See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/347539333

Autonomous Car Driving Using Neural Networks

Article · January 2019

CITATIONS READS

3 1,369

4 authors, including:

Ahmed Kateb Ahmed Radhi

Al-Manara College For Medical Sciences Engineering Technical College of Misan
11 PUBLICATIONS 56 CITATIONS 11 PUBLICATIONS 32 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Ahmed Radhi on 22 December 2020.

The user has requested enhancement of the downloaded file.

 www.gsj-pub.com Journal of
Journal of Global Scientific Research (ISSN: 2523-9376) Global Scientific
2 (2019) 240-248
Research

www.gsj-pub.com/jgsr

Autonomous Car Driving Using Neural Networks

*1
Ahmed Kateb Jumaah Al-Nussairi & 2Ahmed Thamer Radhi
1
Al-Manara College for Medical Sciences, Missan, Iraq
2
Electronic & Communication Department, Southern Technical University, Technical Institute of
Amarah, Missan, Iraq.
*Email: [email protected], (Al-Nussairi), [email protected] (Radhi)

Abstract
The goal of this paper is to design a functional software architecture for fully autonomous vehicles.
Existing literature takes a descriptive approach and presents past experiments with autonomous driving
or implementations specific to limited domains (e.g. winning a competition). The architectural
solutions are often an after-math of building or evolving an autonomous vehicle and not the result of a
clear software development lifecycle. A major issue of this approach is that requirements can not be
traced with respect to functional components and several components group most functionality.
Therefore, it is often difficult to adopt the proposals. In this paper we take a prescriptive approach
starting with requirements from an automotive standard. We use a NIST reference architecture for real-
time, intelligent, systems and well-established architectural patterns to support the design principles.
We further examine the results with respect to the automotive software development life cycle and
compliance with automotive safety standards. Lastly, we compare our work with other proposals.

Kewords: Intelligent vehicles, Autonomous vehicles, Software architecture.

1. Introduction the possibility of autonomous driving in real

We are living in the time has many cars, many life [1].

cars that’s mean collision between them and Google has been testing its autonomous

sometimes leading to die of drivers so think to vehicle in actual traffic conditions. In August

expert to find the idea and they found (car 2012, Google announced that they have

without driver) it is working according to the completed over 500,000 km autonomous

data that we will enter it. driving without any accident [2]. Another

Autonomous driving systems have been research group in Italy, the VisLab in Parma

actively researched. Several works have shown University, did 13,000 km test run for
autonomous vehicles from Italy to China [4].
Al-Nussairi, A. K. J. & Radhi, A. T. Journal of Global Scientific Research 241

Such a vehicles perceive the surrounding Figure 1 below shows the architecture of the
environment by camera sensors and fusion project. In this paper, we focus on the design
with other sensors. The VisLab also tested the and implementation of such a vehicle low-
autonomous vehicle in a real environment, level controller.
together with real traffic on July 2013. In May This paper is focused on the low frame of the
2014, Google presented a new concept for automatic control of the speed and the steering
their driverless car that had neither a steering angle of the car following a predefined path
wheel nor pedals [5], and unveiled a fully with the best performances of stability and
functioning prototype in December 2014 that precision. Aware that a good automatic car
they planned to test on San Francisco Bay control is the basis for achieving the other
Area roads beginning in 2015. Google plans to challenges of the autonomous driving systems.
make these cars available to the public in 2020 We propose two strategies of non-linear
[6]. automatic low level control, based on the
The Computer Vision Center (CVC) is method of Lyapunov [12] and based on sliding
automat zing an electric car within the context mode, and a comparison of both has been
of the project Automated and Cooperative made in a simple simulator (based on
Driving in the City. Simulink). Currently, it is being tested in a
complex simulator developed in Unity 3D2
2. System Architecture of the (see Figure 3).

Autonomous Vehicle

Fig. 1. System architecture of the autonomous vehicle

Al-Nussairi, A. K. J. & Radhi, A. T. Journal of Global Scientific Research 242

3. Model of Control Oriented movement in z axis is also neglected and

angles like steering and yaw are assumed to be
Vehicle
small. Then, the set of kinematic equations of
For control design, the autonomous car has
the cartesian position (xr;yr) and orientation
been considered as a bicycle-like vehicle ([7],
(Ɵr) of the real vehicle is presented as follows:
[11] ) positioned at a nonzero distance with
respect to a dynamic waypoint (virtual car of
reference), whose motion is controlled by the Xr= vrsin(Ɵr)

combined action of both the angular velocity Yr= vrCOS(Ɵr) …………… (1)
wr(t) and the linear velocity vr(t) of the real
Ɵr=vr/I tan(Ør)
vehicle (Figures 2 and 3). This model assumes
where vr and Ør represent the linear
that the vehicle is symmetric, the steering
velocity and the steering angle respectively.
angle is the same in both front wheels, the roll
Consequently, from Fig. 3, the kinematic
and pitch movement are neglected, the linear
equations for the virtual car can be defined as:

Fig. 2. Bicycle model Fig. 3. Vehicle’s position and orientation with

respect to the dynamic target frame (virtual car).

X d= vdsin(Ɵ d)

Y d= vdCOS(Ɵ d) ………………… (2)

Ɵ d= vd/I tan(Ø d)
where xd, yd and Ø d are the position and orientation of the next way point generated by the
trajectory planner.
The error model (3) is defined as the difference between real vehicle position and the desired one
multiplied by the rotation matrix over z axis which is the orthogonal to the road plane:

Xe CosƟ d SinƟd 0 xr - xd

Ye = - SinƟd CosƟ d 0 yr -yd ………... (3)

Ɵe 0 0 1 Ɵd -Ɵr
Al-Nussairi, A. K. J. & Radhi, A. T. Journal of Global Scientific Research 243

That after some algebraic manipulations lead to the following expression:

xe = vrcos(Ɵe)+ ye vd/I tan(Ø d) – vd

ye = vrsin(Ɵe) – xe vd/I tan(Ø d) ………………………………..(4)

Ɵe= vr/I tan(Ør ) - vd/I tan(Ø d)
that can be expressed as follows taking into account the real and reference vehicle models (1) - (2):
xe = vr +wrye - vrcos(Ɵe)
ye = - wrxe+vdcos(Ɵe) ……………………………………………………..(5)
Ɵe = wr –wd

Where : Ɵe = wr –wd

4. Description of Automatic Control define a control law assuring the stability and
Strategies the asymptotic elimination of the following

The autonomous car control objective of path error. On the other hand, the basic idea of

following via way points consists to reach SMC is to reach a sliding surface in finite time

asymptotically to zero the difference between and remain on this. However, this control

the dynamic position and orientation of the approach has a drawback: the chattering, i.e. a

real car respect to the dynamic way points trajectory oscillation over the sliding surface.

(virtual car) position and orientation.

In this paper, two nonlinear automatic control There are several ways of dealing with this

strategies for autonomous vehicles for path problem e.g., using a higher order sliding

following and navigation direct method of surface or smoother functions instead of the

Lyapunov [12] and the other based on sliding common sign function;

mode control (SMC) [10]. Both techniques

will consider as starting point the error model 4.1 Direct Lyapunov approach

derived from the vehicle control-oriented We are going to design a control based on the
model presented in previous section. direct Lyapunov approach. This method
On one hand, the idea of nonlinear control guaranties the asymptotic stability of the
based on the direct Lyapunov method is to vehicle control because:

Xe
Lim Ye = 0 …………………………………..(6)
t ∞ Ɵe

which involves also:

xr - xd
Lim yr –yd =0 ……………………….(7)
t ∞ Ɵr- Ɵr
Al-Nussairi, A. K. J. & Radhi, A. T. Journal of Global Scientific Research 244

As control law we propose to use the non linear law from[12]:

Vr vdcosƟe - k1xe
W = wd- k2vd sinƟe/Ɵe ye - yeƟe ………………………(8)

Given the following Lyapunov function:

V= 1/2 Xe+1/2 Ye+ 1/2 Ɵe ………………………………………………(9)

The stability condition is achieved when ˙V <= 0. Thus, taking into account the error model (5), the
proposed control law (8) and substituting in the derivative of (9) and after some simplifications we
obtain the following expression:
V= - k1k2xe - k3Ɵe <=0 ………………………………………………………………………….(10)

which implies that the control parameters k1, k2 and k3 should be positive to assure the asymptotic
stability of the closed-loop.

4.2 Sliding mode control the mathematical point of view a sliding

The main idea behind this approach is to reach surface is an expression composed by states of
the sliding surface in a finite time and remain the system to be minimized:
on such surfaces where the error is null. From

Fig. 4. Sliding Mode behavior

Fig. 6 shows the basic principle of the method smooth functions instead of the common used
and it can be seen how the trajectory reaches sign function.
the surface s and remains there. However, the In this paper, a stable SMC has been designed
trajectory presents some oscillations around which has as inputs the errors (3), desired
the surface. This is known as the chattering linear velocity (vd), desired angular velocity
phenomena. This drawback can be reduced by (wd), desired linear acceleration(vd) and
using a higher order sliding surfaces or using desired angular acceleration (wd), and as
Al-Nussairi, A. K. J. & Radhi, A. T. Journal of Global Scientific Research 245

outputs the velocity (vr) and the steering angle one sliding surface have to contain two of such
(Ɵr). error variables. It has been decided to couple
In order to implement it, a set of sliding ye and Ɵe in the same sliding equation,
surfaces are chosen. There is a surface si per therefore the resulting surfaces are the
control action. Given that there are three error following:
variables (xe, ye and Ɵe) and only two surfaces,

s1 = Xe + K1+ Xe ………………………………………………………..(11)
s2 = Ye+ k2ye+ k3Ɵe ……………………………………………………(12)
where k1, k2 and k3 are positive defined parameters. According to [9], the dynamics of the sliding
surface is the following, which is called the reaching law:
si= - Qisi - Pisign(si) …………………………………………………..(13)

where Q and P are positive defined parameters and its stability can be proven using Lyapunov
theorem [10]. A Lyapunov candidate function and its time derivative is defined as follows:
V = 1/2 s' s ……………………………………..(14)
Evaluating its derivative:

V= s' s …………………………………………(15)

and considering the control law (13), it can be expressed as:

V= s1(- Q1s1-P1sgn(s1))+ s2(- Q2s2- P2sgn(s2))……………………..(16)

We are developing them, the following control laws are obtained:

Vc= -vrƟesin(Ɵe) – yewd - yewd+vd - k1xe - Q1s1-P1sgn(s1) / cos(Ɵe) ……………..(17)

We will take integration for the equation:

Vc= ∫ vcdt ………………………..(18)
Ɵc= wd+ - P2s2 - Q2sign(s2) - k2ye+wdxe+wdxe - vcsin(Ɵe) / vrsin(Ɵe)+k3 …………………(19)
Øc = atan(1/Vr Øc) ……………………………………..(20)
The finally:
Equations (17) and (18) represent the velocity control law while (19) and (20) represent the
steering control law.

5. Application Matlab/Simulink and it is also currently

In this section, the results of the previous developed in the Unity platform.
control methods are presented in simulation. In parallel with the implementation of the
The simulation has been developed in controller, a trajectory planner has also been
implemented which provides the specific
Al-Nussairi, A. K. J. & Radhi, A. T. Journal of Global Scientific Research 246

instructions to the control area. The next steps takes a sub way point features as a desired
are followed to perform the trajectory tracking: configuration and perform the control.
_ The GPS provides to the vehicle a set of Figure 6 shows the total path to be followed,
forward way points at every segment. The where blue circles represent the way points and
space between two way points is called between them there are a set of sub way points.
segment. Figure 5 shows the desired set of velocities for
_ When a segment finishes, the planner takes the whole path computed by the planner. Input
the next way point and perform the correct disturbances have been included in the sensors
speed profile according to the maximum data and in the model as Gaussian random
acceleration allowed. From this segment a set noises to make the simulation more realistic.
of sub way points are calculated with its Inside the vehicle model, two first order filters
respective position, orientation, linear velocity, have been considered in order to mitigate the
angular velocity, linear acceleration and high frequency terms of the control signals in
angular acceleration. case of the sliding model controller. In next
_ Once such a segment has been sampled, at subsections the results of the control methods
every sample time (Ts = 0.1s) the control area are presented.

Fig. 5. Desired angular and linear velocities Fig. 6. Path proposed to prove control techniques

A. Direct Lyapunov based controller results resulting path, the control actions and the
The controller is designed according the errors. It can be seen that the errors are
procedure described in Section III.A. The sufficiently small. For instance in the case of
design parameters are adjusted as follows: k1 = lateral vehicle error the maximum value
0:9, k2 = 1:1 and k3 = 3. They have been reached when the car arrives to a curve is 5 cm.
adjusted by trial and error. Fig. 6 presents the
Al-Nussairi, A. K. J. & Radhi, A. T. Journal of Global Scientific Research 247

B. Sliding mode control results sliding parameters are shown in Table I, which
The controller is designed according the have been adjusted by trial and error too.
procedure describes in Section III.B. The set of
K1 0.22
K2 2
K3 2.55
P1 0.48
Q1 0.048
P2 3.7
Q2 0.3
TABLE I
SLIDING MODE CONTROL PARAMETERS

6. Results and discussion used in the literature for modeling autonomous

By observing the results, we can admit both cars. The two control strategies follow a model

methods provide similar results and they are reference approach. Using this approach, the

very good. error dynamics model has been developed.

Note that the steering angle control action of Both controllers receive as input the

the Lyapunov technique is smother. It is due to longitudinal, lateral and orientation errors

the fact that the SMC method perform the generating as control outputs the steering angle

steering signal by using the three errors and the and the velocity of the vehicle. The first

acceleration action, and all these variables nonlinear control approach has been designed

have noise. On the other hand, Lyapunov by means of the Lyapunov direct approach.

technique only uses two of the three errors to The second approach has been designed using

compute the steering angle. the sliding mode approach. Both controllers
have been implemented, tested and compared

Notice also that the sliding mode method with different path following scenarios in

achieves smaller errors although such an errors simulation. From the obtained results, both

can be minimized by obtaining a better set of methods provide similar results being quite

parameters in both techniques. Both control robust to uncertainty and noise. As already

methods have demonstrated to be robust with commented, currently, both control strategies

respect to some noise and disturbances, and the are being tested on a virtual reality simulation

obtained results show the effectiveness of such developed in Unity before being tested in a

proposed control. real car available at the Computer Vision

Center.

7. Conclusions
8. References
This work has presented the comparison of
[1] A. Carvalho, S. Lefvre, G. Schildbach, J.
two nonlinear model-based control strategies
Kong and F. Borrelli, Automated driving:
for autonomous cars. Both controllers have
The role of forecasts and uncertaintyA
been designed using a control oriented model
of vehicle based on a bicycle model commonly
 Al-Nussairi, A. K. J. & Radhi, A. T. Journal of Global Scientific Research 248

control perspective, European Journal of [8] B. Siciliano, O. Khatib, and F. Groen, The
Control, 24,pp 14-32, , 2015. DARPA Urban Challenge: and Automation
[2] C. Urmson. ”The self-driving car logs Magazine, pp. 27-35, March 1995.
more miles on new wheels” Googleblog, Autonomous Vehicles in City Traffic, vol.
Aug. 2012. Available: 56. Springer, 2009.
http://googleblog.blogspot.hu/2012/08/the- [9] W. Gao, John Y. Hung and James C. Hung.
self-driving-car-logsmore- miles-on.html Variable Structure Control: A Survey.
[3] J. Muller. ”With Driverless Cars, Once IEEE, 1993.
Again It Is California Leading The Way. ” [10] J. Slotine and W. Li. Applied Nonlinear
Forbes.com, Sep. 2012. Available: Control. Prentice-Hall, 1991.
http://www.forbes.com/sites/joannmuller/2 [11] R. Solea and U. Nunes, ”Trajectory
012/09/26/with-driverlesscars- once-again- Planning and Sliding-Mode Control Based
it-iscalifornia-leading-the-way/ Trajectory Tracking for Cybercars”,
[4] A. Broggi, ”VIAC: The VisLab Integrated Computeraided Engineering, Int.
Intercontinental Autonomous Challenge.” Journal, IOS Press, vol.14, n.1, pp.33-47,
Vislab, Nov. 2010. Available: 2007.
http://viac.vislab.it [12] W. Dixon, D. Dawson, E. Zergeroglu and
[5] L. Gannes. ”Google Introduces New Self A.Behal, Nonlinear Control of Wheeled
Driving Car at the Code Conference - Mobile, Springer-Verlag New York, Inc.
Re/code”. Re/code, 2014. Secaucus, NJ, USA 2001 View publication.
[6] T. Halleck, ”Google Inc. Says Self-Driving
Car Will Be Ready By2020”. International
Business Times, 2015.
[7] M. Aicardi, G. Casalino, A.Bicchi, A.
Balestrino, ”Closed loop stering of
unicycle-like vehicles via Lyapunov
techniques”, IEEE Robotics.

View publication stats