1020 The International Arab Journal of Information Technology Vol. 13, No.

6B, 2016

Unmanned Vehicle Trajectory Tracking by Neural

Networks
Samira Chouraqui and Boumediene Selma
Department of Computer Sciences, University of Sciences and Technologies of Oran USTO’MB, Algeria

Abstract: This paper, deals with a path planning and intelligent control of an autonomous vehicle which should move safely
in its road partially structured. This road, involves a number of obstacles like donkey, traffic lights and other vehicles. In this
paper, the Neural Networks (NN)-based technique Artificial Neural Network (ANN) is described to solve the motion-planning
problem in Unmanned Vehicle (UV) control. This is accomplished by choosing the appropriate inputs/outputs and by carefully
training the ANN. The network is supplied with distances of the closest obstacles around the vehicle to imitate what a human
driver would see. The output is the acceleration and steering of the vehicle. The network has been trained with a set of
strategic input-output. The results show the effectiveness of the technique used, the UV drives around avoiding obstacles.

Keywords: UV, NN, control.

Received March 22, 2013; accepted March 19, 2014;published online June 11, 2015

1. Introduction solution to autonomous navigation problem because of

their ability to learn complex non linear relationships
Modern systems are becoming increasingly more between input values and output control variables. This
complex, making conventional control algorithms that ability of NN has attracted many researchers across the
utilize mathematical models insufficient. This resulted globe in developing NN based controllers for reactive
in the evolution of soft computing techniques [5] based navigation of mobile robots or vehicles in its
on biological processes such as learning and environments.
evolutionary development. Soft computing techniques This paper, investigates the possibility of applying
model human intelligence, providing decisions given NN predictive control to mobile vehicle in its road.
imprecise and uncertain information [1]. The three The main objective is to have a vehicle that drives by
most commonly employed soft computing tools are itself and avoids obstacles in a virtual world. Every
fuzzy logic, Neural Networks (NN) and genetic instant, the vehicle decides by itself how to modify its
algorithms [4, 6, 8]. speed and direction according to its environment. In
One of the most widely researched topics in the soft order to make it more real, the NN should only see
computing community is autonomous vehicles. The what a person would see if it was driving, so the UV
term autonomous represents the elimination of human decision is only based on obstacles that are in front of
control. Thus, autonomous vehicles generally refer to the vehicle. By having a realistic input, the NN could
vehicular systems capable of performing navigation possibly be used in a real vehicle and work just as well.
tasks without human guidance. The organization of the paper is as follows: Section
Autonomous vehicles are a promising technology 2 brings out the research method which describes the
with the ability for reduced traffic congestion and navigation system developed section 3 give the briefly
decreased accidental rates. Common tasks which description of the NN. Section 4 describes the
require automation include lane-following, parking, implementation of the NN controller. Simulation
collision avoidance, vehicle following, traffic signal results and discussions are given in section 5. Section 6
detection and navigation at intersections. will summarize our conclusions and gives the notes for
Unmanned Vehicle (UV) has the ability of a mobile our further research in this area.
robot to reach the set targets by avoiding obstacles in
its way. Thus, essential behaviors for UV navigation
are obstacle avoidance and goal reaching [3].
2. The Research Method
Conventional control techniques can be used to build The objective of the navigation system developed in
controllers for these behaviors; however, the this work consists of driving the vehicle to follow a
environment uncertainty imposes a serious problem in reference path from an initial point to a final one in a
developing the complete mathematical model of the partially structured environment as shown in Figure 1.
system resulting in limited usability of these Unexpected fixed obstacles are considered, shown in
controllers. Amongst the various artificial intelligence Figure 2.
techniques available in literature, NN offer promising
Unmanned Vehicle Trajectory Tracking by Neural Networks 1021

H h (t )F ( Net h (t )) (1)

H h (t )F  (W hi A i (t )) (2)
i

Where Neth represents the total input to the node h in

the hidden layer and F(x) is the activation function,
which has to be differentiable. In this paper, the
activation function is the sigmoid function.
Figure 1. Reference path. x
j
e 1
F ( x )  x
(3)
e 1 1e
x

The output of the node c in the output layer Cc(t) is:

C c (t ) F ( Net c (t )) (4)
H h (t ) F  (W ch F (W hi Ai (t ))) (5)
h

Figure 2. Obstacles avoidance (speed bump, traffic lights and Where Netc represents the total input to the node C in
vehicles exceeding). the output layer.
The second type of operation of the back
3. Neural Networks propagation NN is called error back propagation,
which is marked by dashed lines in Figure 3. The sum
NN [2] are powerful tool for the identification of of the square of the differences between the desired
systems typically encountered in the structural output Lc(t) and NN outputs C(t) is:
dynamics fields. NN were originally developed to 1
E  ( Lc (t)  C c (t))
2
simulate the function of the human brain or neural
system. Artificial Neural Network (ANN) is basically a 2 h

massive parallel computational model that imitates the The adaptive rule for the weight Wch as the connections
human brain. This method does not really solve between the hidden layer and output layer, can be
problems in a strictly mathematical sense, but they are determined as:
one method of relaxation that gives an approximate
W ch (t  t ) W ch (t)  W ch (7)
solution to problems. A number of NN techniques have
been used in system identification such as E
backpropagation network [8], Hopfield network and W ch   (8)
W ch
Kohonen network. In the present paper, the most
widely used technique; the back propagation NN is W ch    c (t ) H h (t ) (9)
adapted for the identification of a structural dynamic i

model. The principles of the back propagation NN are dF ( Net c )

shown in the following. c (t ) ( Lc (t )  C c (t )) (10)
dNet c
A typical three-layer back propagation NN is shown
in Figure 3 and consisted of the next: The input layer The adaptive rule for connections between the input
with a nodes, the hidden layer with b nodes and output layer and the hidden layer Whc as:
layer with c nodes. Between layers there are weights E
Wha and Wch representing the strength of connections of W ha   (11)
the nodes in the network. The first type of operation of W ha
back propagation NN is called feed forward and is W ha     h (t ) Aa (t )
shown as solid lines with arrow in Figure 3. For this i
(12)
operation, the output vector C(t) is calculated by dF ( Net h )
feeding the input vector A(t) through the hidden layer  h (t )  W hc c t  (13)
of the neural network. The output of the node h in the dNet h c

hidden layer Hh(t) for the given input layer A(t) is: The coefficient η is called the learning rate. The error
back propagation rules shown in Equations 8 and 13
with applying the differentiation process successively
can be expanded to the networks with any number of
hidden layers. The weights in the network are
continuously adjusted until the inputs and outputs
reach the desired relationship.

4. The System Model

The neural system to control the vehicle in its road is
Figure 3. Three layer back propagation NN. modelled as shown in Figure 4.
1022 The International Arab Journal of Information Technology Vol. 13, No. 6B, 2016

As input of the neural system the vector of positions

(X, Y) which characterize four positions: The position
of every object, the position of the road, the position of
the vehicle and also, the position of obstacles. The
second information is the velocity V of the vehicle and
finally the Obstacle O.
Input Layer Hidden Layer Output Layer
Figure 5. Learning about sample reference path.

Figure 4. NN Architecture. Figure 6. The path found by the NN.

The output needs to control the vehicle’s speed and

direction. That would be the acceleration, the brake
and the steering wheel. So, three outputs are needed,
one will be the acceleration/brake since the brake is
just a negative acceleration and the others will be the
positions.
In Table 1 different arrangements of obstacle
relative to the vehicle, velocity and the desired reaction Figure 7. The error between the reference path and the path
from the UV are visualized. calculated.
Table 1. Vehicle situations on the road.
Output Neurons Relative
Input Neurons Relative to Obstacle and Velocity
to Velocity
Obstacle Velocity Antecedent Acceleration Consequent
No Obstacle Full Full
No Action
Acceleration Acceleration
Low
No Obstacle Accelerate
Acceleration
Full
Donkey Distance donkey is D Slow down
Acceleration
Traffics Full Color traffic lights is
Lights Acceleration Green
Slow down Figure 8. Generation of rules.
Traffics Full Color traffic lights is
Stop
Lights Acceleration Red The function of the speed controller subsystem is to
Our vehicle speed is Change of
Vehicle
Full
greater than the vehicle
Full
position and
achieve the desired speed that is to say the increase and
Acceleration Acceleration
on the road exceeding decrease in speed on the road and especially in front of
Speed of your vehicle is obstacles, so the accelerator control. Figure 9 shows
Vehicle Full Full
smaller than that of No Action
Acceleration Acceleration
vehicle in road the variation of the UV speed in function of obstacles
come across the road.
5. Simulation Results and Discussions
In this section, to show the contribution of the control
by NN approach, simulation was approved on an UV.
First, Figure 5 shows the path followed by the UV
controlled by the proposed NN including obstacles
described in the above sections.
Figure 6 shows the vehicle’s path obtained after
controlling with the NN and Figure 7 gives the error
Figure 9. Vehicle speed variations at obstacles.
calculated between the reference paths and obtained
one after application of control. From Figure 9, it can be seen that the UV can
As it can be seen from Figures 6 and 7 the path indeed avoid obstacles and reach the targets. To verify
obtained from simulation setup is more close to the the feasibility of proposed method Table 2 shows
reference path which validates the proposed method. results of NN controller.
Figure 8 shows the generations of rules.
Unmanned Vehicle Trajectory Tracking by Neural Networks 1023

Table 2. NN best simulations results. of Information Technology, vol. 7, no. 2, pp. 199-
Datasets Statistical Results 205, 2010.
Rate Accuracy (%)
Set X
Total Classification Accuracy (%)
98.43 98.77
[2] Bajpai S., Jain K., and Neeti J., “Artificial Neural
Set Y 99.11 Networks,” the International Journal of Soft
Computing and Engineering, vol. 1, no. 2011, pp.
Therefore, it can be concluded that the NN 2231-2307, 2011
controller have a good potential to effect fast response [3] Cuesta F. and Ollero A., Intelligent Mobile Robot
to obstacles and reduce errors. Navigation, Springer-Verlag, Berlin Heidelberg.
To show the performance of the results obtained by 2005.
our approach, an approach based on type-2 fuzzy logic [4] Goldberg D., Genetic Algorithms in Search,
theory and genetic algorithm. Optimization and Machine Learning, Addison
Algorithms [7] have been selected for comparison Wesley, New York, 1989.
because of its high capacity of prediction and control [5] Hoffmann F., “An Overview on Soft Computing
in non linear dynamical systems. in Behavior Based Robotics,” in proceedings of
the 10th International Fuzzy Systems Association
Table 3. Comparison of the NN results with GA-FL.
World Congress Istanbul, Turkey, pp. 544-551
Approach Average Error
NN 1.23
2003.
GA-FL 3.81 [6] Holland J., Adaptation in Natural and Artificial
Systems, University of Michigan Press, 1975.
The average error obtained with fuzzy logic [7] Jang J. and Sun C., Neuro-Fuzzy and Soft
combined with the genetic algorithm (GA-FL) is Computing: A computational Approach to
mentioned in Table 3 and by comparing with the Learning and Machine Intelligence, Prentice
results obtained by the NN algorithm, it is Hall, 1997.
demonstrated that neural networks can be used [8] Zadeh L., “Fuzzy Sets,” Information and
effectively for the identification and control of Control, vol. 8, no. 3, pp. 338-353, 1965.
nonlinear dynamical systems precisely in autonomous
vehicle path planning. Samira Chouraqui received her
MSc degree in Satallite and Systems
6. Conclusions Communication from Surrey
University UK and received PhD in
Automatic motion planning and navigation is the
applied mathematics from University
primary task of an automated guided vehicle or mobile
of Science and technology of Oran,
robots. All such navigation systems consist of a data
Algeria. Currently, she is teaching
collection system, a decision making system and a
numerical analysis and systems dynamics at the
hardware control system. In this research, our artificial
University of Oran Mohamed Boudiaf USTO of Oran
intelligence system is based on NN model for
(Algeria).
navigation of an UV in unpredictable and imprecise
environment. Boumediene Selma received his
We have designed a trajectory tracking controller Engineer degree in Computer
taking into account the obstacles using NN and we Science from the University of
have demonstrated that soft computing approaches are Science of Mostaganem Abdelhamid
more preferable over conventional methods of problem Ibn Badis, Algeria and actually is in
solving, for problems that are difficult to describe University of Science and
by analytical or mathematical models. Autonomous technology USTO of Oran, Algeria;
robotics is such a domain in which knowledge about preparing his Magister in systems dynamic attitude
the environment is inherently imprecise, unpredictable estimation and control using evolutionary
and incomplete. Therefore, the features of NNs are of Techniques. His current research interests are in the
particular benefit to the type of problems emerging in area of artificial intelligence and mobile robotics,
behaviour based robotics. Soft computing techniques Pattern Recognition, neural networks, neuro-fuzzy and
contribute to one of the long term goal in autonomous data-mining.
robotics, to solve the problems that are unpredictable
and imprecise namely in unstructured real-world
environments.

References
[1] Awad M., “An unsupervised Artificial Neural
Network Method for Satellite Image
Segmentation,” the International Arabic Journal