy in kinetic
energy, while it is null at peaks. This is an advantage for the TREK1000 since it
can better sample the peaks where the pendulum first decelerate, then stops and
then accelerate again, while between one peak and the other the number of sample
per unit length is lower.
Last limit is represented by the damping factor that reduces the amplitude of the
oscillations in time.

In figure 3.11 are depicted different amplitude oscillation. Number of samples on

the x axis, while amplitude in meters is placed on y axis. Each red cross reprsents
one sample, while the blue lines are the exponential interpolations of the peak values
from which the damping factor is obtained. The amplitude in time gives that sinu-
soidal shape represented in the picture. No filters or any kind of data eleboration is
applied to the amplitude values, they are exactly as the system sampled them.

This experiment proves that the TREK1000 is able to track a moving tag from 1
m
s
to 1.6 ms respect to all limit presented by the model. Considering that iRobot has
a maximum velocity of 0.5 ms and the patient in the apartment cannot move around
at a speed greater than 1.5 ms , this system is suitable for the final application.

61
 3 – Study of the TREK1000 system and errors

Amplitude oscillation 1ms

3

2.5

2
amplitude [m]

1.5

0.5

0
120 140 160 180 200 220
sample

(a)
Amplitude oscillation 1.33 m/s
3

2.5

2
amplitude [m]

1.5

0.5

0
80 100 120 140 160 180 200 220 240 260 280
sample

(b)

Amplitude oscillation 1.47 m/s

3

2.5

2
amplitude [m]

1.5

0.5

0
150 200 250 300
sample

(c)
Amplitude oscillation 1.6 m/s
3

2.5

2
amplitude [m]

1.5

0.5

0
80 100 120 140 160 180 200 220 240 260 280
sample

(d)

Figure 3.11: Amplitude oscillation for different release angles.

62
Chapter 4

Positioning Algorithm

This section is dedicated to the description of the positioning algorithm developed

using MatLab environment. All the parts that compose the final software are ex-
plained step by step presenting the problem and the relative solution.

4.1 Trilateration algorithm

Analyzing the trilateration algorithm implemented by Decawave and explained in

chapter II, section ”Positioning algorithm”, is possible to identify the limits that
affect its performance. In fact this code exploits only three anchors for the trilater-
ation process, while the fourth, when used, selects the right solution among the two
obtained by the spheres intersection. Furthermore the first three anchors must be
positioned at the same height, on the same z plane.

These constrains are limits since don’t allow the code to use more than 3 anchors for
the trilateration, so no redundant information, and intodruce a physical constrain on
the anchor’s arrangement. Since the final application must be used in apartment,
it would be better to have a certain freedom degree on the anchor’s placement.
This means having a system that the customer can place in his home environment
without the help of technicians. This include also a calibration process that will be
discussed in chapter VI.
In order to solve this problem a new positioning algorithm is used. Many different
types of algorithm are used for positioning problem [9], [10], [3]. After consulting
many papers and different types of algorithm a closed form one has been chosen.
Since it’s important to mantain the same performance in terms of sampling fre-
quency, the use of a closed form solution is preferred respect to a iterative one as

63
 4 – Positioning Algorithm

proposed in [17]. In this way the position can be identified in just one code cy-
cle. This positioning algorithm exploits a statistic approach by formulating an error
function and finding the minimum of this one.

4.1.1 Closed-form Localization from Range-difference mea-

surements

N denotes the number of anchors, while dij denote the range difference (RD) between
anchors i and j with i and j varying from 1 to N. The positioning vector(x; y; z) of
the i-th anchor is denoted as xi and the unknown position of the tag is denotes as
xs .
The distance between the i-th anchor and the tag is Di = ||xi − xs || and the distance
from the origin of the cartesian axis to the position of the i-th anchor is denoted as
Ri .
The range differences are denoted as:

dij = Di − Dj (4.1)

where i = 1, . . . , N and j = 1, . . . , N.
The localization problem is to find the tag position xs given dij . In total there are :

N (N − 1)
(4.2)
2

distinc range differences. The formula above calculate all the possible combina-
tion of rage difference combining anchors excluding i=j and each equal pair dij = -
dij .
The position of the tag is estimated as the one that best fits the RD measurements.
It is found in closed form considering the tag’s position as the minimizer of the error
equation, so the minimizer of the difference between functions of RDs and function
of the hypothesized source location.

64
 4 – Positioning Algorithm

Figure 4.1: 3D scheme of the geometric distances between anchors (sensors) and tag
(source).

First step is to map the origin of the cartesian space as the position of one anchor.

Rj = 0
xj = 0 → (4.3)
Dj = Rs

From Pythagorean theorem and using 4.3 in 4.1

(Rs + dij )2 = Ri2 − 2xTi xs + Rs2 (4.4)

and

0 = Ri2 − d2ij − 2Rs dij − 2xTi xs (4.5)

Considering equation 4.5 involving the N - 1 range differences relative to anchor

j, N - 1 equations in the three unknowns are obtained for the position xs .
The delays affecting the range measurements is not null and is not negligible, so the
so called error equation is introduced on the lefthand of eq. 4.5. Then is possible
to obtain a position estimation of xs minimizing the equation using a least square
approach.

65
 4 – Positioning Algorithm

Let be j = 1. Then 4.5 is:

i = Ri2 − d2ij − 2Rs dij − 2xTi xs i = 2,3, . . . ,N (4.6)

where i is the error equation to be minimized. Using the matrix notation:

 = δ − 2Rs d − 2Sxs (4.7)

where

     
R22 − d221 d21 x2 y2 z2
 2
 R3 − d231  d31   x3 y3 z3 
    

δ= ..  d= . 
 S=
 .. .. ..  (4.8)
.  ..   . . . 
   
 
2
RN − d2N 1 dN 1 xN yN zN

Error equation 4.10 is linear respect to xs and to Rs . The closed form solution
is:

1
xs = Sw∗ (δ − 2Rs d) (4.9)
2

where

Sw∗ = (S T S)−1 S T (4.10)

So this algorithm use one anchor as reference point, placing the origin of the axis
in its position. Every range measurement is transformed in a range difference respect
to the reference anchor, then an error equation is formulated considering the fact
that measurements are affected by error. Minimizing this equation an estimation of
the tag’s location is obtained. The validity of this algorithm is proved in [12].
The advantages of using this approach is that now the localization process can
exploit multiple anchors in a redundant way and the anchors can be positioned even
at different height.

66
 4 – Positioning Algorithm

4.2 Corrective function

Looking at the static measurements presented in chapter III, is possible to notice
that the behavior of the anchors is affected by errors. Figure shows the static
measurements obtained on a surface of 6m x 4m using 4 anchors and the positioning
algorithm presented in the previous section. Anchors positioned in:

• anchor 0 (0.0m; 0.0m, 0.09m);

• anchor 1 (6.0m; 0.0m, 0.09m);

• anchor 2 (0.0m; 4.0m, 0.09m);

• anchor 3 (6.0m; 4.0m, 0.09m);

• tag placed on a tripod 1.64m tall.

The tag is moved on each green dot that represents a test point, while the blue point
are clusters of measurements obtained for each test point.

Figure 4.2: Static measurements obtained using 4 anchors and the closed-form al-
gorithm

Taking in consideration one random anchors is possible to notice a cerrtain be-

havior of the measurement clusters. For example let’s consider anchor 0, positioned

67
 4 – Positioning Algorithm

in (0.0, 0.0) in the figure, as reference node, how is possible to see the clusters near
this node are underestimated respect to the relative test points, while distant clus-
ters are overestimated. This problem is related to the fact that ranges coming from
anchors tend to be underestimated for values smaller than 2m, while are overes-
timted for values greater than 2m. Ranges around 2 m are accurate respect to the
real value.
In order to compensate this error a corrective function is built.

(Ideal distance)/(Measured distance)

3
measured points
interpolating curve

2.5
multiplicative factor

1.5

0.5
0 1 2 3 4 5 6 7 8 9 10
Distance [m]

Figure 4.3: Corrective function.

Figure 4.3 shows the corrective function. This function corrects the range value
obtained as a measure from the Decawave TWR process. It is basically a multi-
plicative factor that multiplies the range correcting its value. Since the range is
underestimated for values smaller than 2m and viceversa for values greater than
2m, the multiplicative factor is greater than 1 before 2m and smaller than 1 over
2m. This means that the corrective function is function of the only range value, so
it is a one dimensional function.

range0 = range · f (range) (4.11)

In order to build this function an anchor and a tag is used. Placing the anchor in
a fixed point and moving the tag by steps of 25cm far from the anchor on a straight

68
 4 – Positioning Algorithm

line, range values have been recorded using Log File. This process is done starting
25cm far from the anchor till 10m far, for different orientation of the antennas. Four
different antenna configuration have been used. First the antennas in a parallel con-
figuration, each antenna directed toward the other. Then the antenna of the tag is
rotated of 90 degrees in order to be perpendicular to the other. Third configuration
with both the antenna rotated of 90 degrees, and finally the tag’s antenna directed
toward the anchor and the anchor’s one rotated of 90 degree.
For each configuration a several measurements have been recorded. In figure 4.3
for each step, 4 blue crosses are reported, one for each antenna configuration. Each
blue cross represents the average value of a cluster of points obtained for that step.

The function is obtained dividing the ideal (real) distance values by the measured
ones, and then interpolating those points. It is one dimensional since it only needs
range value as input and gives the multiplicative factor as output. Multiplying the
latter for the initial rage value, the corrected one is obtained.

Results and performance of this operation are presented in the next chapter.

4.3 Double system: 8 anchors

In order to increase the redundant information and try to increase the accuracy of
the positioning process another TREK1000 development kit is used for a total of 8
anchors. The use of more than 4 anchors allows to go deeper into the study of the
localization problem and permit to develop new techniques able to mitigate errors
due to problems that will be explained in the following sections such as the No Line
Of Sight (NLOS) one.

Since the system is designed to allocate a maximum of 4 anchors and also the TWR
algorithm allows only this number of nodes a trick has been used. The system has
two working channels, channel 2 around the frequency of 3.6 GHz and channel 5 of
frequency 6.5 GHz. The trick consists in using 4 anchors working on channel 2 and
the other 4 on channel 5, while the tag switches between these two in order to collect
measurements and save them in Log File format. Then using a MatLab script the

69
 4 – Positioning Algorithm

measurements are analyzed and using the closed-form localization seen above, the
position of the tag is estimated.

For the following experiments the anchors are distributed in a symmetric way, form-
ing an octagon (considering the 2D plane). From now on all the techniques and
methods developed are applied to a system with redundant information, in paticu-
lar with 8 anchors.

4.4 Best anchors filter

As explained in chapter I, ”positioning process” section, the spheres intersection
produces an uncertainty area in which the tag is situated. Using 8 anchors and so
8 range values for the localization problem could lead to a better accuracy, but in
some cases a range value could expand the uncertainty area. This could be caused
by the non ideal position of the tag respect to the tag position, or a peak of noise
component during the measure. In order to tight this uncertainty area a simple
algorithm has been developed.

The approach used is iterative. It consists in non considering one of the range
value and use the others to obtain the tag position. Once estimated the latter,
errors are calculated as difference between the euclidean distance between the tag’s
position and the anchors’ one and the range value. This error is calculated for each
node and the process is repeated neglecting one anchor, and one range value, at
turn. Once identified the range measurement responsible of the bigger error, this
one is eliminated from the positioning algorithm.
This approach is repeated until the maximum error is below a certain static thresh-
old, or the number of remaining range values is 4 (the minimum numbe of anchors
for 2D positioning plus one redundant anchor).

Even if the positioning algorithm is statistical one beased on error equation,

so not based on geometrical and spheres intersection, this approach to select the

70
 4 – Positioning Algorithm

best range values is still valid. In fact instead of tightening the uncertainty area
it minimize the error equation neglecting terms that are affected by a greater error
respect to the others.

4.5 Offset correction

Like every positioning algorithm, even the ”Closed-form Localization from Range-
difference measurements” one introduces some bias and variance in the tag position.
This error derives from linearization and approximation operation that simplifies
the mathematical model, but are an approximated representation of the reality.

The variance is mitigated thanks to a median operation performed on the tag posi-
tion results. The bias instead, is related to which anchor is used as the refence node
in the algorithm. So in order to mitigate this error every anchor is used as reference
one and then the tag position is obtained by means of an average operation. In this
way the bias factor is compensated and redistributed in the 2D plane.
Results and performance are shown in the next chapter.

4.6 NLOS correction

No Line Of Sight (NLOS) is one of the major problem for localization system. If
the direct path between an anchor and the moving tag is obstructed from an object,
obstacle or even a person, the ToA of the signal is delayed, which means that a pos-
itive bias is introduced. The first step in NLOS mitigation is identify the presence
of this problem in the measurements obtained, then try to mitigate it.
In general NLOS identification is done by means of probabilistic techniques [18],
working with probability density function, noise variance and static thresholds. One
of the aim of this thesis is to keep a simple approach to the positioning problem, so
the probabilistic way is discarded and a more simpler method is preferred.

As discussed above, NLOS introduces a bias factor in the range value, so this means
that there is a variation in the range measure when an obstacle moves between the

71
 4 – Positioning Algorithm

achor and the tag. The first operation done by the algorithm is a derivative oper-
ation on the range value. In this way when the derivative value overcome a static
threshold the algorithm understands that there is presence of NLOS.
The static threshold must hold in consideration that both the robot and the patient
can move inside the home environment, so must be able to discriminate between the
variation due to a NLOS and the variation due to the movement of the tag.
Since the robot moves at a maximum speed of 0.5 ms , and the system sampling
frequency is 10 Hz, the static threshold is :

m 1
max − robot − variation = 0.5 · s = 0.05m (4.12)
s 10
The same for the patient:

m 1
max − patient − variation = 1.5 · s = 0.15m (4.13)
s 10
However considering the variance of the tag position and the error due to the
TREK1000 non synchronization, a factor of 10cm must be added to these threshold.
In this way the algorithm doesn’t recognize the error introduced by the system as
a NLOS one. So the thresold for the robot is 15cm, while the one for the patient is
25cm. Adding 10 cm on the thresholds is not a problem since the variation due to
metallic object or walls is in the order of meters.

The derivative method applied above allows the algorithm to identify variation in
the range measure that are not due to the tag movement, but to some object that is
situated between the tag and the anchor LOS. Thanks to the redundant information
coming from the eight anchors, the number of ranges for the trilateration is higher
than the minimum required, so ranges affected by NLOS are discarded from the
positioning computation.

If we consider the case in which an obstacle moves between the tag and an an-
chor and then stay still, the derivative is not able to identify it anymore. In fact the
derivative approach is valid when the range suffers a change. Let’s consider a wall,
the robot moving in the apartment can change room and an anchor that before was
in LOS, now is subjected to NLOS. The derivative notices this only in the moment

72
 4 – Positioning Algorithm

of the change of room, but then, since the wall is still, no variations occur.

After the derivative processs the NLOS mitigation algorithm performs a check on
the consistency of the range measurements. Using the position obtained during
the previous measurement session, the code produces an estimation of the future
range values. Once obtained the new measurements, these are compared with the
estimated one, and the range values that are not consistent are filtered from the
positioning process.
The estimation is obtained using the previous position and the known location of
the anchors. The difference of these two gives a range value that then is compared to
the new one, and if their difference is over a certain threshold, like in the derivative
case, they are discarded. Even in this case the threshold hold in consideration the
possibility that the tag is moving.

4.7 Algorithm overall

RANGES
CORRECTIVE
FUNCTION TAG
DERIVATIVE BEST ALGORITHM
ANCHORS + CLOSED-FORM POSITION
CONSISNTENT
RANGES OFFSET
POSITIONS LOCALIZATION
FILTER FILTER CORRECTION

Figure 4.4: Algorithm overall.

73
 4 – Positioning Algorithm

In figure 4.4 is depicted a scheme relative to the developed algorithm. On the left
are present the inputs of the algorithm that are the measured ranges, and the posi-
tion of the anchors. The first stage is represented by NLOS filter that performes the
above explained derivative and consistent filter. This stage receives as input also
the previous tag position that is represented by the arrow that goes from the output
to this stage.
The second stage is the corrective function that operates only on the ranges since
depends only on them. The third stage is a best ranges filter, that discards range
values affected by errors considered too big respect to a treshold.
The central stage is represented by the closed-form positioning process that iden-
tifies the tag positiong taking as input the value of non discarded ranges and the
relative anchor positions. The last step consists in correcting the offset introduced
by the localization algorithm.
The output is the tag position.

The algorithm has been traslated from Matlab to a java library and then imple-
mented on the existent robot software described in chapter II.

74
Chapter 5

Test and Performance

In this chapter tests performed on the algorithm exposed previously are presented.
All the tests were run in the main lab of the Motion Analysis Lab (MAL) at the
Spaulding Rehabilitation Hospistal in Boston, Massachussets.
The log Files recorded by means of the Decawave system are analyzed using MatLab
scripts.

Let’s start from a system composed by 4 anchors and one tag. Anchors positioned
in:

• anchor 0 (0.0m; 0.0m, 0.09m);

• anchor 1 (6.0m; 0.0m, 0.09m);

• anchor 2 (0.0m; 4.0m, 0.09m);

• anchor 3 (6.0m; 4.0m, 0.09m);

• tag placed on a tripod 1.64m tall.

75
 5 – Test and Performance

Figure 5.1: Static measurements obtained using 4 anchors and the closed-form al-
gorithm

Again the black diamond represent the anchors, the green dots the test points
and the blue cluster are the measured point. Figure 5.1 shows the starting point.

No correction
x[m] y[m] X err [m] Y err [m] Abs err [m]
0 1 -0,1271 -0,1899 0,2285
0 2 -0,1835 -0,0153 0,1841
0 3 -0,1836 0,0388 0,1876
1 0 -0,1200 -0,2121 0,2437
1 1 -0,1437 -0,1514 0,2087
1 2 -0,1734 -0,0569 0,1825
1 3 -0,1567 0,0437 0,1627
1 4 -0,1985 -0,0280 0,2005
2 0 -0,0626 -0,1488 0,1615
2 1 -0,0526 -0,1069 0,1191
2 2 -0,1083 -0,0420 0,1161
2 3 -0,1134 -0,0036 0,1135
2 4 -0,0688 0,0794 0,1051
3 0 0,0038 -0,1322 0,1323
3 1 0,0275 -0,1071 0,1106

76
 5 – Test and Performance

3 2 0,0022 -0,0272 0,0273

3 3 0,0014 0,0138 0,0139
3 4 -0,0054 0,0568 0,0570
4 0 0,1274 -0,1059 0,1657
4 1 0,0971 -0,0795 0,1255
4 2 0,1002 -0,0036 0,1003
4 3 0,0843 0,0187 0,0864
4 4 0,0934 0,0317 0,0987
5 0 0,1598 -0,1508 0,2198
5 1 0,1901 -0,0772 0,2052
5 2 0,2177 -0,0025 0,2177
5 3 0,1645 0,0369 0,1686
5 4 0,2086 -0,1187 0,2400
6 1 0,2222 -0,0345 0,2249
6 2 0,2291 -0,0201 0,2300
6 3 0,3514 -0,2880 0,4544
max err [m] 0,4544 σ[m] 0,0824
RMSE [m] 0,1832

Table 5.1: Static measurement error table.

Table 5.1 shows the static error obtained without any correction function or
filtering. This represents the starting point of the positioning system. At the end
of the table maximum error, Roor Mean Squared Error and variance are reported.
Colors change according to the magnitude and sign of the error as depicted in the
following table.

Legend
color limit [m]
cyan -0.10≤x<-0.05
blue X<-0.10
yellow 0.05<x≤0.10
red x>0.10

77
 5 – Test and Performance

green -0.05≤x≤0.05

Table 5.2: Color legend relative to static error.

Errors from -0.05 to 0.05 meters are green cause considering the final application,
this magnitude is acceptable. In fact since the scope is to track a person and a robot
in a home environment, high accuracy under 5 cm is not required. Errors greater
than 5 cm but smaller than 10 cm in magnitude are still accepted, but they are
borderline, while is better to avoid errors greater than 10 cm. The reason is that
knowing the position of the center of the robot with an uncertainty greater than 10
cm in magnitude could be a problem for the algorithm present on the bot which
duty is to draw the path from the actual position to the patient. Considering that
in a home environment many are the possible obstacles, an uncertainty of the order
of 15-20cm could drive the robot to collide with them.

4 0.25

3.5

0.2

2.5
0.15

0.1
1.5

1
0.05

0.5

0 0
0 1 2 3 4 5 6

(a)
0.25

0.2
0.5

0.45

0.4
0.15

0.35

0.3

0.25 0.1

0.2

0.15
4
0.05
0.1
3
0.05
2

0 1
0
0 1 0
2 3 4 5 6

(b)

Figure 5.2: 3D error surface for no error correction. (a) top view, (b) side view

78
 5 – Test and Performance

Figure 5.2 is a 3D representation of the absolute error. x-y plane shows the
position of the test points, while the absolute error is in the z direction. Each error
related to the relative test point is then interpolated forming a surface.
As it is possible to see the error is higher near the border of the area covered by the
anchors. The explanation is that the ranges are affected by underestimanting/over-
estimating error, and in the center area, due to the simmetry of the anchors position,
these errors compensate. Near the border istead these error tend to add each other.

5.1 Corrective function

This section shows the results obtained after the use of the corrective funtion. The
latter is one dimensional function that depends only on the range value. It operates
as a multiplicative factor that modifies the range value, correcting the underesti-
manting/overestimating effect presented in the previous chapter and noticeable in
figure 4.2

Correction
0 1 0,0442 -0,1340 0,1412
0 2 -0,0056 -0,0141 0,0152
0 3 -0,0049 -0,0098 0,0110
1 0 0,0011 -0,0944 0,0944
1 1 -0,0141 -0,0990 0,1000
1 2 -0,0415 -0,0539 0,0681
1 3 -0,0281 -0,0025 0,0283
1 4 -0,0787 -0,1324 0,1540
2 0 -0,0001 -0,0173 0,0173
2 1 0,0150 -0,0398 0,0426
2 2 -0,0328 -0,0400 0,0517
2 3 -0,0393 -0,0629 0,0741
2 4 -0,0096 -0,0475 0,0484
3 0 0,0037 0,0150 0,0154
3 1 0,0257 -0,0264 0,0369
3 2 0,0020 -0,0250 0,0251

79
 5 – Test and Performance

3 3 0,0013 -0,0605 0,0605

3 4 -0,0051 -0,0863 0,0865
4 0 0,0605 0,0154 0,0625
4 1 0,0264 -0,0113 0,0287
4 2 0,0249 -0,0037 0,0252
4 3 0,0122 -0,0454 0,0470
4 4 0,0331 -0,0852 0,0914
5 0 0,0364 -0,0456 0,0583
5 1 0,0581 -0,0334 0,0670
5 2 0,0843 -0,0023 0,0843
5 3 0,0355 -0,0054 0,0359
5 4 0,0883 -0,2105 0,2283
6 1 0,0460 0,0205 0,0504
6 2 0,0484 -0,0189 0,0520
6 3 0,1623 -0,3240 0,3624
max err [m] 0,3624 σ[m] 0,0706
RMSE [m] 0,1008

Table 5.3

Table 5.3 if compared to the previous one 5.1, shows an increased accuracy. The
maximum error is decreased of almost 10 cm, while the RMSE is 8 cm smaller, almost
halved. The RMSE, that now is 10 cm, is compliant with the final application.

Comparing figure 5.3 with figure 5.2 is possible to notice an important improve-
ment of the absolute error. The top view shows how the error is drastically reduced.
The side view instead is flattered and more uniform respect to the previous one.

80
 5 – Test and Performance

4 0.25

3.5

0.2

2.5
0.15

0.1
1.5

1
0.05

0.5

0 0
0 1 2 3 4 5 6

(a)
0.25

0.5
0.2
0.45

0.4

0.35
0.15
0.3

0.25

0.2
0.1
0.15

0.1

0.05
0.05
0
0 4
3.5
1 3
2 2.5
3 2
4 1.5
1
5 0.5 0
6 0

(b)

Figure 5.3: 3D error surface for error correction. (a) top view, (b) side view

Figure 5.4: Static measurements corrected using the corrective function.

Figure 5.4 shows the clusters of point distribution once applied the corrective

81
 5 – Test and Performance

function. If compared with figure 4.2 is evident how the clusters are moved toward
the test point, decreasing the euclidean distance.

5.2 Double system: 8 anchors

In order to have redundant information, another TREK1000 development kit is
added to the previous one. A total of eight anchors and one tag are used in the
following tests. Anchors are positioned forming an octagon in order to exploit the
symmetry of the constellation.
Anchors positioned in:
first system

• anchor 0 (0.0m; 0.0m, 0.09m);

• anchor 1 (2.82m; 2.82m, 0.09m);

• anchor 2 (-2.82m; 2.82m, 0.09m);

• anchor 3 (0.0m; 5.64m, 0.09m);

second system

• anchor 0 (-2.0m; 4.82m, 0.09m);

• anchor 1 (-2.0m; 0.82m, 0.09m);

• anchor 2 (2.0m; 4.82m, 0.09m);

• anchor 3 (2.0m; 0.82m, 0.09m);

• tag placed on a tripod 1.64m tall.

82
 5 – Test and Performance

Figure 5.5: Static measurements with 8 anchors witout any correction.

Figure 5.5 shows the measurement distribution with 8 anchors. One system is
represented by black diamond, while the second one by cyan diamonds.

5.3 Offset correction

The offset introduced by the closed-form localization algorithm is a bias errors that
pulls the measurement toward the reference anchor. In order to mitigate this error
all the anchors are used as reference one, repeating the positioning algorithm for
each of them. Then the final position is obtained averaging all the results.

83
 5 – Test and Performance

4 0.25 4 0.25

3.5 3.5
0.2 0.2
3 3

2.5 2.5
0.15 0.15

2 2

0.1 0.1
1.5 1.5

1 1
0.05 0.05

0.5 0.5

0 0 0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4

(a) (b)

Figure 5.6: Top view of the error surface. (a) shows the error after the correction,
while (b) is the error before.

In figure 5.6 is reported the top view of the surface error before and after applying
the offset correction. In the first case the error is ditributed along the border far from
the anchor 0 located in (0; 0) that is the reference node. In fact since all the points
are pulled toward this node, the error is present along the border in the opposite
direction. After averaging the results get from all the anchors used as reference, the
error is spread around the 2D area.

No correction
X err[m] Y pos[m] x err[m]2 y err[m] abs err[m]
0 1 0,0570 -0,0278 0,0634
0 2 0,0426 -0,0754 0,0866
0 3 0,0517 -0,1100 0,1216
1 0 0,0095 0,1178 0,1182
1 1 -0,0048 -0,0167 0,0174
1 2 0,0184 -0,0463 0,0499
1 3 0,0210 -0,1720 0,1732
1 4 -0,0335 -0,1956 0,1985
2 0 -0,0149 0,1332 0,1341
2 1 0,0034 -0,0052 0,0062

84
 5 – Test and Performance

2 2 0,0057 -0,0664 0,0666

2 3 -0,0012 -0,1652 0,1652
2 4 -0,0282 -0,1970 0,1991
3 0 -0,0066 0,1215 0,1217
3 1 -0,0744 -0,0213 0,0774
3 2 -0,0354 -0,0844 0,0915
3 3 -0,0774 -0,0216 0,0804
3 4 -0,0168 -0,1619 0,1628
4 1 -0,0668 -0,0543 0,0861
4 2 -0,0655 -0,1357 0,1507
4 3 -0,0680 -0,1664 0,1798
max err [m] 0,1991 [m] 0,0562
RMSE [m] 0,1247

Table 5.4

Correction
X pos[m] Y pos[m] x err[m] y err[m] abs err[m]
0 1 0,0585 0,0022 0,0585
0 2 0,0431 -0,0434 0,0612
0 3 0,0529 -0,0852 0,1003
1 0 0,0114 0,0807 0,0815
1 1 -0,0049 0,0100 0,0112
1 2 0,0184 0,0027 0,0186
1 3 0,0210 -0,1110 0,1129
1 4 -0,0110 -0,1881 0,1884
2 0 -0,0160 0,1135 0,1147
2 1 0,0034 0,0330 0,0332
2 2 0,0057 -0,0115 0,0129
2 3 -0,0012 -0,1179 0,1179
2 4 -0,0177 -0,1857 0,1865
3 0 -0,0142 0,1276 0,1284

85
 5 – Test and Performance

3 1 -0,0744 0,0146 0,0758

3 2 -0,0354 -0,0487 0,0602
3 3 -0,0460 -0,0275 0,0537
3 4 -0,0187 -0,1401 0,1414
4 1 -0,0669 -0,0039 0,0670
4 2 -0,0698 -0,0953 0,1181
4 3 -0,0776 -0,1376 0,1580
max err [m] 0,1884 [m] 0,0528
RMSE [m] 0,1042

Table 5.5

Analyzing tables 5.4 and 5.5 is possible to notice how the error is redistributed
in the area covered by the anchors. Peaks of error are reduced, the maximum error
is smaller after the correction as well as the RMSE.

5.4 NLOS correction

The NLOS is a common problem in environments full of objects and obstacles. In

the following picture is represented a static measure, where the anchors are forming
an octagon and the tag is placed in the center of the constellation in (0; 2.82). During
the measuments a person is walking around the tag moving an obstacle represented
by a metal cylinder. Since they both represent obstacles, some ranges are biased as
effect of the NLOS. The cluster of points is very spread all around the 2D plane and
the measurements don’t convey toward a unique solution.

86
 5 – Test and Performance

NLOS, person+obstacle moving around the tag

6

13

5
20 22

4
y[m]

3
12 11

1
21 23

0 10
-3 -2 -1 0 1 2 3
x[m]

Figure 5.7: Moving obstacle, no NLOS correction applied.

After applying the NLOS filter a big improvement is noticeable. Now the cluster
is compact and the points convey toward the center. All the point are included in a
radius smaller than 20cm from the real position (test point).

NLOS mitigation
6

13

5
20 22

4
y[m]

3
12 11

1
21 23

0 10
-3 -2 -1 0 1 2 3
x[m]

Figure 5.8: Moving obstacle after applying the NLOS filter.

87
Chapter 6

Anchors Calibration

The final application is designed to work in home environment. One aim is to make
this application ready to go as soon as the patient get the system, without any
external interventon by technicians. Once the patient places the anchors inside the
apartment, the position of the nodes are not known to the system, and it would be
unpleasant if the patient has to measure the exact position of them, or ask for the
help of some technicians.
In order to avoid all this a calibration phase has been developed. During this phase
the robot is able to construct a map of the environment and locate the anchors in it.
Thanks to this the final user only needs to know how to place the nodes all around
and start the robot.
The calibration process requires three steps to be performed in sequence in order to
be effective:

• step 1: use on board SLAM process to construct the environment map;

• step 2: during step 1, record where the minimum range values are measured;

• step 3: move in positions recorder during step 2 and perform a reverse trilat-
eration.

6.1 Step 1: mapping

Many vacuum cleaner home robots have the SLAM feature. Exploiting their on
board sensors they are able to draw a map of the environment in which they oper-
ate and also to autolocate theirselves.
The idea is to use this feature to obtain a detailed map of the house, knowing in
this way where obstacles are located and the shape of the apartment. Then thanks

88
 6 – Anchors Calibration

to the map, the robot is able to trace the shorter path between it and the patient
and navigates to him/her.
When the robot has to reach the patient, first of all it establishes a path to follow in
order to travel till destination. Then this path is decomposed in a list of movement
commands and sent to the drive interface that controls the wheels. As the robot
proceeds, it monitors its position and retraces the map, and if it is the case it mod-
ifies the list of commands.

During the acquisition of the shape of the environment the robot is able to lo-
cate also some obstacles such as a couch or a table, and then avoid it during the
path determination process. Furthermore in the case the patient decides to change
the enterior design of the habitation, he/she just needs to lunch the calibration pro-
cess again and the system will be upgraded with the new map. The same if he/she
decides to move the anchors.

6.2 Step 2: locate minimum ranges

As explained above, during step 1, the robot must record its position on the map
when the minimum distance between its tag and an anchor is measured. In this way
is possible to obtain an approximation of the anchor distribution in the house. The
latter is foundamental for the last step that is in charge of locating the anchors with
high accuracy.
Step 2 is a simple operation that can be executed using a simple MatLab script
that analyzes the range values, and every time the minimum distance between the
tag and one anchor si found out, the position obtained from the robot must be saved.

Decawave system has the problem that when the tag and the anchor are closer
than 20 cm, the range value is not consistent anymore, in the sense that the ToA
value is affected by an error that is greater that the measure itself. When two devices
are too close, the ToA measure requires a very high resolution, but since TREK1000
doesn’t have it, the measure is not valid.
This problem can be simply avoided by placing the anchors near the ceiling, above
people head. In this case the signal avoid the interference due to people, and since

89
 6 – Anchors Calibration

the tag is positioned on the robot, that travels on the floor, the range will be always
greater than 20cm.

6.3 Step 3: reverse trilateration

Once obtained an approximation of the position of the anchors, the robot can per-
form the third and last step to locate with an error smaller than 10 cm the anchors
in the home environment.
First it has to move in the same room where the minimum distance with respect to
the anchor has been recorded and then perform the reverse trilateration.

The trilateration process consists in locate a certain tag knowing its distance from
three fixed points positioned in known coordinates. In this case the robot can only
measure the distance from the tag to the anchor, but it has no fixed points. To
overcome this problem, the reference nodes are obtained by moving the robot in
three different points during the calibration process.

2.5

2 Anchor 2 Anchor 3

1.5

1 B C

0.5

0 A

-0.5

-1 Anchor 0 Anchor 1

-1.5
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3

Figure 6.1: Calibration procedure, L movement.

As shown in figure 6.1 the robot starts from point A, travels straight till point B,
then roates of 90 degrees and travels till point C. In each point A, B and C the robot
performs some range measurements. At this point the system knows the distances

90
 6 – Anchors Calibration

from the tag and the anchor in three different points at a known distance. In fact
the location of the point is not known, but the relative distances are known.
So the reverse trilateration is subdivided in several different phases:

• record range distances in point A;

• travel straight till point B,

• record range distances in point B;

• turn 90 degrees clockwise;

• travel staright till point C;

• record range distances in point C;

• trilateration.

The range distances are recorded by the computer connected to the robot and
to the tag. The L movement described by the robot is accurate, with uncertainty of
millimiters thanks to the high resolution of the wheel encoders. So the uncertainty
due to the robot movement is negligible respect to the uncertainty obtained from
the UWB system.

Note: in the following example is reported the case in which the reverse trilater-
ation is performed on 4 anchors. In the final application the reverse trilateration is
applied on one anchor at time, in order to avoid the NLOS problem introduced by
walls.

Point A is considered the origin of the system (xA = 0; yA = 0; zA = rh ), where

rh is the height of the tag positioned on the robot. Thi value is measured once the
structure of the robot is designed and recorded as constant in the software. Range
measurements in pont A are indicated as [d0A , d1A , d2A , d3A ]. Then the robot
travels till point B of a distance equal to mAB provided by the optical encoders.
The new point B is located in (xB = 0; yB = mAB ; zB = rh ) and the ranges are

91
 6 – Anchors Calibration

[d0B , d1B , d2B , d3B ] .

After a rotation of 90 degrees, the robot travels a distance of mBC measured again
by the encoders. So point C is (xC = mBC ; yC = mAB ; zC = rh ). Once last mea-
surements [d0C , d1C , d2C , d3c ] are obtained, reverse trilateration can be performed.
 
d0A d1A d2A d3A
D = d0B d1B d2B d3B  (6.1)
 

d0C d1C d2C d3C

 
0 0 rh
P = 0 mAB rh  (6.2)
 

mBC mAB rh
Where D is the matrix of the range values and P is the matrix of the relative
positions of the anchors. Matrix D is divided as follow:
       
d0A d1A d2A d3A
D0 = d0B  D1 = d1B  D2 = d2B  D3 = d3B  (6.3)
       

d0C d1C d2C d3C

Exploiting the Matrix P and vectors Di−th is possible to perform the reverse
trilateration and figure aout the position of the anchors:

• Anchor 0 = trilateration (P, D0 );

• Anchor 1 = trilateration (P, D1 );

• Anchor 2 = trilateration (P, D2 );

• Anchor 3 = trilateration (P, D3 );

Using three fixed points to trilaterate the position of the anchors gives two so-
lutions, one above and one below the height of the tag positioned on the bot. Since
this height is not modifiable, is not possible to obtain a measure on a different plane,
so the solution to be chosen is the one above the tag. In fact placing the anchors
just under the ceiling allows to know a priori that the anchors are higher than the
tag.

92
 Table 6.1: Calibration error vs relative distance between reference points. Errors obtained as euclidean distance
between real point and measured one.

anchor 0 anchor 1 anchor 2 anchor 3

distance x err0 y err0 abs err0 x err1 y err1 abs err1 x err2 y err2 abs err2 x err3 y err3 abs err3
0,10 0,2205 -1,4320 1,4489 -0,2541 1,1725 1,1997 -0,4379 -0,1890 0,4769 0,8130 -0,4375 0,9232
0,25 -0,5416 0,0797 0,5475 -0,1471 0,2259 0,2695 -0,1798 0,0538 0,1877 -0,0481 0,3795 0,3825

93
0,50 -0,4564 0,0883 0,4649 -0,0399 0,2473 0,2504 -0,1381 -0,0054 0,1382 0,1022 0,0996 0,1427
0,75 -0,3754 0,1048 0,3898 0,0486 0,1095 0,1198 -0,0830 -0,0120 0,0839 0,1367 0,2199 0,2589
1,00 -0,4354 0,0346 0,4368 0,0779 0,0761 0,1089 -0,0726 -0,0241 0,0765 0,1610 0,0987 0,1888
6 – Anchors Calibration

1,25 -0,2589 0,0260 0,2601 0,0960 0,0239 0,0989 -0,0988 -0,0427 0,1077 0,1610 0,0211 0,1624
1,50 -0,1740 0,0180 0,1749 0,1258 0,0170 0,1270 -0,0593 -0,0461 0,0751 0,1595 0,0319 0,1627
 6 – Anchors Calibration

Table 6.1 shows calibration errors for different relative distance between reference
points. Errors decrease as the relative distance increase, this is due to the fact that
since the reference points are more distant, the intersection of the spheres gives a
tight area.
For anchors 0 and 3 the errors are greater. This is due to the fact that for these
nodes the reference points A and C are in line, resulting in a bigger uncertainty
area given by the intersection. Instead for anchors 1 and 2 the reference points are
perfectly distribuited, in fact the error is in the order of 10 cm or less starting from
a relative distance of 50 cm. from this is possible to notice of the reference points
distribution is foundamental to reach an high accuracy.

Reverse calibration measure

2.5

1.5

1 B C

0.5

0 A

-0.5

-1

-1.5
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5

Figure 6.2: Calibration measuremet, L movement.

Figure 6.2 shows the anchor calibration for a relative distance of 1 m. Black
diamonds are the reference points while the green dots represent the ideal position
of the anchors, while the blue dots are the one obtained by means of the reverse
calibration.
The reverse calibration introduces an error on the position of the anchors of the
order of 10 cm that must be add to the error of the algorithm presented in chapter
IV.

94
Chapter 7

Conclusion and future work

The final application is an home robot monitoring system able to assist a patient
in his/her home environment in case of critic health status. The aim of this the-
sis work is to provide the robot with an indoor navigation system based on Ultra
Wide-Band. In chapter IV is shown how the algorithm works and in the following
one are reported test and results.
Starting from a declared error of 20 cm in positioning, due to the UWB Decawave
system, the algorithm is able to reduce this to 10 cm, so basically is able to halve
the initial error. Furthermore it introduces a mitigation of the NLOS error thanks
to redundant information coming from a second system.
Since the final application is designed to become a commercial device one day, it
requires a calibration process able to locate anchors devices inside the house without
the help of external technicians. This calibration phase is presented in chapter VI.
Thanks to its three steps process is able to locate anchors with an accuracy of 10
cm, that is very high considering the dimension of an apartment.

Many improvements can be done to this application in order to transform it in

a commercial device.
A possible future work consists in substituting the PC with Raspberry board that
can reduce the power consumpion, the occupied volume and the total weight.
The current robot prototype is not equipped with a functioning SLAM, so a possible
improvement is to design a SLAM procedure providing the necessary sensors to the
bot. Furthermore SLAM could be usefull to help the autolocalization of the bot,
so could be integrated with the actual localization algorithm in order to make the
software, and so the total system, more robust.
Since the Decawave UWB system is a development kit, it doesn’t allow to connect

95
 7 – Conclusion and future work

an external sensor to its Wireless Sensor Network (WSN). So the next step would
be to design a dedicated hardware PCB that allocates the DW1000 from Decawave
and alows also to communicate with external devices through a certain protocol.
This is foundamental, especially for the tag to be placed on the patient that has the
duty to transmit the health status to the central working unit represented in this
case by the PC or by the raspberry board.
Furthermore even the TWR algorithm could be improved, so instead allowing the
communication only between 4 anchors, it could be extended to an indefinite num-
ber, allowing to use redundand information on the same channel and with more
than one system.

96
Bibliography

[1] Mohammadreza Yavari and Bradford G. Nickerson Ultra Wideband Wireless

Positioning Systems Technical Report TR14-230 March 27, 2014
[2] Bruno Silva 1, Zhibo Pang 2, Johan kerberg 2, Jonas Neander 2 and Gerhard
Hancke 1 Experimental Study of UWB-based High Precision Localization for
Industrial Applications 1 Dept. of EECE, University of Pretoria, South Africa
2ABB AB, Corporate Research, Sweden 2014
[3] Francesco Bandiera,1 Angelo Coluccia,1 Giuseppe Ricci,1 Fabio Ricciato,2 and
Danilo Spano1 TDOA localization in asynchronous WSNs 1. DII, University
of Salento Via Monteroni, 73100 Lecce, Italy [email protected] 2.
Austrian Institute of Technology Vienna, Austria [email protected]
[4] Anastasios I. Mourikis and Stergios I. Roumeliotis Analysis of Positioning Un-
certainty in Simultaneous Localization and Mapping (SLAM) 2004 IEEElRSJ
International Conference on Intelligent Robots and Systems
[5] Edoardo Bonizzoni A remote assistive technology for elderly safety based on
indoor localization Master of Science in Engineering Master Thesis
[6] Decawave SOURCE CODE GUIDE DECARANGE RTLS ARM SOURCE
CODE Understanding and using the DecaRangeRTLS ARM source code Version
2.1 2015
[7] Decawave TREK1000 User Manual HOW TO INSTALL, CONFIGURE AND
EVALUATE THE DECAWAVE TREK1000 TWO-WAY RANGING (TWR)
RTLS IC EVALUATION KIT Version 1.06
[8] Decawave TREK1000 Quick Start Guide Two Way Ranging(TWR) Evaluation
Kit Version 1.3
[9] Ankush Vashistha, Rejina Ling Wei Choi, Choi Look Law A TDOA Measure-
ment Technique for Asynchronous Indoor Localization System using UWB-IR
The Center for Infocomm Technology (INFINITUS) Nanyang Technological Uni-
versity (NTU, Singapore) [email protected], rejinachoi, [email protected]
[10] Jacek Stefanski synchronous time difference of arrival (ATDOA) method
Gdansk University of Technology, Faculty of Electronics, Telecommunications

97
 Bibliography

and Informatics, Department of Radio Communication Systems and Networks,

Narutowicza 11/12, 80-233 Gdansk, Poland
[11] Jia Wang, Asad Khalid Raja, Zhibo Pang Prototyping and Experimen-
tal Comparison of IR-UWB based High Precision Localization Technologies
ABB AB, Corporate Research, 72178, Vsters, Sweden jiaw, [email protected],
[email protected]
[12] Julius O. Smith, Jonathan S. Abel IEEE TRANSACTIONS ON ACOUSTICS,
SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-35, NO. 12, DECEMBER
1987 1661 Closed-Form Least-Squares Source Location Estimation from Range-
Difference Measurements
[13] Zebra technologies dartuwb tech datasheet
[14] iRobot Create Owners guide www.irobot.com
[15] R. Xu, W. Chen, Y. Xu, and S. Ji, A New Indoor Positioning System Archi-
tecture Using GPS Signals, Sensors, vol. 15, no. 5, pp. 1007410087, 2015.
[16] Tingcong Ye, Michael Walsh, Peter Haigh, John Barton, Alan Mathewson,
Brendan OFlynn Experimental Impulse Radio IEEE 802.15.4a UWB Based
Wireless Sensor Localization Technology: Characterization, Reliability and
Ranging Tyndall National Institute University College Ireland, UCC, Cork
[17] Shuai He, Member, IEEE, Xiaodai Dong, Senior Member, IEEE, and Wu-Sheng
Lu, Fellow, IEEE Asynchronous Time Difference of Arrival Positioning System
[18] Ismail Guvenc, Chia-Chin Chong, and Fujio Watanabe NLOS Identification
and Mitigation for UWB Localization Systems DoCoMo Communications Lab-
oratories USA, Inc. 3240 Hillview Ave, Palo Alto, CA, 94304 iguvenc, cchong,
[email protected]
[19] https://www.pozyx.io/Documentation
[20] https://www.terabee.com/time-of-flight-principle/

98