Page 1 of 11

Transactions on Machine Learning and Artificial Intelligence - Vol. 9, No. 5

Publication Date: October, 25, 2021

DOI:10.14738/tmlai.95.10795. Nabil, A., & Kassem, A. (2021). Self-Driving Vehicles Using End to End Deep Imitation Learning. Transactions on Machine Learning

and Artificial Intelligence, 9(5). 33-43.

Services for Science and Education – United Kingdom

Self-Driving Vehicles Using End to End Deep Imitation Learning

Ashraf Nabil

Aerospace Engineering, Cairo University, Cairo, Egypt

Ayman Kassem

Professor in Aerospace Engineering at Cairo University

ABSTRACT

Autonomous Driving is one of the difficult problems faced the automotive

applications. Nowadays, it is restricted due to the presence of some laws that

prevent cars from being fully autonomous for the fear of accidents occurrence.

Researchers try to improve the accuracy and safety of their models with the aim of

having a strong push against these restricted Laws. Autonomous driving is a sought- after solution which isn’t easily solved by classical approaches. Deep Learning is

considered a strong Artificial Intelligence paradigm which can teach machines how

to behave in difficult situations. It proved its success in many differ domains, but it

needs sometimes to prove itself in the automotive applications.The presented work

will use the end-to-end deep machine learning field in order to reach to our goal of

having Full Autonomous Driving Vehicle that can behave correctly in different

scenarios. CARLA simulator will be used to initially train and test the deep neural

networks. The resulting DNN models are further trained using one of the most

popular real datasets of automotive that includes camera images with the

corresponding driver’s control action [1]. This data mixing approach shows

improvements in the DNN models prediction.

Keywords: Autonomous Vehicles; Self-driving; Machine learning; Deep learning;

Convolution Neural Networks; Control.

INTRODUCTION

According to the World Health Organization (WHO), approximately 1.35 million people die in

the world every year due to vehicle accidents and between 20 to 50 million more people are

injured/disabled. Many of these accidents are preventable, and an alarming number of them

are a result of distracted driving [2].

From this perspective the automotive companies and many research centers and universities

started looking for solutions for such drastic issue. An autonomous car is a vehicle that can

guide itself without human interaction.

Autonomous vehicles depend on sensors, actuators, complex control algorithms, machine

learning systems, and powerful processors to run software. Autonomous vehicles contain 4

main blocks from system and software point of view as shown in figure 1. First is the data

sensors level followed by the perception layer then the motion planning layer and finally the

control layer. Each one of those will be discussed in detail in the next section.

Page 2 of 11

34

Transactions on Machine Learning and Artificial Intelligence (TMLAI) Vol 9, Issue 5, October - 2021

Services for Science and Education – United Kingdom

Figure 1 Self-Driving Vehicles Pipeline [3]

LITERATURE SOLUTIONS

Making vehicle autonomous mostly depends on one of 2 approaches:

1. The first is a modular pipeline that relies on dedicated subsystems for visual perception,

planning, and control. This architecture is in line with most existing autonomous driving

systems [4].

2. The second approach is based on a deep network trained end-to-end via imitation

learning. This approach represents a long line of investigation that has recently attracted

renewed interest [1, 5].

Modular Pipeline Approach

First method is a modular pipeline that decomposes the driving task among the following

subsystems:

Sensors

It is the first layer in both approaches which describes how many sensors are mounted on the

vehicle and what are they. Since the autonomous driving is a sensitive field and may lead to

many accidents, usually vehicles must be equipped with different sensors. Each sensor has its

own pros and cons. Some of the most popular sensors are:

• Radar: it is very accurate in getting the object’s velocities (based on doppler effect).

• Camera: it is very precise in classifying the object’s types.

• Lidar (laser scanner): it is very accurate in measuring the object’s position even for a

long-range object.

Perception

Frequently provide and update a map of the surroundings based on a different sensor mounted

in different parts of the vehicle like radars (monitor the position and velocity of nearby

vehicles), cameras (detect traffic lights, read road signs, track other vehicles and pedestrians),