DrivingStyles: a smartphone application to assess

driver behavior
Javier E. Meseguer, Carlos T. Calafate, Juan Carlos Cano, Pietro Manzoni
Department of Computer Engineering
Universitat Politecnica de Valencia
Camino de Vera SIN, 46022, Spain
[email protected], {calafate, jucano, pmanzoni}@disca.upv.es

Abstract-The DrivingStyles architecture integrates both data the type of road on which the vehicle is circulating. Currently,
mining techniques and neural networks to generate a classifica­ this information is being collected and used in applications
tion of driving styles by analyzing the driver behavior along
aimed at improving road safety and promoting eco-driving [3],
each route. In particular, based on parameters such as speed,
acceleration, and revolutions per minute of the engine (rpm), we
thus reducing fuel consumption and greenhouse gas emissions.
have implemented a neural network based algorithm that is able The given recommendations addressing the driving style of
to characterize the type of road on which the vehicle is moving, as each user can save up to 20% of fuel while improving driving
well as the degree of aggressiveness of each driver. The final goal safety.
is to assist drivers at correcting the bad habits in their driving
The rest of this paper is organized as follow: Section
behavior, while offering helpful tips to improve fuel economy.
In this work we take advantage of two key-points: the evolution
II briefly reviews previous related works. Then, Section III
of mobile terminals and the availability of a standard interface provides an overall overview of the DrivingStyles architecture.
to access car data. Our DrivingStyles platform to achieve a Both client and server side applications are described in more
symbiosis between smartphones and vehicles able to make the detail in sections IV and V, respectively. The neural network
former operate as an onboard unit. Results show that neural
we have developed, along with the achieved accuracy results,
networks were able to achieve a high degree of exactitude at
classifying both road and driver types based on user traces.
are presented in section VI. Finally, section VII presents the
DrivingStyles is currently available on the Google Play Store conclusions of our work.
platform for free download, and has achieved more than 1550
downloads from different countries in just a few months. II. RELATED WORK
Index Terms-Driving styles; Android smartphone; OBD-II; Technological achievements in the field of mobile telephony
neural networks; eco-driving.
are making smartphones very powerful devices. In the research
world, this high computing power opens new and attractive
I. INTRODUCTION
research areas where many of them are able to generate
Mobile devices have experienced a technological break­ economic profit and significant advantages for our society.
through in recent years, evolving towards high performance One of the main problems of eco-driving systems [3]
terminals with multi-core microprocessors, being smartphones is identifying the factors that affect energy consumption.
a clear representative exponent of this trend. In addition, Ericsson [4] suggests that, in order to save fuel, sudden
the On Board Diagnostics (OBD-II) [1] standard, available changes in acceleration and high speed driving should be
since 1994, has recently become an enabling technology for avoided. Johansson et al. [5] suggest maintaining low levels of
in-vehicle applications due to the availability of Bluetooth deceleration, minimizing the use of the first and second gears,
OBD-II connectors [2]. These connectors enable a transparent and putting every effort into using the 5th and 6th gears, while
connectivity between the mobile device and the vehicle's avoiding continuous gear changes.
Electronic Control Unit (ECU). There are several proposals that analyze which variables
The range of possibilities that arise when combining cars affect fuel consumption. Kuhler [6] presented a set of ten
and smartphones is endless, allowing, for example, diagnosing variables. These variables are used in laboratories that work
the car via mobile devices which assume the tasks that are with the consumption of fuel and gas emissions. Other authors,
typically performed by the On Board Unit (OBU) of the like Andre [7] and Fomunung [8], increased the number of
vehicle, or sending the collected data to a platform where parameters, or replaced some of them in order to improve the
diagnosis and vehicle maintenance can be done, detecting obtained results. The problem of all the previous proposals is
possible failures automatically. that the environment in which the vehicle is circulating is not
Based on obtained data about the speed, acceleration, and considered, despite it has a very significant influence on the
revolutions per minute of the engine, we have implemented our energy consumption.
DrivingStyles platform, which is based on neural networks and Regarding other works related to the proposal discussed in
permits characterizing the driving style of each user, as well as this paper, it is worth highlighting the work by Chen et al.

978-1-4799-3755-4/13/$31.00 ©2013 IEEE 000535

 ��===="_==---=-==.
.� .. 3 .. A. Route type classification.
Q"1If �
4 ..
B. Driving style maracreri%ation.

'"
I
Figure 1: DrivingStyles: architecture overview.

[9], who proposed an Android/OSGi vehicular platform able information such as speed, acceleration, engine revo­
to diagnose and manage the status of a vehicle remotely, using lutions per minute, throttle position, and the vehicle's
visual intelligence to continuously update their application geographic position. After gathering this information in
services based on context, and without user intervention. real time, the user uploads the route data to the remote
Focusing on road safety, systems such as e-NOTIFY [10] data center for analysis.
allow a fast detection of traffic accidents, improving the 2) A data center with a web interface able to collect large
assistance to injured passengers by reducing the response data sets sent by different users concurrently, and to
time of emergency services. The proposed system requires graphically display a summary of the most relevant
installing OBUs on board vehicles that are in charge of results. Our solution is based on open source software
detecting accidents, as well as notifying them to external tools such as Apache, PHP and Joomla.
control units (CU); the latter estimates the severity of the 3) A neural network, which must be designed and trained
accident and automatically informs the emergency services. using the most representative route traces in order to cor­
Zaldivar et al. [11] propose an Android-based application that rectly identify, for each path segment, the driving style
monitors the vehicle through the On Board Diagnostics (OBD­ of the user, as well as identify the segment profile: urban,
II) interface, being able to detect accidents and sending details suburban or highway. To implement our neural network
about the accident to pre-defined destinations through either we adopted the backpropagation algorithm [16], which
e-mail or SMS; these tasks are immediately followed by an has been proven to provide good results in classification
automatic phone call to the emergency services. problems such as the one associated to this project.
Our solution differs from the former ones by providing an 4) Integration of the neural networks in the data center
analysis of the driving style of each user based on neural net­ platform. The goal is to dynamically and automatically
work techniques in the scope of urban, suburban, or highway analyze user data, allowing users to find out their profiles
scenarios, helping to promote a more responsible and eco­ as a driver, thus promoting a less aggressive and more
driving behavior. ecological driving.
III. DRIV INOSTYLES ARCHITECTURE Figure 1 illustrates the steps followed by users of the Driv­
The proposed architecture applies data mining techniques ingStyles platform to assess their driving styles. The first step
to generate a classification of the driving styles of users based is registering the user at http://www.drivingstyles.info. and
on the analysis of their mobility traces. Such classification is downloading the free Android application.
generated taking into consideration the characteristics of each After installing the Android application in the smartphone,
route, such as whether it is urban, suburban or highway. and after connecting to the bluetooth ELM327 [2] interface
To achieve the overall objective, the system is structured inside the car (this connector is mandatory on all vehicles
around the following four elements, see Figure 1: since 2001), the data acquisition process can start. Once the
1) An application for Android based smartphones. Using route is completed and stored, the user can send it to the server,
an OBD-II Bluetooth interface, the application collects which checks the username and password before accepting the

978-1-4799-3755-4/13/$31.00 ©2013 IEEE 000536

incoming XML file. Through the web interface, the user can
view all the routes sent to the server, and have access to the
different statistics, the map of the route, the characterization
of the driving style, and the recommendations on eco-driving.

IV. ANDROID APPLICATION

The Android application is a key element of the system,

proving connectivity to the vehicle and to the DrivingStyles
web platform. Currently, it can be downloaded for free from
the DrivingStyles website http://www.drivingstyles.info. or
from Google Play https://play.google.com /store/apps /details?
id=com.driving. styles.

A. Configuration options
- , ,

In order to adjust the functionality of our Android appli­

DriVing Styles
...... :='1
cation to the user requirements, several configuration options
,
- ,.,
""'
must be defined related to user creation, connection options, T�
�.
'" '
GPS activation, and sensor sampling.
....
""

The available functionalities are the following:

1) User creation: A username and password matching
those registered in the website must be introduced,
allowing the server to validate the user credentials before
storing any data in the database.
2) OBD Connection: Before starting the sampling process,
the device must be paired with the OBD-II interface.
Once the device is paired with the smartphone, it can
Figure 2: Snapshots of the acceleration, speed, rpm parameters
be selected from the list of available devices within the
and map module.
application.
3) GPS Activation: although GPS data is not mandatory
for data capturing purposes, it is necessary in the broader
• Real time visualization of the speed, rpm, and accelera-
scope where the study is being done, since the variables
tion in a time window of 10 seconds.
retrieved from the vehicle are recorded along with the
The route upload module is in charge of sending the users'
car's geographical position for later analysis.
traces to the website data center for further analysis. The
4) Sensor sampling: The user can individually select the
Graphical Information module is in charge of displaying the
vehicle sensors used by the application for monitoring
relevant information to the user in the most convenient manner.
and storage among the following ones: acceleration, en­
The charts that appear on the screen are the acceleration, the
gine revolutions per minute (rpm), throttle position, and
speed, and the revolutions per minute (rpm). We have chosen
instantaneous speed. Additionally, the user can define
these three parameters since they are the most relevant ones,
the sampling rate (number of samples per sensor, and
and because they are the ones we selected for training our
per second).
neural network. The graphs can be displayed in real-time or
B. Application Modules by selecting data from previously stored paths. Depending on
the device model, the user can also zoom in and out using
The main module of our application launches the back­
the device's touch screen. Finally, the map module allows
ground processes responsible for capturing the data generated
displaying the GPS position on the map, where a green car
by the OBD-II and the GPS interfaces, as well as the phone's
icon indicates the beginning of the route, and a red icon shows
accelerometer.
the current position of the vehicle. The path is shown by using
Besides showing the sensors we are monitoring, we can
different colors depending on the vehicle's speed (see Figure
perform several parallel actions without affecting the data
2).
capture. Possible options are:
• Start and stop data capture of the Electronic Control Unit V. DRIVINGSTYLES WEB INT ERFACE

(ECU), the mobile's accelerometer, and the GPS. The second main component of our architecture corresponds
• List the routes captured by the app. to the data center and its web interface. For this endeavor we
• Show the current position of the vehicle on the map, as have selected open source software such as Apache HTTP, and
well as the detailed route followed whenever an Internet Joomla as the content management system (CMS). The URL
connection is available. of this module is http://www.drivingstyles.info.

978-1-4799-3755-4/13/$31.00 ©2013 IEEE 000537

 � Routes sent C;
backpropagation [16] since this kind of algorithm provides
g l9:
"' ,""
__ ur�
. 'U!
o
n ",
d o ============j �
commer>t J><
very good results in classification problems.
bv the u�r: jmeseguerOvahoo.com
RoutH

St8rbngTmFrishngTn� �SaTo�Trne AveraoeSpMllxSpeed Mileaoe

After considering the many variables that can be obtained
,..,.... , 1)9:07:54 ()9:()9:36 183 )(1]) from the Electronic Control Unit (ECU), we have chosen to
jneseouer5 201Hl3. )C'O> I!.l
--' 08:35:19 08:37:53 263 )(1]) train the neural network using the mean and the standard
,..,.... " 08:50:19 ()9:07:)7 1787 )(1])
ine=>er12 ()9:07:49 C19:H:21 I)6S )(-0. deviation values of: a) the speed, b) the vehicle accelera­
� Route Ford Fiesta 2 0-50 km/h _ ,.120 kmfh tion, and c) the rpm. In all vehicles used for testing, these
. ""
variables were easily obtained, while other variables, such
• +
as the position of the throttle, and despite being able to
provide important information for the neural network training,
finally had to be rejected because not all ECU manufacturers
provide such information. The data input of each parameter
is normalized between 0 and 1; this normalization should
take into consideration the whole range of possible values.
The equation used to normalize the input parameters is the
following:
xP-min(x)
Figure 3: Snapshot of a route map. xP' -
- max(x)-min(x)

Notice that xP' is the normalized value of input variable x

Once the user is logged in, he is asked to record a number of for pattern p, xP is the original value of variable x for that
important data, especially for future data mining studies. The pattern, and min(x) and max(x) represent the minimum and
most relevant items are sex, age, and other details concerning maximum values for the input variable, i.e., the minimum and
the vehicle used: car manufacturer, model, fuel type, and the maximum values of the column corresponding to variable x.
theoretical 0-100 acceleration (important to normalize the user The application used for the creation and training of the
behavior in our study). Finally, a third block allows drivers to neural networks required by this project was JavaNNS [12],
indicate what they feel about their own behavior behind the which is a Java version of the SNNS software developed at
wheel, i.e., whether they perceive themselves as aggressive, the University of Tiibingen.
moderate, or quiet drivers.
A. Neural Network description
In the Routes' section, users can access all the routes
they have uploaded. The first grid shows routes that are in An empty neural network was created by defining the num­
the database, including the name, date, starting time, ending ber of entries mentioned previously, along with the number of
time, samples sent, total time, average speed, and kilometers hidden nodes (nine in our case). A larger number of hidden
travelled. Below the grid, the selected route is shown in a map. nodes can improve the success rate, but it has the negative
The path varies its color depending on the speed of the car effect of increasing the response time. Two similar neural
(see Figure 3). networks were created, each one with three output nodes; the
In addition, the last two graphs show the results returned by first neural network (see Figure 4a) allows to classify the type
the neural network, including the driving style and the route of road (urban, suburban or highway), and the second one
characteristics (see Figures 6b, and 6a). allows characterizing the user's driving style (quiet, normal or
Next, in section VI, we provide detailed information about aggressive). Random weights between -1 and 1 are assigned to
the neural network we proposed for characterizing routes and the links of both neural networks during initialization. Figure
drivers. 4b shows the equivalence color map with their weight values
for the neural network identifying the road type (first one).
VI. NEURAL NETWORKS BASED DATA ANALYSIS
B. Training the Neural Network
In this project we face a classification problem: starting
After normalizing the input variables, the training set was
from some input data, which in our case are the speed, the
generated. This data set is used to adjust the weights of
acceleration, and the revolutions per minute of the engine
the neural network. After the neural network is trained, we
(rpm), we obtain as output the type of road and the driving
generate validation files. This data set is used to minimize
style.
overfitting, verifying that any increase in accuracy over the
We train Artificial Neural Networks to classify the driving
training data set actually yields an increase in accuracy over
sty Ie of each user and the type of route based on a well-defined
a broader data set not used for training.
set of rules and the ECU input variables. There are many
In order to do so, we rely on the average error for the test
different learning algorithms, such as backprop_momentum, L:'" jdP-yPI
Hebbian, or delta-rule, each one having its own advantages set: E P-l where IdP - yPI is the absolute error
=
m
and disadvantages depending on the type of problem to for pattern p, i.e., the absolute value of the difference between
solve [14]. In the scope of this work, we decided to choose the desired output and the output obtained, m is the number

978-1-4799-3755-4/13/$31.00 ©2013 IEEE 000538

 noName
---�,. ·0.34 -----:::;,.,
20

0.52�.0.33 ------- Rate 0.2

nONam� 15 ------- Rate 0.4
:-777<-:'311• 0.332 :---::,.L-;�nE"iI ------- Rate 0.6

- �"'
N 10
Q)
0 332 +".q;,e;z;.§iln

�
v-.:l
• •
�
5

noON!� 332
ol!;
rpm
��
!:
035"��:--:=31nOName 0
332
�/ 32
200 4 0 600 800 1000

Learning cycles
I 200 I 400 1 600 1800 2.00

•
=
! • om 10
0074 0003 052 If
(a) Convergence when training the road type with 2000 cycles.

O !i �
� ::: '" noName
' �6m '" '" 100
'"
c
=>
c
------- Rate 0.2
80·
------- Rate 0.4

�� �
:::�
N 60 ------- Rate 0.6
052 Q)
v-.:l40
noName -
·1 ·10
0..52
(a) Trained neural network. 500 1.000 1.500 2.000 2.500 3�00 3.500 4.000 4.500 5,005
Learning cycles

10 (b) Convergence when training the driving style with 5000 cycles.

15 Figure 5: Neural networks convergence behavior.

�
W
'" 10
0 �
�
achieves the lowest error among all the tested values,
with a mean square error of 0.43 and a sum of squared
errors of 6.11 (SSE).
Once the neural network is successfully trained (see Figure
source 4a), the knowledge obtained must be converted into usable
(b) Weights of the links of the neural network. code. We use the snns2.exe application [15], to convert the
trained neural network into C code, and this code was then
Figure 4: Trained neural network and associated weights.
integrated in to the web platform for data classification pur­
poses.

of patterns, and the sum corresponds to the sum of absolute C. Obtained results
errors for all patterns. With the neural network already implemented, every time
We started training the neural network by adjusting the a route or route segment is selected, the system automatically
learning rate to 0.2 and observing how the error affects our returns the type of road, and the associated driving style.
neural network. The higher the learning rate, the higher the The overall behavior of each user can also be obtained by
modification of the weights for each iteration, so the learning evaluating all the routes sent by the user.
process becomes faster. However, this can cause unwanted The results obtained from a real set of data can be seen
oscillations in the network. The convergence results of the in Figure 6. Figure 6a graphically shows the result returned
two neural networks are the following: by our neural network when determining the route type. The
1) Road type characterization: By iterating with a learning abscissa axis corresponds to the timeline, and the vertical axis
rate of 0.2, we see how the learning process slows down to the degree of matching for the dominant route type. We can
for a local minimum MSE with an error of 0.02l. The see that, in this example, the user starts and ends in an urban
network starts converging quickly again beyond 300 environment, going through a suburban route in between; such
cycles, as shown is the mean square error shown in results represent an accurate classification of the actual route
Figure Sa. followed by the driver. Figure 6b refers to the neural network
2) Driving style characterization: This characterization is results that determine the behavior on that route; as in the
more complex than the road type, and no local minimum previous graph, the abscissa axis corresponds to the temporal
is reached as quickly as in the previous case; also, the line, and the ordinate axis is a behavior correlation percentage,
error of the neural network is much higher (see Figure where each color identifies a different type of driver behavior.
5b). At the end of these tests we chose to use the We can observe that most of the time the driver is showing
trained neural network with a learning rate of 0.4, which an aggressive behavior, which alternates with periods of quiet

978-1-4799-3755-4/13/$31.00 ©2013 IEEE 000539

 Speed Car ecceleromeler Mobil ecceleromeler Rpm Throttle RNA roule type ItKA roule beMvior
RNA route type, route Ford Fiesta 2
The application, which is available for free download in
- .. CitV - .. Suburblln - .. H,,�h_v
100
" the DrivingStyle's website and in the Google Play Store, has
�
" achieved more than 1550 downloads from different countries
"
"
"
" in just a few months. This shows the great interest for
"
"
" applications integrating smartphones with vehicles.
.,
.,

"
"
" As future work we plan to monitor the fuel consumption
"
"
" in order to correlate it with the driver aggressiveness, thereby
,o
�o�, �
,�,��
. , �
,�,�,
� , ��
"
"U,�
, �
"�,,� U,�
,, , �
,, �
,,� L,�
,, o �
,, �
,,��,U
" . �
"� �
,, ,� U
, , ,�
,,��
,, ,�
" �
,�,,��
,. ,,
making the driver more aware of their bad habits while driving.
(a) Route type. We are also extending the platform to provide traffic recom­
mendations based on real-time feedback about the congestion
Speed Car acc eieromeler Mobil acceieromeler Rpm Thrortle AKA route Iype RIIA raUle behavior
on different routes.
RNA DrivingStytes, route FordFiesta2 FordFiesta2(40years)

ACKNOWLEDGMENT S

This work has been sponsored by the Universitat Politecnica

de Valencia through the ABATIS project (PAID-05-12), and
by the Ministry of Science and Innovation through the Walkie­
Talkie project (TIN2011-27543-C03).

(b) Route behavior. REFERENCES

Figure 6: Snapshots of route type and behavior. [I] International Organization for Standardization, ": Road vehicles, Diag­
nostic systems, Keyword Protocol 2000", 1999.
[2] ELM327DS. OBD to RS232 Interpreter. Elm Electronics - Circuits for
the Hobbyist.
behavior. When analyzing the actual route with the driver, we [3] Co Eco-Driving: Pilot Evaluation of Driving Behavior Changes among
find that periods of quiet behavior were associated with slow U.S. Drivers.
[4] Erikcsson, E., "Independent driving pattern factors and their influence
transit areas where vehicle queueing and semaphores impeded on fuel-use and exhaust emission factors" Transportation Research Part
a more aggressive driving style. D: Transport, Vol. 6(5), pp. 325-345, 200I.
Overall, and after analyzing a wide set of routes with [5] Johansson, H., Gustafsson, P., Henke, M., Rosengren, M., "Impact
of EcoDriving on emissions". International Scientific Symposium on
different drivers, the system has shown to provide a correct
Transport and Air Pollution, Avignon, France, 2003.
classification of the different route traces registered in the [6] Kuhler, M., Kartens, D., "Improved driving cycle for testing automotive
database, both in terms of route types and driving styles, exhaust emissions". SAE Technical Paper Series 780650, 1978.
[7] Andre, M., "Driving cycles development: Characterization of the meth­
validating the proposed solution. In particular, we achieve a
ods". SAE Technical Papers Series 961112, 1996.
degree of accuracy of 98% when attempting to characterize [8] Fomunung, I., Washington, S., Guensler, R. "A statistical model for esti­
road types, and a degree of accuracy of 77% when attempting mating oxides emissions from light duty motor vehicles". Transportation
to characterize the driving style of users. These values are Research Part D, pp.333-352, 1999.
[9] Chen, M.-C, Chen, J.-L., and Chang, T.-W., "Android/OSGi-based
easily explained since the latter process is more ambiguous vehicular network management system," Elsevier Computer Commu­
than the fonner. nications, vol. 34, no. 2, pp. 169 - 183, 201I.
[10] Fogue, M., Garrido, P., Martinez, FJ., Cano, 1., Calafate, c.T., Manzoni,
VII. CONCLUSIONS AND FUTURE WORK P., "Automatic Accident Detection: Assistance Through Communication
Technologies and Vehicles", Vehicular Technology Magazine, IEEE,
In this paper we presented the DrivingStyles platform, Sept. 2012.
which integrates mobile devices with data obtained from the [II] Zaldivar, J., Calafate, c.T., Cano, J.C., Manzoni, P., "Providing accident
Electronic Control Unit (ECU) to determine the type of road detection in vehicular networks through OBD-II devices and Android­
based smartphones", Local Computer Networks (LCN), IEEE 36th
where the driver is circulating, as well as his driving habits. Its Conference, 4-7 Oct. 20II.
main goal is to help promoting a safer and more ecological [12] Fischer, I., Hennecke, F., Bannes, C., Zell, A., "JavaNNS: Java Neural
driving style by making drivers more conscious about their Network Simulator", User Manual, Version 1.1.
[13] Haykin, S., "Neural Networks: A Comprehensive Foundation (2nd ed.)",
behavior on the road. Our platfonn offers helpful tips to Prentice Hall PTR, Upper Saddle River, NJ, USA, 1998.
reduce fuel consumption with the consequent impact on the [14] Anand, R.; Mehrotra, K.; Mohan, C.K.; Ranka, S., "Efficient classifi­
environment. We implemented this platfonn using real devices, cation for multiclass problems using modular neural networks," Neural
Networks, IEEE Transactions on, vol.6, no.l, pp.117-124, Jan 1995
and the results we obtained based on real user traces are quite
[15] hltp:llwww-ra.informatik.uni-tuebingen.de/SNNSI
encouraging, showing that the classification of both routes and [16] Hecht-Nielsen, R., "Theory of the backpropagation neural network,"
driving styles using neural networks presents a high correlation Neural Networks, IJCNN., International Joint Conference on, pp.593-
605 vol.I, 1989.
with the actual routes and driver behavior.

978-1-4799-3755-4/13/$31.00 ©2013 IEEE 000540