MATEC Web of Conferences 90 , 01005 (2017) DOI: 10.

1051/ matecconf/20179001005
AiGEV 2016

Simulating autonomous driving styles:

Accelerations for three road profiles
Juffrizal Karjanto1, 2, Nidzamuddin Md. Yusof1, 2*, Jacques Terken1, Frank Delbressine1,
Muhammad Zahir Hassan1, 3 and Matthias Rauterberg1
1
Department of Industrial Design, Eindhoven University of Technology, Eindhoven, The Netherlands
2
Faculty of Mechanical Engineering, Universiti Teknikal Malaysia Melaka, Melaka, Malaysia
3
Faculty of Engineering Technology, Universiti Teknikal Malaysia Melaka, Melaka, Malaysia

Abstract. This paper presents a new experimental approach to simulate

projected autonomous driving styles based on the accelerations at three
road profiles. This study was focused on the determination of ranges of
accelerations in triaxial direction to simulate the autonomous driving
experience. A special device, known as the Automatic Acceleration and
Data controller (AUTOAccD), has been developed to guide the designated
driver to accomplish the selected accelerations based on the road profiles
and the intended driving styles namely assertive, defensive and light rail
transit (LRT). Experimental investigations have been carried out at three
different road profiles (junction, speed hump, and corner) with two
designated drivers with five trials on each condition. A driving style with
the accelerations of LRT has also been included in this study as it is
significant to the present methodology because the autonomous car is
predicted to accelerate like an LRT, in such a way that it enables the users
to conduct activities such as working on a laptop, using personal devices or
eating and drinking while travelling. The results demonstrated that 92 out
of 110 trials of the intended accelerations for autonomous driving styles
could be achieved and simulated on the real road by the designated drivers.
The differences between the two designated drivers were negligible, and
the rates of succeeding in realizing the intended accelerations were high.
The present approach in simulating autonomous driving styles focusing on
accelerations can be used as a tool for experimental setup involving
autonomous driving experience and acceptance.

1 Introduction
In Malaysia, the number of road accidents is steadily increasing while deaths remain static
Recent studies have shown that the development of autonomous vehicle (AV) is rapidly
growing, and most drivers have already accepted Advance Driving Assistance System
(ADAS), which is the foundation of the autonomous driving system [1], [2]. The
automakers promise that the ultimate level of automation is a fully AV where human
drivers/users will no longer have to deal with the operational and tactical levels in driving

*
Corresponding author: [Link]@[Link]

© The Authors, published by EDP Sciences. This is an open access article distributed under the terms of the Creative
Commons Attribution License 4.0 ([Link]
MATEC Web of Conferences 90 , 01005 (2017) DOI: 10.1051/ matecconf/20179001005
AiGEV 2016

anymore [3]. Therefore, future AV users are only expected to decide where they want to go,
select the destination and let the AV does the rest, from driving the car, selecting the best
route, dealing with other road users and all the tasks that the human drivers used to do. In
that sense, an AV is expected to behave in optimized manners that promote safety,
reliability and comfort to the users. However, the acceptance of comfort may vary
depending on users’ perception and can be very subjective due to many factors especially
driving styles [4]. Taubman-Ben-Ari et al. [5] suggested that driving styles vary across
drivers and identified the most driving-specific factors into four categories; reckless and
careless driving style, anxious driving style, angry and hostile driving style, and patient
and careful driving style.
An AV or self-driving car contains thousands of sensors, and the way the vehicle is
driven is strictly based on optimized logic (based on algorithms that analyse the inputs from
sensors). In contrast, human drivers are driving based on their “targeted feeling” and
heavily influenced by their own personality traits [6], [7]. Hence, this may lead to
mismatched in term of driving styles preference for different types of the future AV users.
The way an AV accelerates and decelerates at the intersection, handles a corner and circles
a roundabout will be unfamiliar to the human drivers/users. Furthermore, since the AV
drivers/users are no longer in command (no hands on the wheel, no feet on the pedals), the
AV will probably accelerate and decelerate in the style that did not induce mental and
physical discomfort to the human drivers/users in order to keep safety and comfort intact.
In order for an AV to be an optimized driver without sacrificing the comfort and
enabling the activities other than driving for the users inside the AV, the AV is predicted
and suggested to accelerate and decelerate slower than the typical human driven car and
faster (or maybe the same) than the light rail transit (LRT) [8]. The LRT systems could be
used as guideline such that the ways the system accelerates and decelerates in longitudinal
and lateral direction are within the acceptable level of comfort for the passengers.
Activities such as drinking and eating without spilling, and working or playing with
electronic devices without experiencing physical discomfort like motion sickness as has
been predicted would happen in the future AV riding experience [9], [10]. These activities
were made possible since the induced forces felt by the users inside the LRT were
minimized. In addition, the experience in riding an LRT is likely to be the benchmark used
by the automotive players for their future development of AV since the LRT concept has
been proven as one of the best options for the user to conduct secondary tasks while
travelling in a moving vehicle [8]. On the same hand, future AV users who have had the
experience travelling by train or LRT would draw mental scenarios of how traveling inside
an AV should be as this new technology will also offer the kind of travelling without the
need to drive.
A typical human driven car accelerates around 0.79 m/s2 to 4.86 m/s2 after leaving a
junction or signalized intersection as stated in the work of Bogdanović and Ruškić [11].
They studied the various accelerations of cars, using digital cameras and programming,
starting from the stop line at a juncture with traffic lights for ten intersections in one of the
towns in Serbia. For the deceleration when approaching a junction or signalized
intersection, El-Shawarby et al. [12] found that with 821 events, the range is between 1.51
m/s2 to 7.47 m/s2. The study was done with 60 participants in the enclosed area where it
was safe to conduct the experiment. On the other hand, for the lateral accelerations, the
ranges are influenced by the radius of the corner or roundabout on which the car is
travelling [13]. Hugemann and Nickel (2003) [13] revealed that drivers usually produced
smaller lateral accelerations with smaller radii corners while, in contrast, higher radii would
usually associate with higher lateral accelerations. For the sharp turn or corner with radius
in between 20 to 40 meters, the 10th, 50th and 90th percentile of the lateral acceleration

2
MATEC Web of Conferences 90 , 01005 (2017) DOI: 10.1051/ matecconf/20179001005
AiGEV 2016

found is 3.1 m/s2, 4.1 m/s2 and 5.3 m/s2, respectively. Table 1 shows the list of accelerations
in multi-directions for both the LRT and human driven car.
Table 1. Longitudinal and lateral accelerations and decelerations for light rail transit and human
driven cars [11]–[15].
Typical human- Upper limit in
LRT
driven car human-driven car
Longitudinal
1.34 m/s2 or 0.14 g 2.47 m/s2 or 0.25 g 4.86 m/s2 or 0.5 g
acceleration
Longitudinal
- 1.34 m/s2 or 0.14 g - 3.27 m/s2 or -0.33 g - 7.47 m/s2 or -0.76 g
deceleration
Lateral
1.47 m/s2 or 0.15 g 4.10 m/s2 or 0.42 g 5.30 m/s2 or 0.54 g
acceleration

For the vertical acceleration, Weber and Braaksma [16] stated that the measurement of
discomfort, in terms of vertical force when traversing a hump, is depending on root sum of
square (RSS) acceleration or peak vertical acceleration. The RSS value that indicates
physical discomfort towards the motorist was at 0.17g, while peak vertical acceleration of
0.7g, is identified as heavily associated with human physical discomfort.
This paper reports the development and the execution of the experimental technique
used in simulating the AV driving style in real road scenario that involves cornering,
junction, and speed hump. A significant challenge to this research was to simulate the
autonomous driving styles based on the selected accelerations at different road profiles. The
present approach consists of experimental work that was carried out and validated using on-
vehicle site testing with two designated drivers who attempted to simulate the intended
accelerations using a special device which is called as the Automatic Acceleration and Data
controller (AUTOAccD).
The outcomes of this experiment were mapped with the results of calculated
accelerations measurement from the data gained from the testing vehicle. In addition, the
simulated data from the two designated drivers were tabulated and compared.
The current approach in simulating autonomous driving styles focusing on the
accelerations in the triaxial direction can be used as a tool for experimental research on
future autonomous driving experience and acceptance, especially in the preferred
accelerations and comfort determination.

2 Methodologies
The methodology of the present study is shown in Fig. 1. The process started with the
determination of driving styles to be simulated. Next, the process continued with the
selection of road profiles that consist of a straight lane approaching and leaving a junction,
a cornering, and a speed hump. Afterwards, the determination process of the driving styles
to be simulated and followed by the determination of ranges of the accelerations for the
selected driving styles based on the selected road profiles. Finally, the outcomes were
examined in twofold analyses; first, was the comparison between the measured
accelerations data using AUTOAccD and the calculated accelerations using the On-Board
Diagnostic (OBD) data from the testing vehicle and second, was the comparison between
accelerations simulated by the two designated drivers.

3
MATEC Web of Conferences 90 , 01005 (2017) DOI: 10.1051/ matecconf/20179001005
AiGEV 2016

Fig. 1. The methodology of the present study.

2.1 Driving style

The first step was to determine the type of driving styles to be simulated. Elander et al. [4]
defined driving style as the approach individuals select to manoeuvre their cars and
something they accustom to over the years. In addition, driving styles include the selection
of driving speeds in different circumstances, distance to others, the decision to overtake and
tendency to violate traffic laws. Wang and Lukic [17] reviewed and summarized that
driving styles are usually clustered into three categories; mild drivers (calm driving or
economical driving style), normal drivers (medium driving style), and aggressive drivers
(sporting driving style). Likewise, Taubman-Ben-Ari et al. [5] created the Multidimensional
Driving Style Inventory (MSDI), which was modified from existing questionnaires (e.g.,
[18], [19]). Thus, driving styles vary across drivers and therefore more than one driving
style for autonomous driving experience would be required [20]. In this study, three types
of future autonomous driving styles of interest were selected which are defensive, assertive
and LRT.

2.2 Road scenario

An Audi A3 was used as the testing vehicle for this experiment, provided by the
Automotive Lab, Eindhoven University of Technology. The experimental investigation was
carried out based on three different urban road profiles consisting of a junction (to access
longitudinal accelerations and decelerations), speed hump (to access vertical accelerations)
and corner (to access lateral accelerations). The route is located within the private
compound of the Science Park campus, Eindhoven University of Technology, Eindhoven,
Netherlands.

2.3 Ranges of accelerations determination

The ranges of lateral and longitudinal accelerations and deceleration to be implemented in
order to simulate the LRT, defensive and assertive driving styles are presented in Fig. 2.
These parameters are based on the data from LRT and typical human driven cars as

4
MATEC Web of Conferences 90 , 01005 (2017) DOI: 10.1051/ matecconf/20179001005
AiGEV 2016

reported in previous works [8], [11]–[13], [15]. For the longitudinal acceleration and
deceleration, where the time plays a major role in getting the selected accelerations and
decelerations, data for required time to generate the intended forces was calculated. In
addition, the targeted velocity (the ending velocity in simulating acceleration or the starting
velocity in simulating deceleration) would determine the required time in order to achieve
the intended accelerations or decelerations and depending on where the experiment is done,
the speed limit of that particular route needs to be taken into consideration.

Fig. 2. The range of selection for defensive and assertive driving style for lateral and longitudinal
accelerations and deceleration.
Next, the lateral acceleration, which is the centrifugal acceleration that pushes the
vehicle to the outward of the curve during a cornering process, was determined. The
centrifugal acceleration (ac) formulation is shown in Eq. 1 below:
v2 (1)
ac
r
where v and r represent travelling velocity in the corner and radius of the corner,
respectively. In order to achieve the intended lateral accelerations as has been shown in
Fig. 2, the selection of velocity has to be determined based on the parameters depicted in
Table 2 below.
Table 2. Required velocity to generate lateral acceleration during cornering.

Driving Lateral acceleration Velocity

style during cornering (m/s) (km/h)
LRT 0.15 g or 1.47 m/s2 4.79 17.24
Defensive 0.29 g or 2.84 m/s2 6.66 23.97
Assertive 0.48 g or 4.70 m/s2 8.57 30.85
Radius of the corner = 15.60 m

As for the vertical acceleration, the selections of velocity when approaching the speed
hump for different driving styles are shown in Fig. 3 and Table 3.

5
MATEC Web of Conferences 90 , 01005 (2017) DOI: 10.1051/ matecconf/20179001005
AiGEV 2016

Fig. 3. The range of selection of velocity approaching the speed hump.

Table 3. Required velocity to generate vertical acceleration at the designated speed hump.
Vertical
Velocity
Driving style acceleration
(m/s) (km/h) (m/s2) (g)
Defensive 0 – 6.7 0 – 24.0 0.39 0.04
Assertive 6.7 – 13.3 24.0 – 48.0 3.72 0.38
Height of the speed hump = 0.10 m
Length of the speed hump = 4.07 m

2.4 Automatic acceleration and data controller (AUTOAccD)

A special dedicated device, the Automatic Acceleration and Data controller (AUTOAccD),
was specifically developed with dual capabilities: to assist the designated driver to drive
according to the defined acceleration condition and also act as instrumented processor to
measure and collect the vehicle dynamics data. AUTOAccD consists of an OBD-II adapter
that is connected to Arduino board and LCD display as shown in Fig. 4. OBD is an
automotive term referring to a vehicle's self-diagnostic and is able to monitor real-time
engine parameters such as speed, fuel consumption and transmission of a vehicle. The
AUTOAccD was developed based on the data logger kit from Freematics [21] which
comprised of Global Positioning System (GPS) receiver, triple-axis Micro-Electro-
Mechanical Systems (MEMS) accelerometers and gyro meters. In addition, the data logger
was connected to a live display in the form of LCD screen. All the measured and collected
data are stored in a micro Secure Digital (SD) card for analysis later.
The display was designed to assist the designated driver in maintaining a certain driving
style by showing the specific ranges of lateral and longitudinal forces during the
experiment. The display served as assistance and assurance device to the driver, apart from
the training and his own experience in driving, in order to achieve the required acceleration
at the specific location. There are three colours available to indicate the instantaneous
driving style, in which LRT is indicated by purple colour while defensive and assertive are
shown by green and red colours, respectively. In addition, there are two axes on which the
vertical axis represents the longitudinal direction (upward movements indicates positive
direction and vice versa) while the horizontal axis represents the lateral direction (rightward
movements indicates positive direction and vice versa.

6
MATEC Web of Conferences 90 , 01005 (2017) DOI: 10.1051/ matecconf/20179001005
AiGEV 2016

(a) (b)
Fig. 4. (a) The position of the AUTOAccD in the car; and (b) the placement of the data logger and
OBD.

2.5 Data analysis

Recent developments in Micro-Electro-Mechanical Systems (MEMS) technology have
enabled the development of miniaturized sensors units like accelerometers and gyro meters.
However, working with data obtained from the experiments involving measurements by
these devices usually comes with noise ranging from alignment errors to random noise.
Hence, the noise has to be filtered for further data analysis and interpretation. Rather than
using the traditional filters for de-noising such as time-frequency domain filters like Fourier
transform, a time-scale domain filter known as Wavelet transform was employed. The
Wavelet de-noising method was implemented because its ability to remove noises and
retain the signal regardless the frequency of the signal [22], [23]. For this study, the
Wavelet de-noising was done using Matlab (Wavemenu) and the selection of the
appropriate filter was based on the previous literature using the same equipment. The
Daubechies filter level three (Db3) was used in order to remove the high-frequency noise
from the MEMS accelerometers [24].
After the de-noising process, the acceleration data from the two designated drivers were
compared to check the ability of the AUTOAccD in assisting the drivers to simulate the
intended accelerations. In addition, the accuracies and consistencies of the simulated
acceleration will also be discussed. Next, the comparison between the acceleration data
gained from the AUTOAccD and the calculated data from the testing vehicle’s OBD was
done to check the discrepancies between the two tools.

3 Results and discussion

The analyses started with the process of filtering or de-noising of the signal attained from
the accelerometers. Wavemenu function in Matlab was used for its simplicity and its
elegant visual tools that enable the users to interactively de-noise a signal and
instantaneously display the result on the screen [25]. After the de-noising process, the
acceleration data were analysed in terms of simulations from the two designated drivers and
were compared with the calculated data that were gained from the vehicle’s system.

3.1 Comparison between data measured by AUTOAccD and calculated from

OBD
The comparison between acceleration data (longitudinal and lateral) gained from the
triaxial accelerometers within the AUTOAccD and calculated using the speed extracted

7
MATEC Web of Conferences 90 , 01005 (2017) DOI: 10.1051/ matecconf/20179001005
AiGEV 2016

from the vehicle’s OBD is shown in Table 4. For the longitudinal accelerations and
decelerations, the largest variances were found in the simulation of assertive driving style
where as much as 0.06 g and 0.12 g differences were recorded, respectively. Otherwise, for
the defensive and LRT driving style, the differences in means between AUTOAccD and
OBD were between 0.00 g to 0.03 g, which indicated higher accuracies. Meanwhile, for the
lateral acceleration, between the three driving styles, AUTOAccD for assertive driving style
yielded the highest difference with 0.14 g while the defensive driving style registered a
difference of 0.05 g and an LRT driving style with the smallest difference with 0.01 g.
Table 5 shows the comparison between vertical acceleration data collected from the
AUTOAccD and determined from the vehicle’s OBD. The differences in means for each of
the driving styles were 0.03 g and 0.04 g. In addition, the standard deviations recorded were
between 0.00 and 0.06 which showed great consistencies among the two set of data. From
Table 4 and Table 5, similar patterns can be observed in which all the measured
acceleration data were higher than the calculated ones. These phenomena were parallel with
the previous work of Brol and Mamala [26] who found comparable results with the
measured data having the differences up to 0.10 g. The differences were contributed from
the road irregularities, and some were from the engine noises. The road surface
irregularities, in this study, were the cobblestones road that generated additional
accelerations through the vehicle suspension system and was recorded by the
accelerometers which were attached to the vehicle’s body.
Table 4. Comparison between longitudinal and lateral accelerations data measured by AUTOAccD
and calculated from OBD.
Type of acceleration Assertive DS Defensive DS LRT DS
AUTO OBD AUTO OBD AUTO OBD
AccD AccD AccD
Longitudinal Mean 0.41 0.35 0.22 0.21 0.09 0.07
acceleration SD 0.07 0.06 0.06 0.06 0.04 0.05
Longitudinal Mean -0.48 -0.36 -0.32 -0.29 -0.15 -0.15
deceleration SD 0.07 0.10 0.04 0.09 0.03 0.03
Lateral Mean 0.54 0.39 0.31 0.26 0.12 0.11
acceleration SD 0.12 0.05 0.03 0.02 0.04 0.03
Table 5. Comparison between vertical accelerations data measured by AUTOAccD and calculated
from OBD.
Assertive DS Defensive DS
Vertical
AUTO AUTO
acceleration OBD OBD
AccD AccD
Mean 0.33 0.30 0.07 0.03
SD 0.06 0.01 0.01 0.00

3.2 Comparison between two designated drivers

For the leaving-a-junction scenario, the tabulated accelerations simulated by the two
designated drivers using three driving styles (with five trials for each driving style) are
presented in Fig. 5. Additionally, the means and standard deviations (SD) are presented in
Table 6. For the simulated accelerations by using LRT driving style, both drivers managed
to simulate the accelerations within the desired range, although the means of the drivers
were quite far apart, the SD indicated that they were very consistent in simulating this
driving style. For the defensive and assertive driving styles, both of the drivers had
fluctuated accelerations and were less consistent compared to when simulating the LRT
driving style. Nonetheless, the means for both drivers in simulating the last two driving

8
MATEC Web of Conferences 90 , 01005 (2017) DOI: 10.1051/ matecconf/20179001005
AiGEV 2016

styles were closer to the intended accelerations. For the defensive driving style, out of ten
trials altogether, three of the accelerations fell outside of the desired range, whereas for the
assertive driving style, all the simulated accelerations were within the intended range.
In this particular scenario, the testing car was accelerating in one smooth motion where
there was no jerking or any other traffic that would impede the acceleration process. In
addition, in order to successfully simulate larger acceleration forces especially for the
assertive driving style, the accelerations must be done in preferably with lower transmission
such as the first gear. Higher acceleration forces were usually achieved with lower gears
especially when the vehicle started to accelerate from a completely static position. Higher
gears usually produce acceleration not as large as the acceleration produced from the lower
gears after leaving an intersection as has been shown in the previous study by Bogdanović
and Ruškić [11].

Fig. 5. Tabulated longitudinal accelerations for two designated drivers simulated with three driving
styles.
Table 6. Means and SD of the three simulated driving styles for longitudinal accelerations for each
driver, ntrial = 5.
LRT DS Defensive DS Assertive DS
(Targeted acc. = 0.14g) (Targeted acc. = 0.20g) (Targeted acc. = 0.38g)
Driver Mean (g) SD Mean (g) SD Mean (g) SD
A 0.11 0.01 0.19 0.06 0.42 0.08
B 0.06 0.03 0.25 0.05 0.40 0.06

For the straight lane approaching the junction scenario, the decelerations simulated by
the two designated drivers using three driving styles, with five trials each, are tabulated in
Fig. 6. In addition, the means and SDs are also presented in Table 7. For this scenario, both
drivers showed high consistency in simulating the accelerations although driver B actually
did not manage to register any acceleration within the desired LRT range, whereas, for the
defensive and assertive driving style, much higher variances were produced by both of the
drivers. In addition, the attempted decelerations generally deviated quite far from the
intended means. For the assertive driving style, all the simulated decelerations were within
the intended range, while for the defensive driving style, both of the drivers registered half
of the trials outside the intended range.

9
MATEC Web of Conferences 90 , 01005 (2017) DOI: 10.1051/ matecconf/20179001005
AiGEV 2016

In simulating the deceleration, the drivers were instructed to decelerate, from a

designated speed of 30 km/h, by adjusting the braking force according to the guides as
shown on the AUTOAccD. However, in general, the simulations of deceleration are harder
to generate compared to the simulations of acceleration. One of the reasons was that the
drivers had to control both the clutch and brake pedal in order to achieve the designated
deceleration rate and making sure the engine would not stall whereas, in the accelerations,
the drivers only need to control the accelerator or gas throttle.

Fig. 6. Tabulated decelerations for two designated drivers simulated with three driving styles.
Table 7. Means and standard deviations of the three simulated driving styles for decelerations for
each driver, ntrial = 5.
LRT DS
Defensive DS Assertive DS
(Targeted acc. = -
Driver (Targeted acc. = - 0.26g) (Targeted acc. = - 0.55g)
0.14g)
Mean (g) SD Mean (g) SD Mean (g) SD
A - 0.13 0.01 - 0.31 0.04 - 0.47 0.07
B - 0.17 0.02 - 0.33 0.05 - 0.49 0.06

For the cornering, the lateral accelerations simulated by the two designated drivers
using three driving styles are presented in Fig. 7. In addition, the means and SDs are also
presented in Table 8. In the simulation of the LRT driving style, driver B had better
consistencies compared to driver A as lower SD was registered. For the defensive driving
style, in which both drivers achieved mean values that are higher than the intended one,
both of the drivers registered good consistencies with SD of 0.03 each. Whereas for the
simulation of assertive driving style, there were fluctuations in the simulated accelerations,
as in this driving style the drivers had to generate accelerations that are above the average
of typical human driving style. However, out of the combined ten trials from the two
drivers, only once a simulated acceleration fell outside the lower limit of the intended
assertive range.

10
MATEC Web of Conferences 90 , 01005 (2017) DOI: 10.1051/ matecconf/20179001005
AiGEV 2016

In simulating the lateral accelerations, it was crucial that the radius of the cornering was
maintained to be as constant as possible. By referring back to Eq. (1), it is clear that both
velocity and radius of the corner are playing a big role in determining the induced lateral
acceleration [27]. In this study, the roads which have been used do not have any road
markings on the surface, and therefore, both of the drivers would try to undertake the
cornering by making sure the car was always approximately travelling in the middle of the
road. As mentioned by Charlton [28], for more accurate results in cornering, a clearer and
more accurate guideline or delineation treatment should be developed such as line marking
on the road surface where the drivers can align themselves to make sure that for each trial,
the radius of the travelling car in the corner stays almost the same every single time.

Fig. 7. Tabulated lateral accelerations for two designated drivers simulated with three driving styles.
Table 8. Means and standard deviations of the three simulated driving styles for lateral accelerations
for each driver, ntrial = 5.
LRT DS Defensive DS Assertive DS
Driver (Targeted acc. = 0.15g) (Targeted acc. = 0.29g) (Targeted acc. = 0.48g)
Mean (g) SD Mean (g) SD Mean (g) SD
A 0.12 0.06 0.30 0.03 0.47 0.11
B 0.12 0.02 0.32 0.03 0.46 0.14

For the vertical acceleration, the simulated accelerations are presented in Fig. 8. In
addition, the means and SDs are presented in Table 9. For this study, both designated
drivers realized higher vertical acceleration means than the intended for the simulation of
defensive driving style while for the assertive driving style, they realized lower means than
the intended one. In term of SDs, for the defensive driving style, both drivers managed to
simulate higher consistencies as indicated by low deviation values although for the Driver
A, his accelerations were sometimes lower and some other time were higher than the
intended means whereas, for driver B, all the recorded values were lower than the intended
means. For the assertive driving style, driver A indicated lower consistencies compared to
driver B as the SD recorded were 0.08 and 0.03, respectively. From

11
MATEC Web of Conferences 90 , 01005 (2017) DOI: 10.1051/ matecconf/20179001005
AiGEV 2016

, all of the 20 trials from both drivers were recorded within the intended ranges of
interest.

Fig. 8. Tabulated vertical accelerations for two designated drivers simulated with two driving styles.
Table 9. Means and standard deviations of the two simulated driving styles for vertical
accelerations for each driver, ntrial = 5.
Defensive DS Assertive DS
Driver (Targeted acc. = 0.04g) (Targeted acc. = 0.38g)
Mean (g) SD Mean (g) SD
A 0.07 0.01 0.35 0.08
B 0.07 0.02 0.31 0.03

In simulating the vertical accelerations, the only parameter to be controlled was the
approaching velocity. The main difference between a speed hump and a speed bump is that
the former is designed for a vehicle to traverse at a low speed (25-30 km/h) without the
need to slow down while the latter require a vehicle to decelerate up to 8 km/h when
arriving at the tip of the bump [29]. Therefore, in this study, the intended assertive driving
style required the drivers to traverse at the speed hump at a velocity of 36 km/h which is
relatively high for drivers with defensive or normal driving style. In addition, the available
speed hump for this experiment was relatively short at 4.07 m in length whereas a good
speed hump’s length is usually around 6.70 m or even extending to 9.50 m in some
countries like Denmark and 12.0 m in Australia [16]. A longer speed hump would induce
peak acceleration at a higher speed as usually a car has a wheelbase with a similar length to
the speed hump whereas a speed bump usually has a significantly shorter length compared
to a car’s wheelbase. Therefore, a speed bump would generate high vertical force even at a
low speed and, therefore, would definitely induce discomfort if the experiment were done
using the same conditions as in this study. Since only the velocity parameter was
considered in the current study, the intended accelerations were lower than measured in
defensive driving style simulation because the suspension effects were not included in the
initial calculations. As highlighted by Shirahatti [30], due to the vehicle suspension, when
the vehicle speed was increasing, the vertical acceleration was first increasing and then
decreasing at a high rate before stabilized at around 25 km/h. The study also indicated that

12
MATEC Web of Conferences 90 , 01005 (2017) DOI: 10.1051/ matecconf/20179001005
AiGEV 2016

the vehicle’s vertical displacement and pitch rotation were higher at low speed and came to
be lower at high speed, similar to the finding in the current study in simulating assertive
driving style.

3.3 General discussion

The simulation of autonomous driving styles at three road profiles namely straight lane
approaching and leaving a junction, cornering, and speed hump could be done with the help
of a special device known as AUTOAccD. The experiment set up started with the
determination of interested driving styles and then followed with the selection of routes and
determination of ranges for acceleration in different directions. In this study, the
determination of driving styles was influenced by the intention to simulate autonomous
driving experience with different driving styles. Therefore, three types of driving styles
were determined, namely assertive, defensive and LRT driving style. The accelerations of
LRT was exclusively included as riding in an LRT could already be a mental benchmark
for future AV users as the experience of using an LRT was the closest one would get in
imagining using an AV. This is in line with the public opinion regarding AV reported by
Schoettle and Sivak in 2014 [31], who found that most of their questionnaire respondents
would like to socialize with friends and family, do work-related activities, sleeping or
resting, reading, or watching television or movies inside a future AV.
In this experiment set up, the implementation of AUTOAccD resulted in two main
benefits. First, this device manages to help the drivers in simulating the intended
accelerations by interactively showing and clustering the live acceleration data so that the
drivers could adjust the throttle position and speed in order to get the required simulations.
Second, the device also acts as a data logger that measure and stores various data for further
analysis. Although the AUTOAccD would help the designated drivers in simulating the
intended accelerations, practise is still required as it would help the drivers to get the
feeling on how to generate higher accuracy simulations with fewer deviations.
In terms of comparison between the designated drivers, most of the acceleration data
have small SDs and close to the intended means. In addition, both drivers managed to
simulate accelerations within the intended ranges as out of 110 trials altogether only 18
values were simulated outside the intended ranges. By comparing the three driving styles,
the assertive driving styles were the hardest to be simulated regardless of the road profiles.
It is evidenced by the higher SD values for assertive driving styles compare to the other two
driving styles. For the defensive driving styles, in general, although low deviations were
recorded the average means for the simulated accelerations are usually higher than the
intended ones. For the LRT driving styles, both means and SDs for the simulated
accelerations indicated variations between the two drivers. Unlike the other two driving
styles which distinguished patterns were shown, for the LRT driving styles both of the
drivers did not show any clear trend regarding the simulated accelerations.
A few cautions have to be taken before executing the experiment on the road such as for
the lateral accelerations where the cornering radius has to be maintained in order to
generate the specified centrifugal forces. For the longitudinal accelerations, a lower gear
preferably the first gear is recommended in order to achieve the intended accelerations. For
the longitudinal deceleration, although a specific starting velocity has been identified, the
simulation of deceleration required an act of balance between clutch and brake pedal and
skill to produce a continuous and smooth deceleration. On the other hand, the process of
analysing the accelerometers data began with the de-noising process in order to eliminate
the noises without flattening the valuable data, from the three accelerometers in the tri-axial
direction.

13
MATEC Web of Conferences 90 , 01005 (2017) DOI: 10.1051/ matecconf/20179001005
AiGEV 2016

4 Conclusions
In conclusion, this study of simulating AV driving styles could be extended with the
evaluation of acceptance and preference of future autonomous driving styles such as using
the subjective measurements (unipolar or bipolar comfort rating) and objective
measurements (continuous physiological measures for example the galvanic skin
conductance (GSR) and electrocardiogram (EKG)) [32]. The significant finding of the
present work could be further extended in the area of comfort study in future AV driving
styles and the effects of motion sickness where the role of the driver would change from the
sole controller of the vehicle to a passive user with a much-minimized role. In addition,
comfort in travelling is usually associated with the way the vehicle is controlled and driven.
In the near future, where accelerations are controlled by the algorithms and machines,
autonomous vehicle users who used to be the drivers may feel discomfort unless the driving
styles of the vehicle are somewhat similar to their own. Hence, there is a special need to
study the associations between simulated accelerations and driving styles in order to
enhance the autonomous driving experience for the future AV users.
If such experiments were to be extended, experts from different backgrounds, like
engineering and social sciences among others, should work together in order to solve these
projected issues effectively and efficiently. The rising of AV technology brings along the
benefits such as lowering the fatal accidents by taking the human out of the driving’s loop
and maximizing the traffic flow and usage of the road [33]. At the same time, studies of
human-machine interaction and acceptance should be explored extensively because as
marvellous as AV technology could be, humans are still the end users who will buy, use
and experience the AV in the coming future.

The authors would like to express their gratitude and special acknowledgements to Universiti
Teknikal Malaysia Melaka (UTeM) and Ministry of Education, Malaysia (KPM) for the funding of
the Ph.D. programs of Juffrizal Karjanto and Nidzamuddin Md. Yusof and also the post-doctoral
program of Muhammad Zahir Hassan in the Industrial Design Department, Eindhoven University of
Technology, The Netherlands.

References
1. KPMG, “Self-Driving Cars: Are We Ready?,” (2015)
2. KPMG, “Self-Driving Cars : The Next Revolution,” (2012)
3. T. M. Gasser and D. Westhoff, “BASt-study : Definitions of Automation and Legal
Issues in Germany,” (2012)
4. J. Elander, R. West, and D. French, “Behavioral correlates of individual differences in
road-traffic crash risk: An examination method and findings,” Psychol. Bull., vol. 113,
no. 2, pp. 279–294, (1993)
5. O. Taubman-Ben-Ari, M. Mikulincer, and O. Gillath, “The multidimensional driving
style inventory - Scale construct and validation,” Accid. Anal. Prev., vol. 36, no. 3, pp.
323–332, (2004)
6. T. Vaa, “From Gibson and Crooks to Damasio: The role of psychology in the
development of driver behaviour models,” Transp. Res. Part F Traffic Psychol. Behav.,
vol. 25, pp. 112–119, (2014)
7. H. Summala, “Towards understanding motivational and emotional factors in driver
behaviour: Comfort through satisficing,” in Modelling Driver Behaviour in Automotive
Environments: Critical Issues in Driver Interactions with Intelligent Transport
Systems, (2007), pp. 189–207
8. S. Le Vine, A. Zolfaghari, and J. Polak, “Autonomous cars : The tension between

14
MATEC Web of Conferences 90 , 01005 (2017) DOI: 10.1051/ matecconf/20179001005
AiGEV 2016

occupant experience and intersection capacity,” Transp. Res. Part C, vol. 52, pp. 1–14,
(2015)
9. C. Diels and J. E. Bos, “Self-driving carsickness,” Appl. Ergon., vol. 53, pp. 374–382,
(2016)
10. M. Sivak and B. Schoettle, “Motion Sickness in Self-Driving Vehicles (Report No.
UMTRI-2015-12),” (2015)
11. V. Bogdanović and N. Ruškić, “The research of vehicle acceleration at signalized
intersections,” Traffic Transp., vol. 25, no. 1, pp. 33–42, (2013)
12. I. El-Shawarby, H. Rakha, V. Inman, and G. Davis, “Evaluation of Driver Deceleration
Behavior at Signalized Intersections,” Transp. Res. Rec. J. Transp. Res. Board, pp. 29–
35, (2007)
13. D. Hugemann and D. Nickel, “Longitudinal and Lateral Accelerations in Normal Day
Driving,” in 6th International Conference of The Institute of Traffic Accident
Investigators, (2003)
14. Parsons Brinckerhoff Team, “California High Speed Train Program Environmental
Impact Report/Environmental Impact Statement: Engineering Criteria,” (2004)
15. TCRP, TCRP Report 155 – Track Design Handbook for Light Rail Transit, Second
Edition. (2012)
16. P. A. Weber and J. P. Braaksma, “Towards a North American geometric design
standard for speed humps,” ITE J. (Institute Transp. Eng., vol. 70, no. 1, pp. 30–34,
(2000)
17. R. Wang and S. M. Lukic, “Review of driving conditions prediction and driving style
recognition based control algorithms for hybrid electric vehicles,” 2011 IEEE Veh.
Power Propuls. Conf. VPPC 2011, (2011)
18. E. Gulian, G. Matthews, A. I. Glendon, D. R. Davies, and L. M. Debney, “Dimensions
of driver stress,” Ergonomics, vol. 32, no. 6, pp. 585–602, (1989)
19. J. Reason, A. Manstead, S. Stradling, J. Baxter, and K. Campbell, “Errors and
violations on the roads: a real distinction?.pdf,” Ergonomics, vol. 33, no. 10/11, pp.
1315–1332, (1990)
20. M. Kuderer, S. Gulati, and W. Burgard, “Learning Driving Styles for Autonomous
Vehicles from Demonstration,” 2015 IEEE Int. Conf. Robot. Autom., vol. 134, pp.
2641–2646, (2015)
21. S. Huang, “OBD-II Telematics Advanced Kit (Arduino MEGA 2560),” (2014)
[Online]. Available: [Link]
22. C. Taswell, “The what, how, and why of wavelet shrinkage denoising,” Comput. Sci.
Eng., vol. 2, no. 3, pp. 1–11, (2000)
23. C. W. Kang, C. H. Kang, and C. G. Park, “Wavelet Denoising Technique for
Improving of the Low Cost MEMS-GPS Integrated Navigation Sytems,” Inetrnational
Symp. GPS/GNSS, pp. 3–8, (2010)
24. A. M. Hasan, K. Samsudin, A. R. Ramli, and R. S. Azmir, “Wavelet-based pre-filtering
for low cost inertial sensors,” J. Appl. Sci., vol. 10, no. 19, pp. 2217–2230, (2010)
25. M. Misiti, Y. Misiti, G. Oppenheim, and J.-M. Poggi, Wavelet toolbox. (1996)
26. S. Brol and J. Mamala, “Assessment of passenger car driveability with use of two axis
accelerometer mounted on car body,” J. Ultragarsas, vol. 2, no. 2, pp. 31–36, (2006)
27. G. Reymond, a Kemeny, J. Droulez, and a Berthoz, “Role of lateral acceleration in
curve driving: driver model and experiments on a real vehicle and a driving
simulator.,” Hum. Factors, vol. 43, no. 3, pp. 483–495, (2001)
28. S. G. Charlton, “The role of attention in horizontal curves: A comparison of advance
warning, delineation, and road marking treatments,” Accid. Anal. Prev., vol. 39, no. 5,
pp. 873–885, (2007)
29. Institute of Transportation Engineers, “Traffic Calming: State of the Practice - Chapter

15
MATEC Web of Conferences 90 , 01005 (2017) DOI: 10.1051/ matecconf/20179001005
AiGEV 2016

6: Legal Authority and Liability,” (1999)

30. A. Shirahatti, “Investigation on effect of hump on a vehicle suspension for the
comfortable ride,” in National Conference on Emerging Trends in Mechanical, (2006)
31. B. Schoettle and M. Sivak, “Public Opinion About Self-Driving Vehicles in China,
India, Japan, The U.S., The U.K. and Australia,” (2014)
32. J. a. Healey and R. W. Picard, “Detecting Stress During Real-World Driving Tasks
Using Physiological Sensors,” IEEE Trans. Intell. Transp. Syst., vol. 6, no. 2, pp. 156–
166, (2005)
33. D. J. Fagnant and K. Kockelman, “Preparing a Nation for Autonomous Vehicles:
Opportunities, Barriers and Policy Recommendations for Capitalizing on Self-Driven
Vehicles,” in 93rd Annual Meeting of the Transportation Research Board 2014, (2014)

16