Improving Traffic Flow for Emergency Vehicles

Using Deep Learning Techniques

A Dissertation
Submitted to the Council of the College of Erbil Technical Engineering
College at Erbil Polytechnic University in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy in Information System
Engineering

By
Kamaran Hussein Khdir Manguri
B.Sc. in Computer Systems Engineering (2011)
M.Sc. in Electronics and Computer Engineering (2016)

Supervised by
Prof. Dr. Aree Ali Mohammed

Erbil, Kurdistan
August 2024
 DECLARATION

I declare that the PhD. Dissertation entitled “Improving Traffic Flow for
Emergency Vehicles Using Deep Learning Techniques” is my own original
work, and hereby I certify that unless stated, all work contained within this
dissertation is my own independent research and has not been submitted for the
award of any other degree at any institution, except where due acknowledgment
is made in the text.

Signature:

Student Name: Kamaran Hussein Khdir Manguri

Date: August/ 2024

I
 CERTIFICATE OF PROOFREADING

This is to certify that this dissertation entitled: “Improving Traffic Flow for
Emergency Vehicles Using Deep Learning Techniques” written by the
postgraduate student (Kamaran Hussein Khdir Manguri) has been proofread
and checked for grammatical, punctuation, and spelling mistakes. Therefore,
after making all the required corrections by the student for further
improvement, I confirm that this last copy of the dissertation is ready for
submission.

Signature:
Name: Asst. Prof. Dr. Salih Ibrahim Ahmed
Phone No.: 07701502771
Email Address: [email protected]
Date: 01/07/2024

II
 SUPERVISOR CERTIFICATE

This dissertation has been written under my supervision and has been submitted
for the award of the degree of Doctor of Philosophy in Information System
Engineering with my approval as supervisor.

Signature
Name: Prof. Dr. Aree Ali Mohammed

Date: / / 2024

I confirm that all requirements have been fulfilled.

Signature:
Name: Byad A. Ahmed
Head of the Department of Information Systems Engineering
Date: / / 2024

I confirm that all requirements have been fulfilled.

Postgraduate Office

Signature:
Name:
Date:

III
 EXAMINING COMMITTEE CERTIFICATION

We certify that we have read this Dissertation “Improving Traffic Flow for
Emergency Vehicles Using Deep Learning Techniques” and as an examining
committee examined the student (Kamaran Hussein Khdir Manguri) in its
content and what related to it. We approve that it meets the standards of a
dissertation for the degree of Doctor of Philosophy in Information System
Engineering.

Signature: Signature:
Name: Prof. Dr. Mazen R. Khalil Name: Asst. Prof. Dr. Moayad Y. Potrus
Chairman Member
Date: / / 2024 Date: / / 2024

Signature: Signature
Name: Assist. Prof. Dr. Ismael K. Name: Assist. Prof. Dr. Azhin T. Sabir
Abdulrahman (Member) Member
Date: / / 2024 Date: / / 2024

Signature: Signature
Name: Assist. Prof. Dr. Shahab W. Name: Prof. Dr. Aree A. Mohammed
Kareem (Member) Supervisor
Date: / / 2024 Date: / / 2024

Signature
Name: Prof. Dr. Ayad Z. Sabir Agha
Dean of Erbil Technical Engineering College
Date: / / 2024

IV
 DEDICATION

This Dissertation is dedicated to:

My merciful parents
My best friend and lovely wife (Chopy) for her beliefs and supports
My Son, Kovan
My siblings.

V
 ACKNOWLEDGMENTS

Expressing my gratitude to Allah for His blessings and guidance

throughout this journey, I want to extend deep thanks to all who supported me
in completing my PhD dissertation. A special acknowledgment goes to my
supervisor, Prof. Dr. Aree Ali Mohammed, for their invaluable guidance and
unwavering encouragement.

I also wish to express heartfelt appreciation to my dissertation committee

members for their significant contributions and insightful feedback.

Furthermore, I would like to express my gratitude for the crucial support

provided by the Department of Information System Engineering at Erbil
Polytechnic University, which significantly aided in facilitating my research.
Additionally, I appreciate the opportunity given by the University of Raparin.
I am grateful for the camaraderie and encouragement from colleagues and
friends, as well as the constant support from my family.

Lastly, I wish to thank all individuals who participated in this research

project, as their contributions greatly enriched the outcome of my dissertation.

VI
 ABSTRACT

The world's population has exponentially grown, which has an effect on

usage of vehicles by individuals and leads to an increase in the number of cars
in urbans. With the direct relationship between population and car usage, traffic
management has become an important issue to be solved. For this purpose, an
intelligent traffic signaling with a rapid urbanization is required to overcome
the traffic congestions, and reduce cost and time of traveling. To overcome
these problems, emerging computer vision and deep learning are vital
candidates to handle this issue because they take an important role for
managing and controlling traffic signals with great success. Nevertheless,
detecting and distinguishing between objects are helpful for counting vehicles
and other objects which avoid crowds and controlling signals in the traffic
areas. Besides, detecting emergency vehicles and giving the priority to them is
required for intelligent traffic signaling system.

The main objective of this study is to design and implement an efficient

system for traffic signal systems based on custom vehicle detection.
Furthermore, the proposed system involves four phases; the first one is
capturing images from both simulated and real time cameras from the roads. In
the second phase, different image preprocessing algorithms are performed to
the captured images as a pre-processing step. In addition, the deep learning
techniques are applied to detect objects such as (regular car, police car,
ambulance, and firefighter, etc..). In the last phase, the proposed system is
tested to evaluate the performance accuracy of the detected vehicles.

A modified transfer learning approach has been applied to the DenseNet201

model for multiple classifications, including non-emergency cars, ambulances,
police, and firefighters. The approach involves freezing the architecture of the
model's layers. A high accuracy rate is obtained with this model and reaches
98.6%. Also, various optimization methods, including (Adam, Adamax,

VII
Nadam, and RMSprob) are used to improve the detection performance based
on the best optimizer selection and yielded an accuracy of 98.84%. In addition,
a modified version of YOLOv5 was proposed for vehicle detection, which aims
to enhance the mean average precision (mAP) detection by 3%. Finally, the
proposed system was simulated to reduce the waiting time at traffic signal. The
experimental results demonstrate a significant reduction in waiting time,
ranging from 30 to 100 seconds depending on the status.

VIII
 TABLE OF CONTENTS

CERTIFICATE OF PROOFREADING ..................................................... II

SUPERVISOR CERTIFICATE .................................................................. III

EXAMINING COMMITTEE CERTIFICATION .................................... IV

DEDICATION ................................................................................................ V

ACKNOWLEDGMENTS ............................................................................ VI

ABSTRACT ..................................................................................................VII

TABLE OF CONTENTS .............................................................................. IX

LIST OF FIGURES ................................................................................... XIII

LIST OF TABLES ...................................................................................... XV

LIST OF ABBREVIATIONS.................................................................... XVI

CHAPTER ONE .............................................................................................. 1

1. INTRODUCTION .................................................................................... 1

1.1 Overview .............................................................................................. 2

1.2 Urban Traffic Light System .............................................................. 4

1.3 Problem Statement ............................................................................. 7

1.4 Research Objectives ........................................................................... 7

1.5 Research Contributions ..................................................................... 8

1.6 Research Challenges and Limitations .............................................. 8

1.7 Dissertation Structure ........................................................................ 9

CHAPTER TWO........................................................................................... 11

2. THEORETICAL BACKGROUND AND LITERATURE REVIEW

11

IX
2.1 Introduction....................................................................................... 12

2.2 Theoretical Background .................................................................. 12

2.1.1 Fixed-Time Traffic Signals ........................................................ 12

2.1.2 Actuated Traffic Signals ............................................................ 13

2.1.3 Adaptive Traffic Signals ............................................................ 13

2.1.4 Traffic Control by Law Enforcement ...................................... 14

2.1.5 Smart Traffic Signals ................................................................. 14

2.2 Deep and Transfer Learning Models .............................................. 15

2.2.1 Deep Learning Factors............................................................... 18

2.2.1.a Neural Network Architecture ........................................................... 18

2.2.1. b Activation Functions .................................................................... 18

2.2.1. c Parameter learning .......................................................................... 19

2.2.1. d Optimization algorithms ............................................................... 20

2.2.1. e Data Preprocessing ....................................................................... 22

2.2.1. f Hyperparameter Tuning ................................................................ 23

2.2.2 Deep Learning-Based Classification Techniques .................... 23

2.2.2.a Image Classification......................................................................... 23

I. VGG Networks ....................................................................................... 24

II. ResNet ................................................................................................. 26

III. MobileNet ........................................................................................... 26

IV. DenseNet ............................................................................................. 27

2.2.3 Deep Learning-Based Object Detection Techniques .............. 29

2.2.3.a YOLO............................................................................................... 30

2.3 Simulation Model Components ....................................................... 33

X
 2.3.1 Arduino Board ............................................................................ 33

2.3.2 LED Traffic Light Signal Module ............................................ 35

2.3.3 Breadboard or Prototyping Board ........................................... 36

2.3.4 Jumper Wires ............................................................................. 36

2.3.5 Camera ........................................................................................ 36

2.3 Literature Review ............................................................................. 36

CHAPTER THREE: ..................................................................................... 61

3. METHODOLOGY, RESEARCH DESIGN, MATERIALS AND

METHODS..................................................................................................... 61

3.1 Introduction....................................................................................... 62

3.2 Research Framework ....................................................................... 62

3.2.1 Creating a New Dataset ............................................................. 66

3.2.2 Data Preprocessing ..................................................................... 67

3.2.3 Image Data Annotation.............................................................. 70

3.3 Vehicle Types Classification ............................................................ 71

3.3.1 DenseNet201 ................................................................................ 75

3.3.2 Modified DenseNet201 (Freezing Layers) ................................ 76

3.3.3 Proposed Optimized Selection Algorithm................................ 76

3.3.4 Selecting Optimizers .................................................................. 77

3.4 Vehicle Types Detection ................................................................... 78

3.4.1 Modified YOLOv5s .................................................................... 79

3.4.2 Integrate YOLO with Arduino ................................................. 80

3.4.3 Optimized Traffic Flow ............................................................. 81

3.5 Mathematical Formulation of Optimized Traffic Flow ................ 82

XI
 3.6 Performance Metrics ........................................................................ 83

CHAPTER FOUR: ........................................................................................ 89

4. IMPLEMENTATION, RESULTS AND DISCUSSION .................... 89

4.1 Introduction....................................................................................... 90

4.2 Deep Learning-Based Models’ Results ........................................... 90

4.2.1 Vehicle Types Classification Results ........................................ 91

4.2.2 Vehicle Types Detection Results ............................................... 96

4.2.3 Simulated Environment Results ............................................. 102

4.3 Discussion ........................................................................................ 109

4.3.1 Discussion of Vehicle Types Classification Results ............... 109

4.3.2 Discussion of Vehicle Detection Results ................................. 111

4.3.3 Discussion of Optimized Traffic Flow .................................... 112

CHAPTER FIVE: ....................................................................................... 114

5. CONCLUSIONS AND FUTURE WORKS ....................................... 114

5.1 Conclusions...................................................................................... 115

5.2 Future Works .................................................................................. 116

XII
 LIST OF FIGURES

Figure 1.1 Four lanes intersection ..................................................................... 5

Figure 1.2 Phase Transitions ............................................................................. 6
Figure 1.3 Crossroad Model and Cameras ........................................................ 6
Figure 2.1 Scale driving Deep Learning progress........................................... 16
Figure 2.2 Multi-Layer Neural Networks ...................................................... 17
Figure 2.3 Activation Functions ...................................................................... 19
Figure 2.4 VGG network Architecture ........................................................... 26
Figure 2.5 ResNet Network Architecture ....................................................... 26
Figure 2.6 MobileNet Architecture ................................................................. 27
Figure 2.7 DenseNet201 architecture ............................................................. 29
Figure 2.8 The YOLOv5 network framework ............................................... 31
Figure 3.1 General Framework of Proposed System ...................................... 63
Figure 3.2 Phase 1 Creating the Dataset ......................................................... 63
Figure 3.3 Phase 2 Classification Vehicle Types ............................................ 64
Figure 3.4 Phase 3 Detection of Vehicles ....................................................... 65
Figure 3.5 Phase 4 Design a Traffic Environment .......................................... 65
Figure 3.6 Phase 5 Implement Optimized Traffic Flow ................................. 66
Figure 3.7 Augmented data (Ambulance): (a) Original Image, (b) Vertical
Flip, Sharpen, (c) Horizontal Flip, Sharpen, (d) Sharpen, Random Brightness
Contrast, (e) Random Brightness Contrast, Median Blur ............................... 69
Figure 3.8 Augmented data (Firefighters): (a) Original Image, (b) Vertical
Flip, Sharpen, (c) Horizontal Flip, Sharpen .................................................... 69
Figure 3.9 Augmented data (Police): (a) Original Image, (b) Horizontal Flip,
Sharpen ............................................................................................................ 70
Figure 3.10 Image Labeling ............................................................................ 71
Figure 3.11 Vehicle Classification Process..................................................... 72
Figure 3.12 Transfer knowledge learning-based processes ............................ 74

XIII
Figure 3.13 Steps of DenseNet201 Modification............................................ 75
Figure 3.14 DenseNet201 with CNN for classification .................................. 75
Figure 3.15 Modified DenseNet201 Architecture........................................... 76
Figure 3.16 Proposed Search Mechanism ....................................................... 77
Figure 3.17 Selecting Best Optimizer ............................................................. 78
Figure 3.18 Vehicle Types Detection Processes ............................................. 79
Figure 3.19 Modified YOLOv5 ...................................................................... 80
Figure 3.20 System Architecture..................................................................... 81
Figure 3.21 General Block Diagram of Optimized Traffic Flow ................... 82
Figure 3.22 Intersection Over Union (IOU).................................................... 87
Figure 4.1 Accuracy versus image size (DensNet201-120_freeze_layers) .... 92
Figure 4.2 Precision versus image size (DensNet201-120_freeze_layers)..... 92
Figure 4.3 Confusion matrix for the optimal model and optimizer ................ 95
Figure 4.4 Model’s loss (DensNet201 – 120 layers freezing) ........................ 95
Figure 4.5 Model’s accuracy (DensNet201 – 120 layers freezing) ................ 96
Figure 4.6 Batch Training ............................................................................... 97
Figure 4.7 Batch Prediction............................................................................. 98
Figure 4.8 F1 Score ......................................................................................... 99
Figure 4.9 Precision....................................................................................... 100
Figure 4.10 Recall ......................................................................................... 100
Figure 4.11 [email protected]................................................................................... 101
Figure 4.12 Confusion Matrix ....................................................................... 101
Figure 4.13 The loss functions for the training and validation sets of original
(YOLOv5s).................................................................................................... 102
Figure 4.14 The loss functions for the training and validation sets for
modified (YOLOv5sm) ................................................................................. 102
Figure 4.15 Simulated Environment of Traffic Intersection ......................... 103

XIV
 LIST OF TABLES

Table 2.1 Different Types of Arduino Boards ................................................ 35

Table 2.2 Summary of Different Studies on Traffic Density Estimation ....... 43
Table 2.3 Summary of Various Studies on the Detection and Recognition of
Traffic Signs. ................................................................................................... 48
Table 2.4 Summary of Studies Conducted in the Field of Accident Detection
......................................................................................................................... 53
Table 2.5 Summary of Studies Conducted on Emergency Vehicle Detection 58
Table 3.1Unbalanced Datasets ........................................................................ 67
Table 3.2 Used Preprocessing Techniques ..................................................... 68
Table 3.3 Balanced Datasets ........................................................................... 68
Table 3.4 Results of DL Techniques ............................................................... 73
Table 3.5 Standard tabular confusion matrix .................................................. 86
Table 4.1 Fixed Hyper-parameters for DL Techniques .................................. 91
Table 4.2 Accuracy Tests ................................................................................ 93
Table 4.3 Precision Tests ................................................................................ 94
Table 4.4 Fixed Hyper-parameters for YOLOv5 Techniques ........................ 96
Table 4.5 Object Detection Using YOLOv5 and Modified YOLOv5 Results99
Table 4.6 Lanes and Factors.......................................................................... 104
Table 4.7 Case 1 ............................................................................................ 105
Table 4.8 Case 2 ............................................................................................ 106
Table 4.9 Case 3 ............................................................................................ 108
Table 4.10 Optimized Traffic Flow Results.................................................. 108
Table 4.11 Fixed Hyper-parameters for DL Techniques .............................. 110

XV
 LIST OF ABBREVIATIONS
Abbreviation Meaning
Adagrad Adaptive Gradient Algorithm
Adam Adaptive Moment Estimation
ARM Advanced Reduced Instruction Set Computing
AVR Atmel Alf and Vegard's RISC Processor
BoI Blocks of Interest
BS Background Subtraction
CBL The Convolutional Layer
CCTV Closed-Circuit Television
CNN Convolutional Neural Network
COCO Objects in Context
CSVM Cubic Support Vector Machine
CTSD Chinese Traffic Sign Dataset
CUDA Compute Unified Device Architecture
CVIS Cooperative Vehicle Infrastructure Systems
DenseNet Densely Connected Convolutional Networks
EVs Emergency Vehicles
FPN Feature Pyramid Network
GPU Graphics Processing Unit
GTSDB German Traffic Sign Detection Benchmark
HIS Hue-Saturation-Intensity
HMM Hidden Markov Model
HOG Histogram of Oriented Gradients
ITS Intelligent Transportation Systems
KRG Kurdistan Region of Iraq
LaRA Laboratório de Robótica e Automação
LISA Intelligent & Safe Automobiles
LSS Local Self-Similarity Machine

XVI
mAP Mean Average Precision
MaxGT Maximum Green Time
MFCC-SVM Mel-frequency Spectral Coefficients Combined with
Support Vector Machine
MinGT Minimum Green Time
MobileNet Mobile Networks
MSFF Multi-Scale Feature Fusion
Nadam Nesterov-accelerated Adaptive Moment Estimation
OpenCV Open Source Computer Vision Library
PSO Particle Swarm Optimization
QL Q-learning
R-CNN Region-based Convolution Neural Network
ResNet Residual Network
RF Random Forest
RFID Radio-Frequency Identification
RISC Reduced Instruction Set Computing
RMSprop Root Mean Square Propagation
RNN Recurrent Neural Network
RoI Regions of Interest
SGD Stochastic Gradient Descent
SGW Simplified Gabor Wavelets
SORT Simple Online and Real-time Tracking algorithm
SPP Spatial Pyramid Pooling
SPPF Spatial Pyramid Pooling - Fast
SSD Single-Shot Detection
STL Smart Traffic Light
STS Swedish Traffic Signs
SVM Support Vector Machine
TSC Traffic Signal Control

XVII
TUIC Temporal Unknown Incremental Clustering
VANETs Vehicular Ad-Hoc Networks
VGG Visual Geometry Group
YOLO You Only Look Once

XVIII
CHAPTER ONE

1. INTRODUCTION

1
1.1 Overview

In recent years, there has been a notable increase in the number of cars (Jain
et al., 2019) leading to a widespread problem of traffic congestion (Biswas et
al., 2019) that presents various challenges worldwide. Consequently, there has
been a rise in car accidents and a worrying escalation in yearly CO2 emissions
(Coelho et al., 2005), both of which threaten the sustainability of future
transportation (Guo et al., 2019). Moreover, effective traffic management
heavily relies on the deployment of manpower (Kumaran et al., 2019a). Traffic
control systems, operating on a time-dependent basis, have been designed to
facilitate smooth traffic movement in all directions. However, it is important to
note that the transition of traffic signals from green to red during turns may
sometimes cause traffic congestion in one direction, with minimal
improvement in traffic flow in the opposite direction (Malhi et al., 2011).

Congestions resulting from traffic signals could have adverse effects on the
transportation economy, primarily due to increased fuel consumption (Lakshmi
and Kalpana, 2017) and time expenditure (Jing et al., 2017). Additionally, road
congestion contributes to environmental issues such as noise and air pollution
(Qadri et al., 2020). Moreover, accidents occurring in congested traffic
conditions can lead to injuries or even fatalities (Lakshmi and Kalpana, 2017).
Conversely, reducing congestion offers economic, environmental, and social
benefits. Among various solutions in urban settings, signalized intersections
have been identified as effective means for addressing prevalent bottlenecks,
thus playing a significant role in urban traffic management (Wei et al., 2019).
To this end, the concept of smart cities revolves around Intelligent
Transportation Systems (ITS), which are pivotal in modern urban planning
(Yuan et al., 2019, Zhang et al., 2011). Throughout history, transportation
systems have been vital components of national infrastructure. Research
indicates in 2011, approximately 40% of the global population was spent at

2
least an hour commuting daily (Zhang et al., 2011). Consequently, managing
the increasing number of vehicles has become increasingly challenging without
the assistance of technology (Veres and Moussa, 2019).

Furthermore, emergency vehicles are crucial in critical situations, but

traffic congestion can be a significant threat to patients' lives, causing more
than 20% of fatalities during ambulance transport. In densely populated areas,
traffic jams are frequent during peak hours, slowing down emergency vehicles
such as police cars, ambulances, and firefighters and worsening life-threatening
situations. To address this problem, it is essential to prioritize these vehicles
and introduce an automated traffic system that can recognize and clear their
path (Roy and Rahman, 2019).

In general, in order to make the optimization problem manageable, several

assumptions have to be made. The main problem here that arises is that these
assumptions deviate and sometimes do so significantly from the real world.
Meanwhile, many factors have effects on drivers in real word traffics such as
driver’s preference interactions with vulnerable road users (e.g., pedestrians,
cyclists, etc.), weather and road conditions (Wei et al., 2019). Also, this system
needs to accurately detect emergency vehicles among regular ones. Advanced
techniques like Convolutional Neural Networks (CNNs) based on deep and
transfer learning have shown great potential in computer vision, delivering
results comparable to, and sometimes even better than, human expertise.

Computer vision is crucial in effectively managing and regulating traffic

signals (Kumaran et al., 2019a), contributing to significant success (Jeon et al.,
2018). In general, in busy urban areas, the best way to control traffic flow is by
using intelligent traffic signal systems(Kumaran et al., 2019a). These systems
can estimate density, detect and recognize traffic signals, identify emergency
and police vehicles, and detect accidents. Even though a better infrastructure

3
can improve the traffic flow(Wang et al., 2018), quieter intersections are
usually managed by human controllers or automated systems (Kumaran et al.,
2019a). Cameras are already present in many congested areas for various
purposes, but they can be utilized for analyzing traffic scenes with specific
hardware, instead of replacing existing closed-circuit television (CCTV)
cameras. These cameras are commonly used for security and vehicle detection.
This multi-functional use of cameras optimizes resources and enhances overall
traffic management efficiency (Khan and Ullah, 2019).

Besides on computer vision, deep learning techniques have capability to

distinguish vehicle classes, and counting number of vehicles for different
classes in a traffic video (Asha and Narasimhadhan, 2018). Scientists and
researchers have proposed many approaches for traffic density estimation
which will be done by using images acquired from live cameras positioned near
the traffic junction and several machine learning algorithms (Ikiriwatte et al.,
2019).

1.2 Urban Traffic Light System

Usually, each traffic light contains three color lights: green, yellow and red
lights. These lights are positioned in the four directions—north, south, east, and
west—at intersections (Huang and Chung, 2009b). Figure 1.1 illustrates a
typical intersection formed by two perpendicular and parallel lanes. The traffic
signal comprises three signal lights: red (R), yellow (Y), and green (G), and
additionally features five lights: red (R), yellow (Y), a left turn arrow
illuminated during green (GL), a right turn arrow illuminated during green
(GR), and a straight-ahead arrow illuminated during green (GS) (Huang and
Chung, 2009a).

4
 Figure 1.1 Four lanes intersection (Manguri et al., 2023)
The configuration of traffic lights can lead to different phase transitions
based on the interactions among the number of signal lights (Huang and Chung,
2009a). A state diagram in Figure 1.2 displays the sequence of 4 phase
transitions. The diagram shows a cycle that starts in phase 1 and ends in phase
4, providing an alternative series of operations. During phase 1, the southern
traffic light enables the GL, GR, and GS signals, allowing vehicles to turn left,
right, or go straight in the southbound direction. Similarly, during phase 4, the
GL, GR, and GS signals are activated at the southern traffic light, allowing
vehicles to turn left, right, or go straight in the southbound direction. The
system follows predetermined timing intervals with fixed durations.

5
 Figure 1.2 Phase Transitions (Huang et al., 2005)

Internationally, drivers understand traffic lights. A green signal allows

vehicles in that lane to proceed while all other directions show red, signaling to
stop (Wei et al., 2019). The management of traffic lights and monitoring of
traffic conditions using computer vision technologies require CCTV cameras
at intersections. Figure 1.3 shows a simulation of traffic control at a crossroad
(Manguri, 2016).

Figure 1.3 Crossroad Model and Cameras (Manguri et al., 2023)

6
1.3 Problem Statement

The global population is rapidly increasing and is projected to exceed 8

billion by 2023 and reach 10 billion by 2056 (Chamie, 2020). As a result,
existing transport infrastructures are being strained, leading to worsened traffic
congestion (Ariffin et al., 2021). This congestion has significant consequences,
including high financial costs (Lee and Chiu, 2020) and decreased quality of
life (Ghazali et al., 2019). One critical issue that arises from this situation is the
inadequate response of traditional traffic systems to emergency vehicles, such
as ambulances, fire trucks, and police cars. These vehicles often experience
delays due to congestion, which can hinder their ability to quickly reach
emergency scenes. The challenge lies in developing traffic systems that
effectively prioritize emergency vehicles by coordinating traffic signals to stop
non-priority traffic, ensuring safe and uninterrupted passage. Addressing this
problem is crucial for improving emergency response times and overall public
safety.

1.4 Research Objectives

This study aims at designing and implementing an efficient traffic signaling

system based on custom vehicle detection.

The objectives include:

1. Saving time by reducing waiting time at intersections, especially for

emergency vehicles.
2. Density estimation and counting cars on the road for reducing traffic
congestion.
3. Changing side signal to green signal in case of detection ambulance,
firefighters and police cars.

7
 4. Testing the proposed system and evaluating its performance with other
related works.

1.5 Research Contributions

Researchers mostly use deep learning architecture for vehicle classification

and detection. However, in this study, a transfer based deep learning system
using Densely Connected Convolutional Networks with 201 Layers
(DenseNet201) was used as a main contribution to this work. The significant
key contributions of the proposed model are as follows:

1. Testing varying dimensions of input images (64*64, 128*128, and

224*224) to select the optimal image size for the training phase.
2. Applying data augmentation to make the datasets balanced among
emergency vehicles’ types (police, ambulance, and firefighters).
3. Modifying some layers in DenseNet-201 to improve performance
accuracy.
4. Testing other deep learning models such as (MobileNet, VGG-19,
ResNet-101, etc.) on the same datasets to produce a fair comparison.
5. Choosing the best optimizer that increases the detection accuracy of the
model.
6. Applying You Only Look Once version 5 (YOLOv5) for controlling real
time signaling.
7. Modifying the YOLOv5 to improve the model accuracy.
8. Proposing a mathematical formulation for optimizing traffic signaling.
9. Designing and implementing an integrated system for smart traffic
signaling based on Arduino.

1.6 Research Challenges and Limitations

8
 Collecting data from real traffic area officially requires a permission from
traffic police center. Due to the non-availability of public datasets, a customized
dataset has been created. The main challenge of the intersection traffic signal
control (TSC) problem is to determine an optimal configuration of traffic
signaling system that allows a maximum traffic flow in a network. Creating a
standard dataset for emergency vehicles is not feasible due to the lack of
standardized colors, sizes, and other characteristics. In this regard, an accurate
system for traffic congestion reduction is required to build, which is the main
goal of this study.

Furthermore, deep learning methods such as Region-based Convolution

Neural Network (R-CNN), Fast R-CNN, and Faster R-CNN necessary large
amount of training data and computational power. The limitation of Faster R-
CNN was its inference time. Moreover, SSD (Single-Shot Detection) and
YOLO (You Only Look Once) were developed to solve the Faster R-CNN
limitation. Generally, SSD and YOLOv5 methods have fast detection speed and
high efficiency but a low accuracy.

1.7 Dissertation Structure

The rest of this dissertation is organized as follows:

Chapter Two: Literature Review and Theoretical Background:
In this chapter, a comprehensive review has been done based on recent
researches and describes the theoretical background of all traffic signal phases
(custom dataset preparation, data preprocessing, train DenseNet201 model,
object detection, and system evaluation).
Chapter Three: Methodology, Research Design, Materials and Methods:
The proposed method and used materials are given in chapter three.
Chapter Four: Implementation, Results and Discussion:

9
 This chapter provides an implementation of the developed intelligent traffic
signaling system. Also, it discusses the obtained results.
Chapter Five: Conclusions and Future Works:
The summary of this study can be found on the chapter five, as well as
suggestions for future work.

10
 CHAPTER TWO

2. THEORETICAL BACKGROUND AND LITERATURE

REVIEW

11
2.1 Introduction

This chapter provides the theoretical background of this study such as object

detection and its techniques are discussed. Furthermore, literature review of the

recent studies carried out on computer vision for managing and controlling

traffic signals in this chapter.

2.2 Theoretical Background

Traffic signaling control can be categorized into several types based on the

technology and methods used. This study provides detailed information on

common types of traffic signaling control. However, it is important to note that

some traffic signaling control systems, such as Pedestrian Signals, Traffic Signs

and Road Markings, Traffic control for construction zones, Railway Crossings,

and Traffic Roundabouts, are not directly related to this study.

2.1.1 Fixed-Time Traffic Signals

Fixed-time control involves the use of predetermined green and cycle

durations that remain constant, regardless of variations in traffic volumes

throughout the day. This approach primarily assigns more green time to the

busiest traffic movements based on historical data. Some fixed-time systems

may incorporate distinct preset time intervals for morning rush hours, evening

peak hours, and other periods of high traffic. Nevertheless, this method is not

well-suited for handling unexpected shifts in traffic demand. To address this

limitation, vehicle-actuated traffic signals are implemented. Traditional

12
vehicle-actuated signals essentially respond to the need for green time and

green extension based on vehicle detection, provided it falls within

predetermined limits and as long as traffic flow allows for accurate

measurement. The cycle duration becomes variable, and consequently, adjacent

signalized intersections can only be coordinated in specific scenarios

(Wahlstedt, 2013).

2.1.2 Actuated Traffic Signals

The actuated control method allows for dynamic modification of the green

time duration and total cycle length in response to real-time traffic needs. These

needs are determined using loop detectors or other conventional traffic sensors.

Within this control scheme, critical parameters such as minimum green time

(MinGT), maximum green time (MaxGT), and cycle duration can be adjusted

based on actual traffic conditions, often using data collected by loop detectors

(Shirvani Shiri and Maleki, 2017, Zheng et al., 2010, Zhang and Wang, 2010,

Viti and Van Zuylen, 2010). Furthermore, actuated signal phase timing

schemes are constrained by their inability to accurately adapt to changing traffic

conditions. This means that parameters such as MinGT, MaxGT, and cycle

length are still determined using predefined values. These values rely on a

limited amount of traffic data and recommended experimental values (Nie et

al., 2021).

2.1.3 Adaptive Traffic Signals

The adaptive traffic control system, known as UTOPIA, is designed to

13
optimize traffic flow while granting selective priority to public transport, all

without compromising travel times for private vehicles (Wahlstedt, 2013,

Samadi et al., 2012). UTOPIA represents an innovation in Urban Traffic

Control (UTC) and is characterized by its hierarchical, adaptive, distributed,

and open traffic control system (Taranto, 2012).

Hierarchical: UTOPIA employs goal-related coordination and incorporates

cooperative control.

Adaptive: The system continually monitors real-time traffic conditions and

adjusts signal plans to ensure rapid responses to changes in traffic demand.

Distributed: This approach involves problem decomposition, forward-

looking strategies, strong interaction, and a focus on terminal costs. The

network optimization is broken down into coordinated junction problems,

which are addressed by the intersection units (SPOT) in collaboration with the

central system.

2.1.4 Traffic Control by Law Enforcement

Law enforcement officers are entrusted with the duty of manually regulating

traffic, particularly in situations such as events, emergencies, or construction

zones. They rely on hand signals and stop signs to provide guidance and

exercise control over the flow of vehicles.

2.1.5 Smart Traffic Signals

The definition of a Smart Traffic Light (STL) involves using sensors,

cameras, and artificial intelligence to control the flow of traffic. This novel

14
method depends on the traffic light playing a central role in overseeing the

movement of vehicles and reducing congestion on the roads. To meet the

criteria of being labeled as a smart traffic light, it must incorporate a mechanism

for automatically adjusting signals in real-time.

Various techniques can be employed for developing smart traffic signals,

including the installation of cameras at traffic intersections. These cameras

capture images of the vehicles on the road from a preset distance away from

the traffic signal. Subsequently, image analysis techniques are applied to assess

the number of vehicles present on the road. The information derived from this

analysis is then transmitted to the traffic light, enabling it to adjust signals (red,

yellow, or green) accordingly (Choudekar et al., 2011).

Under the concept of STL, sensors are used on the sides of roads and

vehicles are interconnected. The collected data undergoes analysis, and based

on this information, concert commands are generated. These commands are

then transmitted to the traffic light, facilitating a responsive adjustment of

signals to optimize traffic flow (Hussain et al., 1995).

2.2 Deep and Transfer Learning Models

Deep learning, a subset of AI, is specialized in building substantial neural

network models proficient in making accurate decisions from data. Its strength

lies in situations marked by complex data and access to large datasets (Kelleher,

2019). Therefore, conducting training with a wide range of potential outcomes

15
can reduce the chances of making errors during testing (Dong et al., 2021). But

the current moment, the limitation on neural network complexity was tied to

available computing power. However, progress in big data analytics has

empowered the development of more intricate neural networks. As a result,

computers can now rapidly observe, learn, and respond to complex situations

more swiftly than humans (Kuutti et al., 2020).

Figure 2.1 Scale driving Deep Learning progress (Dong et al., 2021).
Figure 2.1 illustrates how the performance of the Deep Neural Network

improves as the data size increases (Dong et al., 2021).

The main focuses of deep learning techniques in the ITS including Density

estimations, traffic signal control, accident detection, traffic sign detection and

recognition, etc.(Dong et al., 2021).

Deep learning neural networks use both data inputs and weights, which can

be numerical or learned by machines, then multiplied by inputs. They adjust

themselves based on differences between predicted output and training inputs.

16
Additionally, they make use of activation functions, mathematical equations

controlling neuron activity, and bias, in algorithms skews results by either

encouraging or discouraging certain ideas. These elements collectively mimic

the human brain (Kelleher, 2019, Chen et al., 2021), allowing for precise

recognition, classification, and description of data objects. As a result, multiple

layers serve as the backbone of all neural network types, working together to

construct a neural network model (Dong et al., 2021), however, the exact

number of layers can vary.

Input layer: Depicts the dimensions of the input vector.

Hidden layer: Intermediate nodes that partition the input space into distinct

boundaries. An activation function generates an output based on weighted

inputs.

Output layer: Represents the final output of the neural network.

Figure 2.2 Multi-Layer Neural Networks (Ozturk and Fthenakis, 2020)

17
2.2.1 Deep Learning Factors
Deep learning techniques encompass various factors such as:

2.2.1.a Neural Network Architecture

The Neural Network Architecture is defined as arrangement and

composition of layers, nodes, and interconnections within the network. The

Architecture relies on three primary elements that emulate the usual procedures

of feature extraction, matching, simultaneous inlier detection, and model

parameter estimation. These components are designed to be trainable in an end-

to-end manner. To enhance performance, it usually requires a larger network

capacity. A high-performing network architecture often encompasses a wide

range of potential setups concerning the number of layers, hyperparameters

within each layer, and the layer types. This complexity makes a thorough

manual search impractical, relying significantly on expert knowledge and

experience to craft successful networks. As a result, the automated and

intelligent construction of networks remains an ongoing challenge yet to be

resolved (Zhong et al., 2018).

2.2.1. b Activation Functions

Another crucial element within a neural network involves activation

functions, modeled after the firing mechanism of human neurons—signaling

either activation or inactivity. These functions serve to establish nonlinear

connections between input and output, mirroring the structure of the human

brain due to their nonlinearity, multiple nodes, and layers, thus earning the term

18
"neural network." Several activation functions, as illustrated in Figure 2.3,

exist, including commonly used ones like Sigmoid, Hyperbolic tangent, and

Relu. Their primary function lies in transforming and abstracting data into a

format more conducive to classification (Dong et al., 2021).

Figure 2.3 Activation Functions (Dong et al., 2021).

2.2.1. c Parameter learning

As a traditional machine learning classifiers, deep learning classifiers also

require learning parameters utilizing mathematical tools such as gradient

descent. The gradient descent algorithm is particularly valuable for parameter

learning in convex functions, which possess a single absolute minimum or

maximum. Convex functions simplify parameter learning, while nonconvex

functions require mathematical techniques to transform them into convex ones.

Neural network optimization is considered a non-convex optimization problem,

implying the existence of multiple optimum points (minima/maxima). Learning

occurs by minimizing the error between predicted and actual values (Dong et

al., 2021).

19
2.2.1. d Optimization algorithms

Optimization algorithms, crucial in improving neural network performance,

play a significant role in adjusting a model's hyperparameters in a standard

fashion. Hyperparameters, such as the learning rate, direct the update method,

defining the optimizer's behavior. Any two optimizers can be distinguished by

the combination of their hyperparameters and update rule (Fatima and Journal,

2020).

In the Deep Learning models, loss function is provided (could be considered

as the function we aim to optimize for our specific problem). This, coupled with

the optimization algorithm, becomes a mandatory parameter for compiling our

model. The optimizer's function involves adjusting the weights and learning

rate of our model's nodes during training to effectively minimize the loss

function. In essence, the primary objective of an optimizer is to decrease the

training error (Fatima and Journal, 2020). All optimizers use a mathematical

formula to update the weights, along with specific learning rate values.

𝜕𝑒𝑟𝑟𝑜𝑟
𝑤𝑥′ = 𝑤𝑥 − ∝ ( ) 2.1
𝜕𝑤𝑥
Where wx′ denotes the updated weights while wx represents the old weights,

∂error
and ∝ indicates the learning rate. The term ( ) represents the derivative of
∂wx

the error concerning the weights. Different algorithms implement adjustable

learning rates.

I. Root Mean Square Propagation (RMSprop)

20
 RMSprop, despite its widespread use, is an optimizer that was never

officially published. Geoff Hinton, a prominent figure in backpropagation,

introduced it during his online course on Neural Networks for machine

learning. RMSprop and Adadelta emerged concurrently but separately, aiming

to address the diminishing learning rates observed in Adagrad. RMSprop, a

gradient-based optimizer, diverges from considering the learning rate as a fixed

hyperparameter, opting for an adaptive learning rate that is adjusted

dynamically during training.

II. Adaptive Moment Estimation (Adam)

Adam is one of the most popular optimizers and is closely related to

RMSprop and Adagrad. It is known for its high efficiency, adaptability, and

fast convergence. Adam uses the L2 norm or Euclidean norm for optimization.

Like RMSprop, it incorporates squared gradients to adjust the learning rate.

Additionally, it leverages momentum by calculating the moving average of the

gradient, which is similar to Stochastic Gradient Descent with momentum. The

name "Adam" stems from "adaptive moment estimation," signifying its

computation of individual learning rates for diverse parameters. This is

achieved by calculating the first and second moments of gradients and adjusting

the learning rate for each weight in the network.

III. Adamax

Adamax could be described as an Adam optimizer variant that utilizes the

infinity or max norm for optimization. In scenarios where data exhibits

21
traditional noise in gradient updates, such as datasets with multiple outliers,

Adamax tends to outperform Adam.

IV. Nesterov-accelerated Adaptive Moment Estimation (Nadam)

Nadam amalgamates Adam and NAG (Nesterov Accelerated Gradient)

techniques. It incorporates Nesterov to update the gradient a step ahead, hence

named "Nesterov-accelerated Adaptive Moment Estimation." Nadam finds

utility in handling noisy or highly curved gradients. Accelerating the learning

process, it speeds up by aggregating the exponential decay of moving averages

from the previous and current gradients.

In addition, there are other optimization techniques available that were not

detailed in this study due to the good results obtained from tests, such as

Stochastic Gradient Descent (SGD), Adaptive Gradient Algorithm (Adagrad),

and AdaDelta.

2.2.1. e Data Preprocessing

Preparing data involves converting raw information into a format suitable

for training machine learning models, which is a time-consuming process. In

the case of image data, preprocessing typically encompasses basic

transformations like cropping, filtering, rotating, or flipping images. Presently,

data scientists manually determine, based on their expertise, which

transformations and their sequence are best suited for a specific image dataset.

This manual approach not only poses a bottleneck in real-world data science

projects but can also result in suboptimal outcomes. Relying on intuition or

22
trial-and-error methods to explore various image transformations might hinder

the discovery of the most effective ones (Minh et al., 2018).

2.2.1. f Hyperparameter Tuning

Optimizing or tuning hyperparameters is a core technique for improving the

performance of machine learning models. Hyperparameters are parameters that

are adjusted during the learning process to fine-tune and optimize the model's

performance. Various constraints, weights, or learning rates may be required

for the same machine learning model to effectively capture different patterns in

data. These adjustable parameters, known as hyperparameters, are determined

through a process of trial and error to ensure that the model performs at its best

in carrying out the machine learning task (Pon et al., 2021).

2.2.2 Deep Learning-Based Classification Techniques

Image classification refers to sorting an entire image into predefined classes

or categories, like determining if an image depicts a cat or a dog. Conversely,

object detection not only categorizes objects within an image but also pinpoints

their positions by outlining bounding boxes around them. Object detection

offers a more detailed comprehension by identifying and localizing multiple

objects within an image, whereas image classification concentrates on

assigning a single label to the entire image.

2.2.2.a Image Classification

Image classification is an essential part of image processing. Its objective

23
 involves assigning images to predefined categories (Lillesand et al., 2015).

There are two primary types of classification: supervised and unsupervised

classification. Image classification involves a two-step procedure: training and

testing. During training, the framework extracts distinctive attributes from the

images to form a unique representation for each class. This process is repeated

for all classes, depending on the nature of the classification problem—be it

binary or multi-class (Jaswal et al., 2014, He et al., 2016, Liu et al., 2015, He

and Sun, 2015). During the testing phase, the aim is to classify test images into

the classes for which the system was trained. The class assignment is based on

the partitioning between classes, considering the training features (Aggarwal

and vision, 2018).

In recent decades, CNNs have demonstrated significant potential across

various domains within the computer vision community (LeCun et al., 2015).

Also, many techniques are available for image classifications such as

DenseNet, ResNet, Mobile Net. VGG, Inception, AlexNet, and etc. The used

deep learning techniques during this study to standardize the proposed dataset

are detailed below:

I. VGG Networks

The VGG network is a convolutional neural network architecture known for

its depth and simplicity. The VGG network is composed of very small

convolutional filters, comprising 13 convolutional layers and three fully

connected layers (Simonyan and Zisserman, 2014). Its design emphasizes

24
simplicity in constructing the system. The VGG network is primarily a deep

convolutional neural network created with careful consideration of the optimal

layer depth to avoid excessive network complexity (Haque et al., 2019). The

architecture of the VGG network is illustrated in Figure 2.4.

25
 Figure 2.4 VGG network Architecture (Tammina and Publications, 2019)

II. ResNet

Deep residual networks show significant performance in computer vision

and image processing tasks like image classification and segmentation. These

networks resemble a set of filters, utilizing smaller convolutional filters to

simplify the overall network architecture. ResNet is structured with a series of

1x1 and 3x3 convolutional layers, functioning as filters (He et al., 2016). These

layers play a crucial role in reducing network complexity while extracting high-

level features. The network's output relies on a wealth of information used to

activate fully connected layers, effectively preserving extensive information

that pinpoints the precise location of objects (Haque et al., 2019). Figure 2.5

shows the Basic ResNet Architecture.

Figure 2.5 ResNet Network Architecture (Cai et al., 2021)

III. MobileNet

The MobileNet model uses depthwise separable convolutions, which are a

type of factorized convolution. This convolution splits a standard convolution

into two parts: a depthwise convolution and a 1×1 pointwise convolution. In

MobileNets, the depthwise convolution applies a single filter to each input

26
 channel, and then the pointwise convolution combines the outputs from the

depthwise convolution. Unlike a standard convolution that filters and combines

inputs in one step, the depthwise separable convolution separates this process

into two layers: one for filtering and one for combining. This factorization

significantly reduces both computational load and model size (Howard et al.,

2017). The architecture of MobileNet presented in Figure 2.6.

Figure 2.6 MobileNet Architecture (Shobeiri et al., 2021)

IV. DenseNet

The evolution of DenseNet201 proposes the remarkable potential of the

dense architectural framework to achieve cutting-edge outcomes. This

advancement is particularly evident in scenarios where a modest growth rate is

adopted, resulting in the perception of feature maps as a comprehensive

network-wide resource. As a result, each sequential layer enjoys unfettered

27
access to the entirety of feature maps stemming from preceding layers. The

incremental contribution of K feature maps to the global network state

transpires within each layer, defining the cumulative input feature maps at the

1th layer as:

(𝐹𝑀)𝑙 = 𝐾 0 + 𝐾(𝑙 − 1) 2.2

where, 𝐾 0 signifies the input layer's channels. The enhancement of

computational efficiency is portrayed in Figure 2.7, wherein a (1*1)

convolutional layer precedes each (3*3) convolutional layer, thereby curtailing

the volume of input feature maps. Serving as a bottleneck layer, the (1*1)

convolutional layer engenders the production of 4K feature maps. For

classification purposes, the incorporation of two dense layers housing 128 and

64 neurons, respectively, is realized. The modified feature extraction network,

encompassing a truncated DenseNet201 architecture, is augmented by a

sigmoid activation function for binary classification, thereby supplanting the

conventional softmax activation function is employed within the established

DenseNet201 framework (as depicted in Figure 2.7). The sigmoid function's

definition is as follows:

1
𝑦=
1+𝑒 −(∑𝑖 𝑤𝑖 ∗𝑥𝑖 )
2.3

Where y is the output of the neuron, 𝑤𝑖 𝑎𝑛𝑑 𝑥𝑖 represent the weights and

inputs, respectively.

28
 Figure 2.7 DenseNet201 architecture (Jaiswal et al., 2021)

2.2.3 Deep Learning-Based Object Detection Techniques

Object detection is a highly researched subject in the field of computer

vision. Numerous methods have been developed to address the growing

demand for precise models in object detection (Hu et al., 2004). The Viola-

Jones framework (Viola and Jones, 2001) gained popularity primarily for its

successful implementation in solving face-detection challenges (Padilla et al.,

2012). Subsequently, it was adapted for various subtasks such as pedestrian

(Ohn-Bar and Trivedi, 2016) and car (Sun et al., 2006) detections. In recent

29
times, with the rise of CNN (Krizhevsky et al., 2012, Szegedy et al., 2015,

LeCun et al., 1998, He et al., 2016) and GPU-accelerated deep-learning

frameworks, an innovative perspective emerged in the development of object-

detection algorithms (Hinton et al., 2006, Hinton and Salakhutdinov, 2006).

Works like Overfeat (Sermanet et al., 2013), R-CNN (Girshick et al., 2014),

Fast R-CNN (Girshick, 2015), Faster R-CNN (Ren et al., 2015), R-FCN (Dai

et al., 2016), SSD (Liu et al., 2016), and YOLO (Redmon et al., 2016)

significantly elevated the performance benchmarks in this field.

2.2.3.a YOLO

YOLO is an object detection technique known for its efficiency and

accuracy in real-time object detection tasks. YOLO distinguishes itself from

earlier methods used in object detection. Unlike previous practices that

repurpose classifiers for detection tasks, YOLO reframes object detection as a

regression challenge to predict bounding boxes and class probabilities within

an image. This method employs a single neural network to directly generate

these predictions from entire images in a single evaluation. As the entire

detection process operates within a single network, it allows for direct

optimization specifically for detection performance.

The unified architecture is highly efficient in terms of speed. The standard

YOLO model is capable of processing images in real-time at 45 frames per

second. In comparison, a smaller version of this network, called Fast YOLO,

achieves an impressive speed of 155 frames per second while also improving

30
the mean average precision (mAP) compared to other real-time detectors.

Though YOLO may demonstrate more localization errors when compared to

cutting-edge detection systems, it is less prone to making false positive

predictions in the background (Redmon et al., 2016).

The YOLOv5 network is shown in Figure 2.8 and comprises three main

segments: the backbone, neck, and head (output result). The backbone,

positioned immediately after the input image, primarily handles feature

extraction. Following this, the CNN's neck processes data to perform resolution

feature aggregation, and the final predictions based on object resolution are

generated in the head (Khasawneh et al., 2023, Kateb et al., 2021).

Figure 2.8 The YOLOv5 network framework (Guo and Zhang, 2022)

31
 In the backbone part, the input image, initially sized at 640 × 640 × 3, is

transformed by the focus layer for space-to-depth conversion. Using the slice

operator, this process reshapes the image to 320 × 320 × 12, creating a feature

map. This map is then passed through the convolution operator, which utilizes

32 convolution kernels, resulting in a feature map of size 320 × 320 × 32. The

Convolutional Layer (CBL) involves the application of the convolutional

kernel to the input layer, resulting in Conv2D outputs, batch normalization

(BatchNormal), and LeakyRELU activation (Zhou et al., 2021). BottleneckCSP

plays a pivotal role in feature extraction from the map and stands out for its

ability to diminish information gradient duplication during neural network

optimization, constituting a substantial part of the entire YOLOv5 network's

parameters (Wu et al., 2021).

The research explores the variations in width and depth of the

BottleneckCSP segment, which led to the development of four YOLOv5

models: YOLOv5s (small network), YOLOv5m (medium-sized network),

YOLOv5l (large network), and YOLOv5x (very large network) (Glučina et al.,

2023, Zhou et al., 2021) (See Figure Y). The Spatial Pyramid Pooling (SPP)

module enhances the network's receptive field and captures additional features

of diverse scales. YOLOv5 also integrates a bottom-up pyramid structure based

on the feature pyramid network (FPN) (Chen et al., 2022). The FPN layer is

essential for transmitting semantic features and enhancing the robustness of

bottom-up positioning features. It aggregates features from various layers to

32
improve the network's performance and its ability to detect targets at multiple

scales. This is evident towards the end of the image, where the classification

results and object coordinates are shown (Glučina et al., 2023).

2.3 Simulation Model Components

The simulation model requirements for a smart traffic system based on

Arduino and computer vision vary. They include a range of essential

components that are crucial for the system to operate effectively and efficiently

which include Arduino UNO board, a breadboard, LED Traffic Light Module

Board, three 220-ohm resistors, and jumper wires for connectivity.

2.3.1 Arduino Board

The Arduino microcontroller is a platform that is open-source and designed

to be easily programmed and updated at any time. Originally intended for

professionals and students, it enables the creation of devices that can interact

with the environment using sensors.

Arduino microcontrollers feature both inputs and outputs that enable the

acquisition of information, and based on the received data, Arduino can

generate the output signal.

The Arduino platform comprises two main components: hardware and

software. The hardware component involves the use of the Arduino

development board, while the software component involves the Arduino IDE

(Integrated Development Environment) for code development. Arduino utilizes

33
either 8-bit Atmel Alf and Vegard's RISC processor (AVR) microcontrollers or

32-bit Atmel Advanced Reduced Instruction Set Computing (RISC) Machine

(ARM) microcontrollers. These microcontrollers can be easily programmed

using the C or C++ language within the Arduino IDE (Ismailov et al., 2022).

There are several advantages to choosing Arduino microcontrollers over

other options, as highlighted by Massimo Banzi, the Co-founder of Arduino

(Ismailov et al., 2022):

 Active User Community: Arduino has a vibrant user community

where individuals can share their experiences and seek help when

encountering issues. This collaborative environment promotes

problem-solving and knowledge-sharing among users.

 Inexpensive Hardware: Arduino microcontrollers are known for

their affordability, making them a perfect choice for beginners who

want to start projects without spending too much money.

 Growth of Arduino: Another benefit is that the Arduino platform

itself is free to use through the official website, and users only need

to invest in Arduino hardware.

 Multi-platform Environment: The Arduino Integrated Development

Environment (IDE) is compatible across multiple platforms,

including Microsoft, Linux, and Mac OS X, which further expands

the user base and accessibility of the platform.

Various types of Arduino boards are available (As shown in Table 2.2), each

34
differs from others in terms of their microcontroller types, crystal frequencies,

and the presence of auto-reset functionality.

Table 2.1 Different Types of Arduino Boards

Arduino Microcontroller Crystal Auto-Reset Memory

Board Type Frequency Availability Type

Arduino Uno ATmega328P 16 MHz Yes Flash: 32 KB

Arduino ATmega2560 16 MHz Yes Flash: 256 KB

Mega 2560

Arduino Nano ATmega328P 16 MHz Yes Flash: 32 KB

Arduino Due ATSAM3X8E 84 MHz Yes Flash: 512 KB

Arduino Pro ATmega328P 16 MHz No Flash: 32 KB

Mini

Arduino ATmega32U4 16 MHz Yes Flash: 32 KB

Leonardo

For this study, the Arduino Uno has been selected as it meets the

requirements of the traffic signaling model in terms of processing power, size

and form factor, cost, features and connectivity.

2.3.2 LED Traffic Light Signal Module

The LED traffic light signal module is a digital traffic light module that

provides a signal output. This module is known for its high brightness, making

it perfect for creating models of traffic light systems. Additionally, it is small

35
in size, easy to wire, and comes with a targeted design and customizable

installation options.

2.3.3 Breadboard or Prototyping Board

A breadboard or prototyping board is an essential tool for assembling

circuitry, allowing you to connect LEDs, resistors, and microcontroller

elements. It provides a convenient and adaptable platform designed specifically

for testing and prototyping electronic circuits.

2.3.4 Jumper Wires

Jumper wires are used to establish connections between components on a

breadboard or prototyping board. They provide a versatile and easily adaptable

method for interconnecting circuitry.

2.3.5 Camera
Cameras are essential components in smart traffic systems where they are

responsible for managing traffic signals. By continuously capturing images of

roads and lanes, they provide valuable data that allows for precise control of

traffic signals.

2.3 Literature Review

In efforts to improve the control and monitoring of traffic signaling,

scientists and researchers have put forth numerous methodologies drawing

upon machine vision techniques. The architecture of traffic signal control and

monitoring based on computer vision includes various stages, such as image

acquisition, preprocessing, and the application of advanced computer vision

36
techniques like density estimation, traffic sign detection and recognition,

accident detection, emergency vehicle detection, and the like. This section

explores selected papers during recent years that introduce proposed methods

for controlling and monitoring traffic signals.

(Kumaran et al., 2019b) introduced a smart traffic signal control system that

utilizes computer vision algorithms. They also proposed a new method for

traffic signal timing based on traffic flow, using the Temporal Unknown

Incremental Clustering (TUIC) model to cluster moving vehicles according to

optical flow features. Test results show that their approach leads to reduced

waiting times and increased throughput compared to other signal timing

algorithms. (Jing et al., 2017) conducted a thorough review of adaptive traffic

signal control in congested traffic conditions. Their system effectively reduces

urban traffic congestion by minimizing delays. They evaluated performance by

comparing different methods based on prior research findings. Moreover, they

developed a systematic framework for adaptive traffic signal regulation that

uses linked cars to inform future research paths. More research is required to

create more widely applicable adaptive traffic signal control algorithms for a

connected car environment.

In real-time wireless network simulation, (Faraj et al., 2020) proposed an

intelligent microcontroller circuit-based system for controlling cars in traffic

congestion. This system is both efficient and cost-effective, offering an

improvement over traditional traffic signal systems in terms of managing and

37
controlling traffic flow. Notably, it can dynamically adjust the timing of traffic

light signals and rapidly respond to road traffic conditions, particularly during

peak hours, with the aim of reducing congestion. The implementation

incorporates hardware components such as a server, cameras, and a

microcontroller board, along with a wireless network infrastructure connecting

traffic lights. Experimental findings demonstrate enhanced accuracy when

utilizing YOLOv3 as a machine learning tool and OpenCV as a preprocessing

algorithm. Additionally, the system successfully reduces average waiting times

at traffic intersections by over 55%.

An efficient reinforcement learning approach for traffic management

systems introduced by (Joo et al., 2020). Their method tackles the issue of

traffic congestion through an adaptive Traffic Signal Control (TSC) technique,

designed to optimize the number of cars crossing an intersection and distribute

signals evenly across roads using Q-learning (QL). Their research findings

indicate that this proposed method outperforms other studies utilizing QL in

terms of reducing waiting times and minimizing the standard deviation of the

shortest queue lengths.

(Sharma et al., 2021) presented an intelligent framework for traffic light

control systems using deep learning techniques. Their method combines the

YOLO deep learning model to identify objects in a video stream with the

Simple Online and Real-time Tracking algorithm (SORT) to track these objects

across consecutive frames. Their primary focus is on tackling the intricate

38
traffic conditions found on roads in India. The system was tested in real-time

traffic situations, showing positive results. Additionally, they efficiently

managed traffic issues at a low cost by adjusting signal timings dynamically

based on the current road conditions. In cases of heavy traffic congestion in

interconnected road networks, the system successfully reduced the risk of total

jam. To address this challenge, (Sun et al., 2020) recommended a predictive

model using the Hidden Markov Model (HMM) to predict congestion patterns

in highly congested traffic areas. This model establishes a connection between

observed external traffic conditions and hidden internal traffic conditions in

these busy zones. By refining and analyzing floating vehicle trajectory data, the

HMM was adjusted to forecast congestion patterns. Experimental results

indicate a predictive accuracy of up to 83.4%, outperforming moving average

and autoregressive models by 5.8%. These findings highlight the effectiveness

and feasibility of this method in predicting congestion patterns.

(Ke et al., 2018) have demonstrated an innovative method for detecting road

congestion in intelligent transportation systems, using multidimensional visual

characteristics and CNN. The approach starts by determining foreground object

density with a gray-level co-occurrence matrix. Next, the Lucas-Kanade optical

flow, along with a pyramid scheme, is used to evaluate the speed of moving

objects. Subsequently, a Gaussian mixture model is applied for background

modeling, enabling accurate identification of the main foreground amidst

potential foreground candidates through CNN. The efficacy of the proposed

39
technique is confirmed through thorough quantitative and qualitative

evaluations, showing significant improvements over existing road-traffic

congestion detection methods due to the integration of multidimensional

features via CNN. Regarding traffic flow optimization, this section is

subdivided into density estimation, traffic sign detection and recognition,

accident detection, and emergency vehicle detection.

2.3.1 Density Estimation

Estimating traffic density is essential for automating traffic signal control

and reducing congestion at intersections. The following section provides an

overview of the different methods researchers use to estimate traffic density:

(Garg et al., 2016) presented a method for estimating traffic density through

vision, which is a fundamental component in traffic monitoring systems. In

response to the limitations of current techniques, such as inaccuracies in vehicle

counting and tracking, vulnerability to lighting changes, occlusions, and

congestion, the authors developed strategies to address these challenges. They

also tackled the issue of high computational complexity found in holistic

approaches, which hindered real-time implementation. To overcome this

hurdle, they suggested a block processing method for calculating density in

busy road segments. This method involves two main steps: first, identifying

Regions of Interest (ROI), creating Blocks of Interest (BOI), and establishing

background; second, using an iterative process that includes updating the

background, detecting occupied blocks, removing shadow blocks, and

40
estimating traffic density. The effectiveness of the proposed methods was

assessed and verified using the TrafficDB dataset.

According to (Biswas et al., 2017), density estimation is conducted through

car counting, employing the Background Subtraction (BS) method and the

OverFeat framework. The accuracy of the proposed system is assessed through

manual car counting, and a comparative study is carried out before and after

the implementation of the OverFeat framework. The results show a significant

improvement in accuracy, with an average accuracy of 96.55% achieved after

integrating the OverFeat framework, compared to 67.69% for Placemeter and

63.14% for BS, respectively. Additionally, this study demonstrates the

versatility of the OverFeat framework, as it is shown to have other applications

beyond density estimation. Furthermore, the study examines the advantages

and limitations of the BS method, analyzing six individual traffic videos from

various perspectives, including camera angles, weather conditions, and time of

day, in conjunction with the OverFeat framework.

(Biswas et al., 2019), applied SSD and MobileNet-SSD to estimate traffic

density using fifty-nine individual traffic cameras. They evaluated the strengths

and weaknesses of these frameworks by comparing their accuracy with

manually estimated density. The experiments showed that the SSD framework,

in particular, demonstrated significant potential in traffic density estimation,

achieving high detection accuracy rates of 92.97% for SSD and 79.30% for

MobileNet-SSD.

41
 K.-H. N et al (Bui et al., 2020), developed a method that uses advanced

computer vision technologies to analyze traffic flow by extracting data from

video surveillance. The technique involves collecting data from video

surveillance to estimate traffic density at intersections. YOLO and DeepSORT

techniques were utilized to detect, track, and count vehicles, allowing for the

calculation of road traffic density. The effectiveness of the method was tested

using real-world traffic data obtained from CCTV footage gathered over a day.

A new technique for estimating traffic density utilizing a macroscopic

approach has been developed by Kurniawan et al. (2017). The method

proposed comprises two main components: background construction and a

traffic density estimation algorithm. Background construction involves

identifying non-moving vehicles in front of or behind others, while the image

background is determined using edge detection techniques. Density estimation

is achieved by calculating the ratio between the number of Regions of Interest

(ROI) containing objects and the total number of ROI.

(Eamthanakul et al., 2017) introduced an image processing technique for

congestion detection comprising three components: (1) image background

subtraction to distinguish vehicles from the background, (2) application of

morphological techniques to eliminate image noise, and (3) calculation of

traffic density based on the obtained image from CCTV. The outcomes of this

process are subsequently transmitted to the transport plan database.

Finally, Table 2.2 presents a summary of various studies on traffic density

42
estimation.

Table 2.2 Summary of Different Studies on Traffic Density Estimation

Reference(s) Algorithm(s) Dataset(s) Accuracy % Contribution(s)

Traffic density
(Garg et al., The estimation with the
Block Variance 93.70
2016) TrafficDB low computational
cost.
Background Defining ROI by
(Biswas et Subtraction, Over Over Feat
ImageNet 96.55
al., 2017) Feat framework, framework
and Place meter
Data New path opened for
92.97 (SSD),
Detection (SSD) collected real time traffic
(Biswas et 79.30 (
and MobileNet- from cameras density estimation
al., 2019) MobileNet-
SSD with different
SSD)
places
87,88 (Day, Detecting, tracking
Congestion) and counting
vehicles
Collected
(Bui et al., YOLO and 93,88 (Day,
data from
2020) DeepSORT Normal)
CCTV
82,1 (Night,
Normal)
New technique
(Kurniawan ROI and edge developed for
- N/A
et al., 2017) detection estimating traffic
density.
Background Traffic density
(Eamthanak
subtraction and estimated.
ul et al., - N/A
Morphological
2017)
techniques

2.3.2 Traffic Sign Detection and Recognition

Recognition of traffic signs is integral to driver assistance systems and

intelligent autonomous vehicles, contributing significantly to road safety.

Additionally, it can facilitate the automation of traffic signals, thereby helping

to prevent intersections from being crossed when signals indicate red.

Novel approaches proposed in (Kaplan Berkaya et al., 2016) for the

43
detection and recognition of traffic signs. Specifically, a novel technique called

the circle detection algorithm has been developed to identify traffic signs. In

addition, RGB-based color thresholding technique was proposed by Kaplan

Berkaya et al. (2016). Moreover, three algorithms, namely histogram of

oriented gradients, local binary patterns, and Gabor features, were utilized

within a Support Vector Machine (SVM) classification framework to recognize

traffic signs. The effectiveness of these methods for both detection and

recognition was assessed using the German Traffic Sign Detection Benchmark

(GTSDB) dataset. According to the experimental findings, the proposed system

surpassed existing literature and could be employed in real-time operations.

(Yang et al., 2016) presented a method for detecting and recognizing traffic

signs, consisting of three primary steps. Firstly, they use thresholding of HSI

color space components to segment the image. Secondly, they use blobs

extracted from the previous step to detect traffic signs. A significant aspect of

their approach in the initial step is the use of invariant geometric moments for

shape classification instead of machine learning algorithms. Thirdly, they

propose a novel recognition method inspired by existing features. They extend

the Histogram of Oriented Gradients (HOG) features to the HSI color space

and combine them with LSS features to create the descriptor. They test Random

Forest (RF) and SVM classifiers in conjunction with the new descriptor. The

effectiveness of the proposed system is evaluated using the GTSDB and STS

datasets, and its performance is compared with that of existing techniques.

44
 Salti et al. (2015), combined solid image analysis and pattern recognition

techniques to identify traffic signs in mobile mapping data. Unlike other

existing systems that use sliding window detection, their system focuses on

extracting regions of interest. Despite facing challenges like varying

illumination, partial occlusions, and large scale variations, the system showed

strong performance. It targeted the detection of three main categories of traffic

signs - mandatory, prohibitory, and danger signs - following the experimental

setup of the recent GTSDB competition. The system not only excelled in the

online competition but also proved effective when tested on a demanding

dataset of Italian signs in mobile mapping, indicating its potential for successful

real-world deployment.

(Du et al., 2017) developed a robust and efficient classifier-based detector

aimed at achieving fast performance. They introduced two algorithms for

detection and classification. Firstly, they proposed aggregate channel features,

which are based on three types of features: color features, gradient magnitude,

and gradient histograms. Secondly, they presented a boosted trees classifier for

a multi-scale and multi-phase detector, based on the Real AdaBoost algorithm.

Experimental results from this study demonstrated high average-recall and

speed when evaluated on the Daimler, Laboratory for Intelligent & Safe

Automobiles (LISA), and Laboratório de Robótica e Automação (LaRA)

datasets.

Real-time detection and recognition of traffic signs are crucial for improving

45
the intelligence of smart vehicles. To tackle this challenge, (Shao et al., 2018)

proposed a new approach with two main steps. Firstly, images captured from

the road scene are converted into grayscale images. Then, Simplified Gabor

Wavelets (SGW) filters are used to optimize the parameters of the grayscale

images. Additionally, traffic signs are delineated using edge detection to

prepare the data for the following step. In the second stage, the maximally

stable extremal areas approach is used to identify the region of interest, and the

superclass of traffic signs is classified using SVM. CNN is used for subclass

classification of traffic signs, using input from simplified Gabor feature maps

and the same parameters used in the detection phase. Finally, the proposed

method is assessed on the GTSDB and Chinese Traffic Sign Dataset (CTSD)

datasets, with experimental results showing rapid processing speed at 6.3

frames per second and high accuracy of 99.43%.

(Kaplan Berkaya et al., 2016), introduced innovative approaches to improve

traffic efficiency by enhancing object recognition and problem detection using

colorful graphics. They improved two digital image processing methods, the

Circle Detection Algorithm and RGB, which rely on simple image

segmentation techniques to boost traffic sign detection capabilities.

Additionally, they developed a classification framework called the SVM by

integrating three main attributes - Gabor features, histogram of oriented

gradients, and local binary patterns - into the intelligent system. The

effectiveness of their proposed technique was confirmed using the GTSDB

46
datasets. Practical results showed that their technique significantly

outperformed the methods mentioned in the paper and performed consistently

in real-time operations.

A new method for detecting and identifying traffic signs, as suggested in

(Ellahyani et al., 2016) consists of three main steps. Firstly, the image is

segmented based on color space components using Hue-Saturation-Intensity

(HIS) thresholding. Next, blobs identified in the previous step are used to detect

traffic signs. Finally, the recognized traffic signs are classified in the last step.

In their research, two different methods are used for sign classification.

Initially, forms are classified using invariant geometric moments rather than

machine learning methods. New recognition features are then suggested,

drawing inspiration from current ones. The HSI colour space, obtained from

HOG features, is merged with local self-similarity (LSS) features to provide

the descriptor for the new approach. Finally, the suggested method's efficacy is

evaluated and tested against the German Traffic Sign Recognition Benchmark

(GTSRB), German Traffic Sign Detection Benchmark (GTSDB), and Swedish

Traffic Signs (STS) datasets.

The CNN machine learning algorithm is known for its high effectiveness in

object recognition, thanks to its superior recognition rate and efficient

execution time. In a study by (Shustanov and Yakimov, 2017) traffic sign

recognition was implemented using CNN, where different CNN architectures

were compared. The training was conducted using the TensorFlow library, and

47
a massively parallel architecture was adopted for multithreaded programming

with Compute Unified Device Architecture (CUDA). The entire process of

detecting and recognizing traffic signs was carried out in real-time on a mobile

Graphics Processing Unit (GPU). The method's efficiency was then assessed

using the GTSRB dataset, resulting in an impressive classification accuracy of

99.94% for images.

Table 2.3 provides a summary of various studies that focus on different

approaches for detecting and recognizing traffic signs.

Table 2.3 Summary of Various Studies on the Detection and Recognition of

Traffic Signs.

Reference(s) Algorithm(s) Dataset(s) Accuracy % Contribution(s)

Developed Circle
98.24 detection algorithm
(Yang et al., HSI, HOG, LSS GTSDB,
(GTSDB), and an RGB-based
2016) and SVM CTSD
98.77 ( CTSD) color thresholding
technique.
In the first step,
machine learning
algorithms not used
classify shapes instead
(Kaplan HOG, LSS,
of this invariant
Berkaya et Random Forest GTSDB
geometric moments
al., 2016) and SVM 97.04
have been used.
Second, method has
been proposed for the
recognition.
99.43 On-line detecting
(Prohibitory) mandatory,
ROI, HOG, SVM
prohibitory and
(Salti et al., and Context
GTSDB 95.01 danger traffic signs
2015)
(Mandatory)
Aware Filter
97.22 (Danger)
Aggregate Proposed the high
Daimler, 84.314 (
(Du et al., Channel Features average-recall and
LISA and Daimler),
2017) and Boosted speed method
LaRA 90.33 ( LISA),
Trees Classifier

48
 92.048 (
LaRA)
Shapes classified by
GTSRB, using invariant
(Ellahyani et HOG, LSS and
GTSDB and 97.43 geometric moments
al., 2016) SVM
TST

Speed of detection
and classification
(Shao et al., GTSDB and
SGW and SVM 99.43 improved which is
2018) CTSD
more than 6 frames
per second
(Shustanov CNN process
and described
CNN GTSRB 99.94
Yakimov,
2017)
The proposed
(Liu et al., Proposed model CapsNet is employed
TL_Dataset
2021) named CapsNet for traffic sign
recognition.

2.3.3 Accident Detection

One crucial aspect of traffic monitoring beside of others is identifying and

tracking vehicles, which helps in reporting and detecting incidents at traffic

junctions. This section also covers methods for predicting and detecting

accidents.

(Tian et al., 2019) developed a Cooperative Vehicle Infrastructure Systems

(CVIS) and introduced a machine-based vision system capable of automatically

detecting car accidents. The study consisted of two phases: In the first phase,

the CAD-CVIS database was created to improve the accuracy of accident

detection. This database, CAD-CVIS, includes various types of accidents,

weather conditions, and accident locations, representing different traffic

scenarios. In the second phase, a deep neural network model named YOLO-

49
CA, based on CAD-CVIS and deep learning algorithms, was created for

accident detection. Moreover, to enhance the model's performance in detecting

small objects, Multi-Scale Feature Fusion (MSFF) and a loss function with

dynamic weights were utilized. The results showed that the proposed method

surpassed previous approaches, being able to detect car accidents within

milliseconds with a very high average precision of 90.02%. Finally, the

proposed method was compared with existing approaches, demonstrating

improved accuracy and real-time performance compared to other models.

Neoteric framework is presented in (Ijjina et al., 2019), for accident

detection. This framework introduces the use of Mask R-CNN for accurate

object detection and is supported by a centroid-based object tracking algorithm

for surveillance footage efficiency. The main idea is to identify accidents by

spotting irregularities in vehicle speed and trajectory once vehicles intersect.

This framework proves to be superior and paves the way for real-time versatile

vehicular accident detection algorithm development. Performance evaluation

and validation of this framework were conducted using a dataset with diverse

weather conditions.

(Saini et al., 2017), suggested a novel vehicle tracking technique utilizing

image processing, which eliminates the need for background subtraction to

delineate the region of interest. Instead, the study advocates for a hybrid

methodology that integrates feature detection and region matching, facilitating

the estimation of vehicle trajectories across successive frames. As vehicles

50
traverse through an intersection, the system monitors the tracked direction for

any potential events. The research concludes that the proposed method exhibits

proficiency in detecting accidents involving two vehicles.

According to (Wenqi et al., 2017), the TAP-CNN model was introduced for

accident prediction on highways by utilizing convolutional neural networks.

This model combines traffic state and CNN architecture to create a state matrix

that includes various accident factors like traffic flow, weather, and lighting.

The researchers also investigated ways to improve the TAP-CNN model's

accuracy through multiple iterations. They gathered accident data for training

purposes and to assess the model's performance. The experimental results

conclusively show that the TAP-CNN model outperforms conventional neural

network models in accurately predicting traffic accidents.

(Dogru and Subasi, 2018) proposed an intelligent accident detection system

in which automobiles share microscopic vehicle variables. Using vehicle speed

and coordinates obtained from Vehicular Ad-Hoc Networks (VANETs), data

are collected and simulated in the proposed system, which then sends traffic

alerts to drivers. The study also demonstrates the utilization of machine

learning techniques for accident detection on freeways within ITS. Position and

velocity values of each vehicle serve as crucial parameters for easy accident

analysis and detection. Moreover, the proposed method is evaluated using the

OOB dataset, with results indicating that the RF algorithm outperforms ANN

and SVM algorithms, achieving accuracies of 91.56%, 88.71%, and 90.02%,

51
respectively.

As can be seen in (Yu et al., 2019), vision-based algorithms are used to

detect traffic accidents by applying an ST-IHT algorithm to improve the

robustness and sparsity of spatio-temporal features. Furthermore, a weighted

extreme learning machine detector is used to distinguish between traffic

accidents and normal traffic. The study also presents a two-point search

technique designed to dynamically locate candidate values for Lipschitz

coefficients to improve tuning accuracy. To assess the efficiency of the

suggested approach, 30 traffic videos from the YouTube website are used for

testing and evaluation. The results indicate that the suggested technique

surpasses current approaches in terms of traffic accident detection performance.

The accelerometer method is widely utilized for crash detection. In this

respect (Borisagar et al., 2018) state that the accelerometer values undergo

calibration to detect accidents based on acceleration. However, due to the

limitations in accelerometer accuracy and the need for efficient accident

detection, the researchers turned to the CNN machine learning algorithm.

While image classification techniques are typically used for accident detection,

CNNs require significant time, data, and computational resources for training.

To address these challenges, transfer learning techniques were creatively

employed. This involved retraining a pre-trained network, specifically the

Inception-v3 classifier developed by Google for image tasks for accident

detection purposes. The efficiency of the proposed method was then compared

52
to traditional accelerometer-based techniques, resulting in an accuracy of

84.5% for the Transfer Learning algorithm.

Summary of Various Studies Conducted in the Field of Accident Detection

Presented in Table 2.4

Table 2.4 Summary of Studies Conducted in the Field of Accident Detection

Accuracy
Reference(s) Algorithm(s) Dataset(s) Contribution(s)
%
CAD-CVIS dataset
Deep neural
(Tian et al., created and the
network model CAD-CVIS 90.02
2019) proposed method more
YOLO-CA
fast and accurate.
Developing vehicular
(Ijjina et al.,
Mask R-CNN Proposed 71 accident detection
2019)
algorithms in real-time.
Hybrid of feature Accident detection
(Saini et al., Real world
detection and N/A between two vehicles
2017) dataset
Region matching
(Wenqi et al., Accident data Accident predicted by
CNN 78.5
2017) collected using CNN
The proposed method
can provide estimated
ANN, SVM and
(Dogru and 91.56 (RF), geographical
Random
Subasi, OOB data set 88.71
2018) (ANN), location of the possible
Forests (RF)
90.02 accident
(SVM)
(i) Robust Fractures
87.4 ± 0.3 extraction proposed
(SVM), based on OF-DSIFT
ST-IHT, Spatio-
(Yu et al., Collected 94.3 ± 0.2 and ST-IHT
Temporal Features
2019) dataset (ELM),
and W-ELM
95.5 ± 0.3 ii) detect imbalance
(W-ELM) between traffic accident
and normal traffic
video presents a new efficient
(Ghahreman sequences framework for accident
nezhad et al., YOLOv4 collected detection
2022) from
YouTube N/A

53
2.3.4 Emergency Vehicles Detection

The success of law enforcement and public safety is the timely arrival of

first responders at emergency scenes. These responders usually consist of

ambulance, firefighter, and police vehicles. The following section reviews

several suggested techniques for detecting these emergency vehicles.

(Raji et al., 2022) have developed an innovative strategy to manage

emergency situations, especially in congested traffic scenarios. Addressing the

problem of blocked emergency vehicle movement during peak traffic hours,

the system successfully improves timing inefficiencies and reduces traffic

congestion. As an emergency vehicle moves through a specific lane, a Radio-

Frequency Identification (RFID) transmitter captures and transmits signal data.

This allows an RFID receiver to change the traffic signal from red to green,

thus clearing the lane for the emergency vehicle. To adapt to changes in traffic

density, the system dynamically adjusts signal timing intervals to either 10 or

6 seconds. This adjustment leads to decreased congestion, reduced travel time,

and potentially saving lives. By utilizing RFID technology for accurate motion

detection and identification, the proposed system can consistently evaluate

vehicular density and grant automatic priority to emergency vehicles. Both

traditional and deep neural networks are commonly used to classify regular and

emergency vehicles. In this respect, (bin Che Mansor et al., 2021) presented a

classification technique specifically designed to identify emergency vehicles

that frequently get stuck in congested traffic areas. Detecting emergency

54
vehicles on city roads can help improve their prompt arrival. They employed

the VGG-16 model as a pre-trained base, adjusting the convolutional layer and

filter size to boost performance. The experimental results showed that the

proposed method achieved an accuracy rate of 95%. In (Haque et al., 2022),

developed and implemented an automated system to identify emergency

vehicles, distinguishing them from non-emergency vehicles. They used

YOLOv4 for initial object detection with the ROI strategy, then trained the

detected objects using CNN and VGG-16 by fine-tuning the model parameters.

The system reached an average accuracy of 82.03% when tested on the

Emergency Vehicle Identification v1 dataset.

According to (Kaushik et al., 2020), two computer vision techniques are

used to identify and locate emergency vehicles. These methods include object

detection and instance segmentation. More precisely, the process includes

using Faster RCNN for object detection and Mask RCNN for instance

segmentation. The results demonstrate the proposed approach's effectiveness,

particularly its accuracy and suitability for detecting emergency vehicles in

chaotic traffic conditions. Furthermore, a custom dataset of 400 images was

employed for emergency vehicle detection, carefully labeled using the LabeMe

tool.

(Roy and Rahman, 2019) have developed a model aiming at identifying

emergency vehicles like ambulances and fire trucks in crowded road CCTV

footage. This model gives priority to these vehicles, ensuring the emergency

55
lane is cleared to help them pass through traffic intersections smoothly. When

traffic police encounter difficulties in determining which lanes to open for

emergency vehicles, this model provides an automated solution. By employing

deep convolutional neural networks and the Common Objects in Context

(COCO) dataset, the method they propose shows promising results in

effectively detecting and recognizing different types of emergency vehicles.

(Jonnadula and Khilar, 2020b) introduced a hybrid architecture for

emergency vehicle detection, blending image processing and computer vision

elements. They also reduce the search space by using region of interest

techniques.

To reduce casualties in road emergencies, (Lin et al., 2020) introduced a

novel approach that utilizes machine learning techniques. Various features are

extracted through multi-faceted methods to accurately represent ambulance

characteristics. The effectiveness of predicting next-day demand was tested

through experiments involving cutting-edge machine learning techniques and

ambulance demand prediction methodologies. This testing was conducted

using actual ambulance and demographic data from Singapore. Additionally,

the accuracy of the proposed method was validated across different machine

learning techniques and data types, using the SCDF-Engineered-Socio dataset.

The current traffic light system often lacks responsiveness during

emergencies involving ambulances, firefighters, and police vehicles. In

response to this issue, (Suhaimy et al., 2020) have developed an embedded

56
machine learning application. This application involves data acquisition,

feature extraction, exploration of various algorithms, tuning, and model

deployment to achieve optimal performance in a simulation environment.

Specifically, they created a classifier for ambulance siren sounds, sorting them

into 'Ambulance Arrive' and 'No Ambulance Arrive' categories. This classifier

allows the traffic light system to detect and monitor ambulance arrivals during

emergencies. The approach proposed utilizes Mel-frequency Spectral

Coefficients Combined with Support Vector Machine (MFCC-SVM)

implemented on MATLAB R2017b. Additionally, other researchers have

investigated incorporating optimization algorithms into deep learning models

in similar research efforts.

(Alhudhaif et al., 2022) utilized a hybrid approach that combined a pre-

trained CNN GoogleNet and particle swarm optimization (PSO) - an algorithm

inspired by nature - to classify autonomous vehicles. They trained the model

using a Kaggle dataset comprising vehicle images that were enhanced through

various transformations. Subsequently, the model underwent classification

using different classifiers, with the Cubic Support Vector Machine (CSVM)

emerging as the most effective model. The CSVM exhibited superior

performance in terms of both time efficiency and accuracy, achieving an

impressive accuracy rate of 94.8%. Both empirical and statistical evaluations

confirm the model's superiority over similar approaches, not only in accuracy

(94.8%) but also in training duration (82.7 seconds) and speed in forecasting

57
(380 observations per second).

The studies conducted in the field of emergency vehicle detection are

summarized in Table 2.5.

Table 2.5 Summary of Studies Conducted on Emergency Vehicle Detection

Reference(s) Algorithm(s) Dataset(s) Accuracy % Contribution(s)

81 (Object The computational

and accuracy for
Detection), emergency vehicle
(Kaushik et Faster RCNN and Custom
detection are suitable
al., 2020) Mask RCNN dataset
92 (Iinstance
Segmentation
)
(Roy and Deep Detecting and
Rahman, convolutional COCO 97.97 identifying all kinds
2019) neural network emergency cars
Hybrid architecture
presented for
(Jonnadula detection of
and Khilar, YOLO + ResNet COCO N/A emergency vehicles
2020a) in a real time

Varying degrees to
SVR, MLP, SCDF- the model training in
(Lin et al.,
RBFN, and Engineered- N/A LightGBM
2020)
LightGBM Socio

Effectively
(Suhaimy et distinguish audio
MFCC-SVM - 97
al., 2020) events from audio
signals

2.3.5 Transfer Learning and Optimization Techniques

Researchers have conducted some studies in order to improve the accuracy

58
of emergency vehicle detection. Their main focus has been on models that

utilize deep learning techniques. Also, they use of transfer learning for traffic

signal systems and reducing traffic congestion. In this respect, some other

researchers have conducted researches of using optimization algorithms in deep

learning models. (Alhudhaif et al., 2022) proposed a method that combines a

pre-trained CNN GoogleNet with particle swarm optimization (PSO), a nature-

inspired optimization algorithm, to classify autonomous vehicles. The model

was trained using a Kaggle dataset that included vehicle images enhanced with

various transformations. After training, the model was subjected to

classification using different classifiers, with the Cubic support vector machine

(CSVM) proving to be the most effective. It exhibited superior performance in

terms of both time efficiency and accuracy, achieving an accuracy rate of

94.8%. The results of empirical and statistical evaluations clearly demonstrate

that this model not only surpassed similar approaches in terms of accuracy

(94.8%) but also excelled in training duration (82.7 seconds) and speed

forecasting (380 observations per second).

In another work, (Haque et al., 2022) automated an approach which is

deployed to detect emergency vehicles. Ambulances and fire trucks are

categorized as emergencies, while other vehicles are considered non-

emergency. The process begins by identifying multiple vehicles within an

image using the YOLOv4 object detector. These identified vehicles are then

further investigated. Also, the method distinguishes between emergency and

59
non-emergency vehicles. Finally, the research contributes by developing a

model that incorporates a convolutional neural network (CNN) with a viral

algorithm in deep learning. Transfer learning with a fine-tuned VGG16 model

is also utilized for emergency vehicle detection. On the Emergency Vehicle

Identification v1 dataset, this model achieves an average accuracy of 82.03%.

60
 CHAPTER THREE:

3. METHODOLOGY, RESEARCH DESIGN, MATERIALS

AND METHODS

61
3.1 Introduction

This chapter introduces a comprehensive methodology that specifically

addresses vehicle classification and object detection using advanced deep

learning techniques. The primary objective is to enhance accuracy and offer

readers a practical guide for efficient research design and material utilization.

Moreover, this chapter provides detailed explanations of the methods used and

the intricate modifications applied to deep learning techniques. Additionally,

the chapter offers thorough explanations of the employed methods and the

modifications made to deep learning techniques.

3.2 Research Framework

The process of implementing research is divided into five distinct phases,

each with its own objectives and tasks. Visual representations are utilized to

enhance the clarity and flow of each phase within the framework. Table 3.1

presents a block diagram that provides an overview of the entire study

framework.

62
 Figure 3.1 General Framework of Proposed System
Figure 3.2 specifically illustrates Phase 1, where the customized dataset is

created, highlighting the inclusion of various vehicle types such as police cars,

ambulances, firefighters, and non-emergency vehicles.

Figure 3.2 Phase 1 Creating the Dataset

63
 Phase 2 illustrates the development of the classification system. This system

utilizes a modified deep learning technique to accurately distinguish the

different types of vehicles shown in Figure 3.3.

Figure 3.3 Phase 2 Classification Vehicle Types

Figure 3.4 represents Phase 3, which focuses on refining the vehicle

detection system using YOLOv5. This system plays a vital role in the decision-

making process for traffic signals.

64
 Figure 3.4 Phase 3 Detection of Vehicles
Figure 3.5 details Phase 4, where the simulation model is designed. This

figure illustrates the various components and requirements incorporated into

the simulated traffic environment.

Figure 3.5 Phase 4 Design a Traffic Environment

65
 Phase 5 is dedicated to showcasing the implementation of the optimized

traffic signal system. Its primary objective is to improve traffic flow and

prioritize emergency vehicles, as depicted in Figure 3.6.

Figure 3.6 Phase 5 Implement Optimized Traffic Flow

3.2.1 Creating a New Dataset

This research used a dataset that included images of emergency and non-

emergency vehicles. Since there was no existing dataset available specifically

for emergency vehicles like police cars, ambulances, and firefighters, a custom

dataset was created for this study. The vehicle images were gathered from

different sources, including Kaggle (www.kaggle.com), Fatkun Batch, and the

66
Rania traffic directorate in the Kurdistan region of Iraq. It is important to

mention that the vehicle images vary in dimension and have an unbalanced

distribution across different classes. For more information about the dataset for

emergency and non-emergency vehicles, please refer to Table 3.1.

Table 3.1Unbalanced Datasets

Vehicle types Total number of images
Ambulance 322
Firefighters 526
Police car 700
Non-emergency 1670
Total 3218

3.2.2 Data Preprocessing

Preprocessing algorithms involve a range of techniques that aim to improve

the quality of images through specific operations. The initial steps include

resizing the images to meet model requirements, such as dimensions of 64x64,

128x128, and 224x224. After resizing, the images are categorized into

emergency and non-emergency vehicle classes. Often, the quality of the images

is affected by factors like electronic device effects and lighting conditions.

Therefore, implementing preprocessing algorithms is crucial for preparing the

images for the classification process.

This study highlights the importance of image sharpening, smoothing, and

contrast enhancement in enhancing image quality. These enhancements

provide essential support for downstream tasks such as image segmentation,

detection, and classification. For these purposes, Albumentations library

67
python data augmentation is used for balancing the dataset. Albumentations

library has 12 transformations which can be applied to easy and fast

augmentation. Table 3.2 shows a balanced dataset consisting of both

emergency and non-emergency vehicles.

Table 3.2 Used Preprocessing Techniques

Vehicle types Used Techniques

Vertical Flip, Sharpen
Horizontal Flip, Sharpen
Sharpen, Random Brightness
Ambulance Contrast
Random Brightness Contrast,
Median Blur

Vertical Flip, Sharpen

Horizontal Flip, Sharpen
Firefighters Sharpen, Random Brightness
Contrast

Police car Horizontal Flip, Sharpen

Because our dataset is unbalanced, balancing scaling is used for balance

classes and is shown in Table 3.3. The obtained dataset size after the

augmentation process is 6222 images (An example of augmented data for

Ambulance, Firefighters, and Police Car is shown in Figures 3.7, 3.8, and 3.9,

respectively). The results achieved by applying preprocessing techniques to

balance the dataset collected for this study are presented in Table 3.3.

Table 3.3 Balanced Datasets

68
 Total number of
Vehicle types Balancing Scale
Images
Ambulance 5.19844358 1610
Firefighters 3.165876777 1682
Police car 2.385714286 1260
Non-emergency 1 1670
Total 6222

(a) (b) (c)

(d) (e)

Figure 3.7 Augmented data (Ambulance): (a) Original Image, (b) Vertical
Flip, Sharpen, (c) Horizontal Flip, Sharpen, (d) Sharpen, Random Brightness
Contrast, (e) Random Brightness Contrast, Median Blur

(a) (b) (c)

Figure 3.8 Augmented data (Firefighters): (a) Original Image, (b) Vertical
Flip, Sharpen, (c) Horizontal Flip, Sharpen

69
 (a) (b)
Figure 3.9 Augmented data (Police): (a) Original Image, (b) Horizontal Flip,
Sharpen

3.2.3 Image Data Annotation

Annotations play a crucial role in object detection tasks. There are several

free tools available for annotating datasets, such as MakeSense, LabelImg,

CVAT, LabelMe, VoTT, ImgLab, and some more. For this study, the

MakeSense online tool was used to label and annotate the images. The

annotated information is stored in both text and XML files, which can be used

with various object detection techniques like YOLO, R-CNN, SSD, and others.

Figure 3.10 provides an illustrative example of image labeling.

70
 Figure 3.10 Image Labeling

3.3 Vehicle Types Classification

To improve the accuracy of the vehicle classification system, several

classification techniques are applied to the balanced dataset. These techniques

encompass ResNET, MobileNet, VGG16, VGG19, DenseNet201. Transfer

learning was valuable in object classification, particularly when working with

limited datasets, as it provided valuable insights. A visual representation of the

proposed system’s classification can be seen in Figure 3.11. This study

successfully employed a variety of deep learning techniques to precisely

classify vehicles as they pass through a traffic intersection.

71
Figure 3.11 Vehicle Classification Process

72
 Table 3.4 showcases the top results attained, along with the optimizers

utilized for this dataset.

Table 3.4 Results of DL Techniques

Techniques Optimizer Accuracy (%)

VGG16 RMSprop 95
MobileNet RMSprop 95.3
ResNet RMSprop 95.3
DenseNet Adam 97

Based on the results obtained from the tests and the selection of the best

technique for vehicle classification, DenseNet has been chosen for

modification.

Furthermore, fine-tuning the deep transfer learning DenseNet201-based

model can be employed to further improve the results. This process involves

repurposing a pre-trained model, allowing the application of knowledge gained

from a larger dataset to a smaller one, as depicted in Figure 3.12. The

methodologies utilized in this study encompass various processes, as illustrated

in Figure 3.13. The block diagram of DenseNet 201 modification comprises the

following stages:

1. Data Compilation: images are collected from publicly available datasets

and local traffic offices in the KRG, Iraq, to compile the data on vehicles.
2. Image Labeling: Vehicles are categorized into emergency (Ambulance,
Police, and Firefighter) and non-emergency groups through annotation.
This results in four types of vehicles that are used for model training.

73
3. Improvement of Image Quality: Preprocessing algorithms are applied to
enhance the quality of the images. This includes resizing, sharpening,
smoothing, and contrast enhancement.
4. Data Augmentation: Various image transformations are utilized to
address overfitting and balance the datasets.
5. Dataset Partitioning: The data is divided into training and
testing/validation sets for cross-validation. 80% of the data is allocated
for training, while 20% is reserved for testing and validation.
6. Proposed Transfer Model Training: The model is constructed and certain
parameters are adjusted for training.
7. Vehicles Classification: Vehicles are classified into emergency and non-
emergency categories.
8. Evaluation of Performance Metrics: Performance is assessed using loss-
accuracy curves, a confusion matrix, precision, recall, F1-score, and
average accuracy.

Figure 3.12 Transfer knowledge learning-based processes

74
 Figure 3.13 Steps of DenseNet201 Modification
3.3.1 DenseNet201
The proposed model uses the DenseNet201 for classification. The design

and implementation of the proposed approach consists of three steps: apart of

preprocessing, feature extraction, classification, and optimization. Figure 3.14

displays the stages of the proposed model architectures.

Figure 3.14 DenseNet201 with CNN for classification

75
3.3.2 Modified DenseNet201 (Freezing Layers)
In the field of neural networks, freezing layers refers to the control of weight

updates. When a layer is frozen, its weights remain unchanged during further

processing. This simple, yet effective, technique helps reduce computational

costs during training without significantly affecting detection accuracy.

Therefore, freezing layers is a strategic approach to speed up neural network

training by gradually immobilizing hidden layers. In this study, we will apply

multiple freezing layers to the DenseNet201 layers to show their impact on

performance accuracy as depicted in Figure 3.15.

Figure 3.15 Modified DenseNet201 Architecture

3.3.3 Proposed Optimized Selection Algorithm

We train each model using input sizes of 64x64, 128x128, and 224x224.

The models are then simulated with different epoch numbers. Performance

metrics such as accuracy, precision, recall, and F1-score are calculated for each

model. Next, we sort the results to find the optimal values of accuracy and

precision. Our selection process is based on choosing the best optimizer when

accuracy or precision reach their peak, as shown in Figure 3.16.

76
 Figure 3.16 Proposed Search Mechanism

3.3.4 Selecting Optimizers

Many optimization techniques discussed in the existing literature have been

widely used in recent studies on deep transfer learning. These techniques are

designed to reduce the loss function and modify the weights during back-

propagation. Gradient descent is a commonly used method for finding local

minima of different functions. It calculates the gradient by evaluating the loss

function across the entire dataset.

This study investigates how different optimizers affect the accuracy of deep

77
learning DenseNet201 models, both with and without freezing layers. Figure

3.17 depicts the selection of the optimal optimizer.

Figure 3.17 Selecting Best Optimizer

3.4 Vehicle Types Detection

Deep learning techniques can be employed for the purpose of detecting and

categorizing vehicles, with a specific focus on distinguishing emergency

vehicles (EVs) such as police cars, ambulances, and firefighters from non-

emergency vehicles. This involves training a deep learning model to not only

detect vehicles, but also identify them accurately. Figure 3.18 displays a

flowchart depicting the processes used to detect vehicle types.

78
 Figure 3.18 Vehicle Types Detection Processes

3.4.1 Modified YOLOv5s

The changes are made to the YOLOv5s backbone are intended to improve

the network's ability to learn and represent features more efficiently. In this

study, the original C3 layers are replaced with BottleneckCSP layers and

substituted the Spatial Pyramid Pooling - Fast (SPPF) module with the Spatial

Pyramid Pooling (SPP) module in the modified backbone of the YOLOv5s

79
architecture and named to YOLOv5sm. Modified version of YOLOv5 is

illustrated in Figure 3.19.

Figure 3.19 Modified YOLOv5

3.4.2 Integrate YOLO with Arduino

The architecture of the proposed system is divided into two distinct

components hardware and software. In terms of hardware, it includes a webcam

for video input, a PC with a GPU for training and executing the YOLO model,

an LED Traffic Light Signal Module, and an Arduino Uno for signal control.

As for the software component, it involves the use of the Python programming

language and its associated libraries. The working mechanism of the proposed

system for signalling control is illustrated in figure 3.20.

80
 Figure 3.20 System Architecture

3.4.3 Optimized Traffic Flow

The use of deep learning methods allows for the optimization of traffic flow

and the distinction between emergency and non-emergency vehicles. By

utilizing advanced neural network architectures, such as YOLO, it becomes

possible to analyze traffic dynamics and classify vehicles in real-time. This

allows for the accurate identification of emergency vehicles, such as police

cars, ambulances and firefighters, within regular traffic, which in turn enables

the implementation of dynamic traffic management strategies. These strategies

prioritize the passage of emergency vehicles, thereby improving response times

during critical situations. In addition, deep learning techniques can analyze

traffic flow patterns and adjust signal timings at intersections to reduce

congestion and improve overall traffic efficiency. By integrating deep learning

into traffic management systems, smarter and more adaptable control

mechanisms can be implemented, leading to safer and more efficient

transportation networks. A general block diagram of using optimized traffic

81
flow is shown in the Figure 3.21.

Figure 3.21 General Block Diagram of Optimized Traffic Flow

3.5 Mathematical Formulation of Optimized Traffic Flow

This study presents a novel mathematical formulation that aims to optimize

the traffic flow. The proposed approach utilizes advanced algorithms and real-

time traffic data analysis to minimize congestion by adjusting traffic signals in

the presence of emergency vehicles.

Normally, allowed time (ATi) for one lane is equal to:

𝐴𝑇𝑖 = 𝐺𝑖 + 𝑌𝑖 3.1
Where,

82
  ATi : Allowed time for passing vehicles of Lane i (in seconds)

 Gi : Duration of green signal for Lane i (in seconds)

 Yi : Duration of yellow signal for Lane i (in seconds)

That means the total times for one cycle is equal to:

𝐶 = ∑𝑛𝑖=0 𝐴𝑇𝑖 3.2

Where

 C: Total cycle length (in seconds)

The expanded equation of 3.2 is equal to:

𝐶 = ∑𝑛𝑖=0 𝐺𝑖 + 𝑌𝑖 3.3
To incorporate the preemption (Pi) duration for emergency vehicles into the

equation for the total cycle length C, we need to add the preemption duration

for each lane i when an emergency vehicle appears.

+ 𝑃𝑖 𝑖𝑓 𝐸𝑉𝑠 𝑎𝑝𝑝𝑒𝑟𝑎𝑟𝑒𝑑
𝐶 = ∑𝑛𝑖=0 𝐺𝑖 ± 𝑃𝑖 + 𝑌𝑖 { 3.4
− 𝑃𝑖 𝑖𝑓 𝑟𝑜𝑎𝑑 𝐶𝑙𝑒𝑎𝑟𝑑 𝑎𝑛𝑑 𝑛𝑜 𝑣𝑒ℎ𝑐𝑖𝑙𝑒 𝑜𝑛 𝑡ℎ𝑒 𝐿𝑎𝑛𝑒

Then, the static signaling time is converted from static to dynamic value based

on factors including appearance EVs and not cars in the lane while still in green.
𝑛
+ 𝑃𝑖 𝑖𝑓 𝐸𝑉𝑠 𝑎𝑝𝑝𝑒𝑟𝑎𝑟𝑒𝑑
𝐶 = ∑ ∆(𝐺𝑖 ± 𝑃𝑖 + 𝑌𝑖 ) { 3. 5
− 𝑃𝑖 𝑖𝑓 𝑟𝑜𝑎𝑑 𝐶𝑙𝑒𝑎𝑟𝑑 𝑎𝑛𝑑 𝑛𝑜 𝑣𝑒ℎ𝑐𝑖𝑙𝑒 𝑜𝑛 𝑡ℎ𝑒 𝐿𝑎𝑛𝑒
𝑖=0

3.6 Performance Metrics

To evaluate the performance of the models, several metrics are used.

Machine learning tasks are typically categorized into either classification or

object detection. It is crucial to select appropriate metrics to evaluate

83
performance, as not all metrics are suitable for every type of problem. The

metrics used to evaluate the proposed models in this study are divided into two

parts: classification performance and object detection.

The precision-recall curve offers a comprehensive assessment of a system's

performance, often consolidated into a single metric by calculating the average

precision across various standard recall levels or document numbers (Goutte

and Gaussier, 2005).

Precision stands for a model's capability to accurately identify pertinent

objects, denoting the percentage of correct positive predictions (Kamal and

Hamid, 2023).
𝑇𝑃
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 3.6
𝑇𝑃+𝐹𝑃

Sensitivity (Recall) represents a model's ability to detect all relevant

instances, signifying the percentage of accurate positive predictions among all

provided ground truths (Kamal and Hamid, 2023).

𝑇𝑃
𝑅𝑒𝑐𝑎𝑙𝑙 = 3.7
𝑇𝑃+𝐹𝑁

Accuracy is defined as the proportion of correctly predicted observations in

a dataset compared to the total number of observations (Kamal and Hamid,

2023).

𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑜𝑟𝑟𝑒𝑐𝑡 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 3.8
𝑇𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠

Or
𝑇𝑃+𝑇𝑁
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 3.9
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁

84
 F1 Score is a metric that combines precision and recall, providing a

balanced evaluation by considering both metrics (Kamal and Hamid, 2023).

2.𝑇𝑃
𝐹1 𝑠𝑐𝑜𝑟𝑒 = 3.10
2𝑇𝑃+𝐹𝑃+𝐹𝑁

Specificity which refers to the rate of true negatives, is a measure used to

assess the accuracy of correctly identifying negative data (Kamal and Hamid,

2023).
𝑇𝑁
𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 = 3.11
𝑇𝑃+𝐹𝑃

The Confusion Matrix is a structured tabular representation displaying the

results of predictions in binary classification. It offers a comprehensive insight

into a classification model's performance when evaluated against a dataset with

known true values (Padilla et al., 2020, Gong, 2021). This metric helps to assess

and delineate the accuracy and efficiency of classification predictions. Table

3.5 demonstrates a typical confusion matrix for binary classification, yet it can

be expanded to accommodate classification involving more than two classes.

85
 Table 3.5 Standard tabular confusion matrix (Narkhede, 2018)

Actual

Positive (P) Negative (N)

Positive True Positive (TP) False Positive (FP)

Predicted

Negative

False Negative (FN) True Negative (TN)

In the displayed table, columns represent the predicted values, while rows

indicate the actual values. The table distinguishes between two potential

classes: Positive and Negative. For example, in predicting the presence of an

emergency vehicle in an image, a positive prediction column suggests the

image containing a change signal to green, while a negative prediction column

implies the no changing signals. The table is divided into four categories:

 True Positive (TP): Signifies a correct prediction aligned with reality.

 True Negative (TN): Indicates an accurate negative prediction.

 False Positive (FP): Represents an incorrect positive prediction.

 False Negative (FN): Denotes an incorrect negative prediction.

Average Precision (AP), in object detection challenges and across the

scientific community, the AP (Average Precision) stands out as the primary

metric used to assess the accuracy of detections among diverse annotated

datasets. It is important to note that in object detection, a true negative (TN)

86
result is irrelevant. This is due to the fact that there are numerous bounding

boxes in any given image that should not be detected (Padilla et al., 2020).

The provided definitions necessitate defining "correct detection" and

"incorrect detection." A prevalent method to achieve this is by utilizing the

intersection over union (IOU), which is a metric derived from the Jaccard

Index—a measure of similarity between two data sets (Jaccard, 1901).

In object detection, the IOU measures the overlap between the predicted

bounding box Bp and the ground-truth bounding box Bgt, defined as the ratio

of the area of their intersection to the area of their union

𝑎𝑟𝑒𝑎(𝐵𝑝 ∩𝐵𝑔𝑡 )
𝐽(𝐵𝑝 , 𝐵𝑔𝑡 ) = 𝐼𝑂𝑈 = 3.12
𝑎𝑟𝑒𝑎(𝐵𝑝 ∪𝐵𝑔𝑡 )

as illustrated in Figure 3.22.

Figure 3.22 Intersection Over Union (IOU).

To determine whether a detection is correct or incorrect, we compare the

IOU with a given threshold, t. If the IOU is greater than or equal to t, the

detection is considered correct. If the IOU is less than t, the detection is

87
considered incorrect.

As mentioned earlier, object detection frameworks do not utilize true

negatives (TN). Therefore, metrics such as true positive rate (TPR), false

positive rate (FPR), and ROC curves (Hanley and McNeil, 1982) are not used.

Instead, the evaluation of object detection methods primarily revolves around

precision (P) and recall (R) (as previous mentioned). Precision is defined as the

proportion of correctly detected objects, while recall is the proportion of actual

objects that are correctly detected.

Precision measures a model's ability to correctly identify relevant objects,

expressed as the percentage of accurate positive predictions. Recall, on the

other hand, gauges a model's ability to detect all relevant instances, including

all ground-truth bounding boxes, and is calculated as the percentage of correct

positive predictions relative to all ground truths provided.

The mean average precision (mAP) is a metric utilized to assess the

accuracy of object detectors across all classes in a given database. The mAP is

essentially the average precision calculated for each class (Ren et al., 2015, Liu

et al., 2016), that is

1
𝑚𝐴𝑃 = ∑𝑁
𝑖=1 𝐴𝑃𝑖 3.13
𝑁

with APi representing the average precision in the i-th class, and N being

the total number of classes evaluated.

88
 CHAPTER FOUR:

4. IMPLEMENTATION, RESULTS AND DISCUSSION

89
4.1 Introduction

This chapter focuses on using advanced deep learning techniques to achieve

vehicle classification and detection. The main goal is to achieve high accuracy

and provide valuable insights to readers. In addition, simulation environments

of real traffic intersections are developed, which are crucial for the application

of these techniques in real situations. Additionally, this chapter provides an

overview of the architecture of implemented and modified deep learning

techniques. It contains a detailed description of the methods used and changes

made. Finally, the results of this study are presented and discussed.

4.2 Deep Learning-Based Models’ Results

The problem of vehicles passing through a road junction was considered,

and the proposed model, which comprises three main elements, was introduced.

First, a robust deep learning-based classification method was developed. More

precisely, the deep learning model was improved using the efficiency of

DenseNet201, which achieved greater accuracy than other models. The purpose

is to enhance vehicle classification during crossing at a junction. In the second

part, YOLOv5 was employed for vehicle identification and detection tasks.

With this approach, it is possible to accurately and efficiently determine the

vehicle type at any given moment when they pass through the intersection,

achieving complete traffic analysis of the area. Finally, the system features an

innovative algorithm for efficient signal optimization that aims to minimize

90
average travel time by dynamically regulating traffic signals to match observed

congestion levels in order to maximize flow rate and prevent queue build-up.

4.2.1 Vehicle Types Classification Results

As explained in previous sections, one of the main goals of this work is to

find high accuracy DL technique and an optimal optimizer. Accordingly, the

pre-trained models are compiled using the customized dataset for different

epoch numbers (i.e., iterations). Based on the results other hyper-parameters

are fixed for all methods, such as (filter size, dropout rate, batch size, and

learning rate). The description of value of the fixed hyper parameters are shown

in the Table 4.1

Table 4.1 Fixed Hyper-parameters for DL Techniques

Hyper Parameter Description

Filter Size 7*7
Dropout Rate 0.5
Batch Size 16
Learning Rate 0.0001

The tests conducted involved evaluating performance metrics such as

average accuracy, precision, recall, and F1-Score. These metrics were

measured over a span of 15 to 30 epochs, a range chosen based on observations

from preliminary experiments. When the model was trained for fewer than 15

epochs, its performance consistently suffered. Conversely, extending training

beyond 30 epochs did not lead to noticeable improvement in performance

91
metrics. Figure 4.1. and 4.2. show the accuracy and precision results for the

different input image sizes when the best optimizer has been chosen. In this

case, a modified DenseNet201 with 120 layers freezing is used.

Accuracy (%)
1.1

0.9

0.8

0.7

0.6

0.5
15 20 25 30

Accuracy Image Size 64 Accuracy Image Size 128 Accuracy Image Size 224

Figure 4.1 Accuracy versus image size (DensNet201-120_freeze_layers)

Precision (%)
1.1

0.9

0.8

0.7

0.6

0.5
15 20 25 30

Precision Image Size 64 Precision Image Size 128 Precision Image Size 224

Figure 4.2 Precision versus image size (DensNet201-120_freeze_layers)

Experimental results in Figures 4 and 5 reveal that the image size 224*224

gives better accuracy and precision values. Therefore, for further tests, merely

92
used this size with all techniques and optimizers in this research.

Table 4.2 and 4.3 present the accuracy and precision metrics results to select

the best optimizer using our new search strategy and other tests are detailed in

the Appendix A.

Table 4.2 Accuracy Tests

Image Size Epoch Accuracy
Models Optimizers
No. (%)
15 VGG19 Adam 90.37
20 VGG19 Adam 90.04
64
25 VGG16 RMSprop 90.78
30
VGG19 Nadam 90.69
15 DenseNet201 Freeze 0 Nadam 96.37
20 DenseNet201 Freeze 0 RMSprop 96.79
DenseNet201 Freeze 0 -
128 25 Adam 96.21
30
DenseNet201 Freeze 0 -
30 Adam 96.21
90
DenseNet201 Freeze 0 -
15 Adam 98.06
30
DenseNet201 Freeze 0 -
20 RMSprop 98.27
60
224
DenseNet201 Freeze 0 -
25 RMSprop 98.68
150
DenseNet201 Freeze 0 -
30 RMSprop 98.84
120

The highlighted rows in the above tables indicate that the more accurate

model is the DenseNet201 with 120 layers freezing and the best optimizer is

RMSprop for both accuracy and precision tests when the input image size is

224*224 and the epoch No. is 30.

93
 Table 4.3 Precision Tests
Image Epoch Accuracy
Models Optimizers
Size No. (%)
15 VGG16 Adam 90.96
20 VGG19 Adam 91.01
64
25 VGG16 RMSprop 91.63
30 VGG19 Nadam 91.65

15 DenseNet201 Freeze 0 Nadam 96.37

20 DenseNet201 Freeze 0 RMSprop 96.84
DenseNet201 Freeze 0 Adam 96.21
128 25
- 30
30 DenseNet201 Freeze 0 Adam 96.26
- 90

DenseNet201 Freeze 0 Adam 98.61

15
- 30
DenseNet201 Freeze 0 RMSprop 98.27
20
- 60
224
DenseNet201 Freeze 0 RMSprop 98.71
25
- 150
DenseNet201 Freeze RMSprop 98.85
30
0 - 120

In addition, the confusion matrix of the proposed emergency vehicles

classification transfer-based model is plotted, which is displayed in Figure 4.3.

The matrix diagonal represents the proposed model performance accuracy for

different types of emergency cars.

94
 Figure 4.3 Confusion matrix for the optimal model and optimizer

The model’s (DenseNet201 with 120 layers freezing) loss and accuracy for

the training and test validation are depicted in Figure 4.4 and Figure 4.5

respectively.

Figure 4.4 Model’s loss (DensNet201 – 120 layers freezing)

95
 Figure 4.5 Model’s accuracy (DensNet201 – 120 layers freezing)

4.2.2 Vehicle Types Detection Results

Another objective of this study is to apply the YOLO deep learning

technique for detecting vehicle types. The hyperparameters, such as filter size,

dropout rate, batch size, and learning rate, are kept constant across all methods

using the YOLO model, except image size. The Table 4.4 displays the

Description and Value of the Fixed Hyperparameters.

Table 4.4 Fixed Hyper-parameters for YOLOv5 Techniques

Hyper-parameter Description
Batch Size 16
Learning Rate 0.01
Optimizer SGD

Figures 4.6 and 4.7 show examples for batch training and prediction,

respectively.

96
Figure 4.6 Batch Training

97
 Figure 4.7 Batch Prediction
The results are obtained from evaluating different input image sizes using

performance metrics such as mAP, recall, and F1-Score are presented in Table

4.5. The highlighted rows in the tables indicate the superior performance of the

YOLOv5sm model.

98
Table 4.5 Object Detection Using YOLOv5 and Modified YOLOv5 Results
Model Image No. of F1 Score Precisio Recall [email protected]

Size Epochs (%) n (%) (%) (%)

YOLOv5s 640 50 86 100 99 91

75 87 100 99 90.2

100 88 100 98 92.2

512 100 89 100 99 91.7

YOLOv5sm 640 100 0.87 100 98 90.9

(Proposed) 416 100 89 100 99 93.2

The tests conducted evaluate performance metrics, including average F1-

Score, precision, recall, mAP, and confusion matrix are shown in Figures 4.9,

4.10, 4.11, and 4.13 Respectively.

Figure 4.8 F1 Score

99
Figure 4.9 Precision

Figure 4.10 Recall

100
 Figure 4.11 [email protected]

Figure 4.12 Confusion Matrix

Figures 4.14 and 4.15 show the loss functions for the training and validation

sets. These figures depict both the unmodified (YOLOv5s) and modified

(YOLOv5sm) versions, utilizing the same hyperparameters.

101
 Figure 4.13 The loss functions for the training and validation sets of original
(YOLOv5s)

Figure 4.14 The loss functions for the training and validation sets for

modified (YOLOv5sm)

4.2.3 Simulated Environment Results

Due to difficulties implementing smart traffic system on the roads and

obtaining permission from the relevant security offices, a simulated version of

102
a real traffic environment was developed for this study. As previously

mentioned, the equipment in the simulated environment includes a webcam for

acquiring images from simulated roads, a PC for processing the captured

images and uploading them to the model, and an Arduino with LEDs for

controlling traffic signals as shown in the Figure 4.16.

Figure 4.15 Simulated Environment of Traffic Intersection

A semi-truth table has been used to control and optimize traffic flow. This

table represents the values of A (Ambulance), F (Firefighter), P (Police Car),

and C (Crowded Lane) in relation to Lanes (L1, L2, L3, and L4), with the goal

of determining availability.

The factors, which can affect the traffic signals with respecting lanes have

been coded and read by Arduino as shown in the Table 4.5.

103
 Table 4.6 Lanes and Factors

Traffic Lanes

L1 L2 L3 L4

A 1 5 a A
Factors

F 2 6 b B

P 3 7 c C

C 4 8 d D

For testing the proposed traffic flow some scenario created which are detailed

below:

 Case 1: Assume that:

o Set the duration of one cycle = 140 Seconds (120 for Green

Signals with 30 Seconds for each Lane and 20 Seconds for Red

with 5 Seconds for each Lane).

o The green signal turned on for Lane 3, indicating that it can

continue and try to finish within its allocated time.

o An ambulance appeared in Lane 2, which is currently under the

red signal as shown in Table 4.7.

o The proposed method involves changing the signal of L3 to

104
 yellow for 5 seconds, which will enable the ambulance to pass

through the traffic area by opening L2.

o After the yellow signal has finished, the green signal returns to

L3 and traffic cycles continue until another factor appears.

Table 4.7 Case 1

Traffic Lanes

L1 L2 L3 L4

A X ✓ X X
Factors

F X X X X

P X X X X

C X X X X

Where:

✓ represents that the factor occurs on the level.

X represents that the factor does not occur on the level.

 Case 2: Assume that:

o Set the duration of one cycle = 140 Seconds (120 for Green

Signals with 30 Seconds for each Lane and 20 Seconds for Red

with 5 Seconds for each Lane).

105
 o The green signal turned on for Lane 3, indicating that it can

continue and try to finish within its allocated time.

o A Firefighter appeared in Lane 1, which is currently under the

red signal as shown in Table 4.8.

o The proposed method involves changing the signal of L3 to

yellow for 5 seconds, which will enable the Firefighter to pass

through the traffic area by opening L1.

o After the yellow signal has finished, the green signal returns to

L3 and traffic cycles continue until another factor appears.

Table 4.8 Case 2

Traffic Lanes

L1 L2 L3 L4

A X X X X
Factors

F ✓ X X X

P X X X X

C X X X X

Where:

✓ represents that the factor occurs on the level.

X represents that the factor does not occur on the level.

106
 Case 3: Assume that:

o Set the duration of one cycle = 140 Seconds (120 for Green

Signals with 30 Seconds for each Lane and 20 Seconds for Red

with 5 Seconds for each Lane).

o The green signal turned on for Lane 4, indicating that it can

continue and try to finish within its allocated time.

o A Police Car appeared in Lane 1, which is currently under the

red signal as shown in Table 4.9.

o The proposed method involves changing the signal of L4 to

yellow for 5 seconds, which will enable the Firefighter to pass

through the traffic area by opening L1.

o After the yellow signal has finished, the green signal returns to

L4 and traffic cycles continue until another factor appears.

107
 Table 4.9 Case 3

Traffic Lanes

L1 L2 L3 L4

A X X X X
Factors

F X X X X

P ✓ X X X

C X X X X

Where:

✓ represents that the factor occurs on the level.

X represents that the factor does not occur on the level.

The results for cases are shown in the Table 4.10.

Table 4.10 Optimized Traffic Flow Results

Case Normal Time to Optimized Time Reduced Time

Pass (in Seconds) (in Seconds) (in Seconds)

1 105 5 100

2 70 5 65

3 35 5 30

108
4.3 Discussion

The analysis of the obtained results focuses on three key areas: classification

processes, vehicle detection processes, and traffic flow optimization. Each

process demonstrates significant improvements in accuracy due to the

implementation of smart traffic signaling.

4.3.1 Discussion of Vehicle Types Classification Results

The performance of the DenseNet201 model can be greatly affected by the

number of frozen layers during training and the quality of the training data. This

impact is evident when comparing two configurations of DenseNet201, as

shown in Table 4.2, which achieved higher accuracy compared to others. The

results highlight the notable effect of augmented data on increasing accuracy.

In the first configuration, the DenseNet201 model was trained by freezing

the first 30 layers using an imbalanced dataset of 3218 images. Despite the

imbalance, the model achieved an impressive accuracy of 96.7%.

In contrast, the second configuration which involved training the model by

freezing the first 120 layers and using a balanced dataset of 6222 images. This

setup resulted in a significantly higher accuracy of 98.84%, showcasing the

combined benefits of freezing more layers and using balanced data. A balanced

dataset ensures equal representation of each class, which generally improves

model generalization. It is worth noting that the same RMSProb optimizer was

used in both scenarios. The fixed Hyper-parameters for DL techniques are

109
shown in Table 4.11.

Table 4.11 Fixed Hyper-parameters for DL Techniques

Model Optimizer Accuracy (%)

DenseNet201 with RMSProb Imbalanced

Freezing 0-30 data 96.7
Layers (3218 images)
DenseNet201 with RMSProb Balanced data
Freezing 0-120 (6222 images) 98.84
Layers

The significant improvement in accuracy, from 96.7% to 98.84%, can be

attributed to two key factors. First, freezing a larger number of layers,

specifically 120 instead of just 30, allows the model to retain more of its pre-

trained features. This aids in better feature extraction and reduces the risk of

overfitting. Second, using a balanced dataset with data augmentation

techniques likely contributed to the improvement. Data augmentation enhances

the diversity and representativeness of the training data, allowing the model to

learn more generalized features through varied training examples.

In summary, the results clearly demonstrate the crucial roles played by layer

freezing and the quality of the training data in the performance of the

DenseNet201 model. Freezing more layers (0-120) and using a balanced

dataset significantly increase accuracy, as evidenced by the rise to 98.84%. This

highlights the effectiveness of these strategies in improving model performance

and emphasizes the importance of data augmentation and appropriate model

adjustments in achieving high accuracy in machine learning tasks.

110
4.3.2 Discussion of Vehicle Detection Results
The performance of the YOLOv5s model and its proposed modified

YOLOv5sm, has been analyzed across different image sizes and training

epochs using key metrics such as F1 Score, Precision, Recall, and [email protected].

The YOLOv5s model, tested with an image size of 640*640, shows a

gradual improvement in F1 Score from 86% at 50 epochs to 88% at 100 epochs.

Precision remains consistently perfect at 100% across all epochs, indicating no

false positives. Recall slightly decreases from 99% to 98% as epochs increase,

while [email protected] improves from 91% at 50 epochs to 92.2% at 100 epochs. This

technique demonstrates that longer training leads to better overall performance

accuracy.

When the image size is reduced to 512*512 and the model is trained for 100

epochs, the F1 Score reaches 89%, Precision remains at 100%, Recall at 99%,

and [email protected] at 91.7%. This configuration suggests that a smaller image size

with sufficient training can yield high performance, but it does not surpass the

performance metrics of the proposed model.

The proposed YOLOv5sm model, tested with an image size of 640* 640 for

100 epochs, achieves an F1 Score of 87%, Precision of 100%, Recall of 98%,

and [email protected] of 90.9%. However, with an image size of 416*416 at 100

epochs, it significantly outperforms the other configurations with an F1 Score

of 89%, Precision of 100%, Recall of 99%, and [email protected] of 93.2%. This

demonstrates that the proposed model's modifications are particularly effective

111
at this smaller image size, achieving the highest [email protected] among all tested

configurations.

In summary, both models exhibit high precision across all configurations,

but the proposed YOLOv5sm with an image size of 416*at 100 epochs delivers

the best balance of metrics, particularly excelling in [email protected]. This indicates

its superior capability in object detection tasks, making it a preferred choice for

applications requiring high accuracy and reliability. Further validation on

diverse datasets and real-world conditions would ensure the robustness of these

findings.

4.3.3 Discussion of Optimized Traffic Flow

The analysis presents three scenarios that compare the time it takes to

complete a process before and after optimization. These scenarios provide

insights into the efficiency gains achieved.

In the first scenario, the initial time to complete the process is 105 seconds.

After optimization, this time is drastically reduced to 5 seconds, resulting in a

reduction of 100 seconds. This indicates an approximately 95.24%

improvement in efficiency, showcasing the significant impact of the

optimization technique.

The second scenario shows a similar trend. Initially, the process takes 70

seconds. After optimization, the time drops to 5 seconds, achieving a 65-second

reduction, which translates to a 92.86% increase in efficiency. This further

underscores the effectiveness of the optimization method employed.

112
 In the third scenario, the process originally takes 35 seconds to complete.

With optimization, the time is reduced to 5 seconds, resulting in a 30-second

decrease and an 85.71% improvement in efficiency. This again highlights the

substantial benefits of the optimization process.

The obtained results clearly illustrate the significant impact of optimization

on reducing process times across all scenarios. The optimized time consistently

drops to 5 seconds in each case, suggesting a standardized optimization

approach.

The value of improvement is noteworthy, with time reductions ranging from

85.71% to 95.24%. This consistency in achieving a 5-second completion time

indicates the effectiveness of the optimization process and its potential for

broad application.

In summary, the analysis demonstrates the profound effect of optimization

on reducing process time with significant efficiency gains observed in each

scenario.

113
 CHAPTER FIVE:
5. CONCLUSIONS AND FUTURE WORKS

114
5.1 Conclusions

In recent years, the issue of traffic congestion has become increasingly

critical due to the growing number of vehicles on the roads. This study presents

its findings in three main phases:

The first phase focused on classifying vehicle types using various deep

learning techniques. Among the models tested, the DenseNet201 model proved

to be the most effective, outperforming the others. By implementing

enhancements such as freezing the first 120 layers and using a balanced dataset

with data augmentation, the accuracy increased significantly from 96.7% to

98.84%. This improvement highlights the importance of balanced data and

model optimization for achieving high accuracy in machine learning tasks.

The second phase involved vehicle detection using the YOLOv5s and

YOLOv5sm models. The YOLOv5s model demonstrated high precision and

recall across different image sizes and training epochs. Specifically, the

YOLOv5sm model, trained with an image size of 416x416 for 100 epochs,

achieved impressive results: An F1 score of 89%, precision of 100%, recall of

99%, and [email protected] of 93.2%. These findings highlight the effectiveness of

the YOLOv5sm model, particularly when using smaller image sizes, and

recommend its use for accurate and reliable vehicle detection.

The final phase involved testing the proposed system in a simulated traffic

environment with the goal of optimizing traffic flow. Three scenarios were

evaluated based on the presented mathematical formulation, and each scenario

115
showed significant reductions in process times. In the first scenario, the time

was reduced from 105 seconds to 5 seconds, representing a 95.24%

improvement. The second scenario achieved a reduction from 70 seconds to 5

seconds, resulting in a 92.86% increase in efficiency. Lastly, the third scenario

reduced the time from 35 seconds to 5 seconds, leading to an 85.71%

improvement. These findings emphasize the substantial impact of the proposed

system in reducing process times, suggesting that its standardized application

could effectively optimize traffic flow in real-world settings.

5.2 Future Works

In order to improve the intelligent traffic systems, the following future

works must be taken into consideration:

 Enhanced Dataset Development: It is important to continuously

update and expand datasets to include a wider range of diverse and

challenging scenarios. This will help to improve the robustness of

trained models.

 Applying another DL technique: It is necessary to apply and explore

more deep learning architectures and optimization techniques. This

exploration will may enhance the accuracy and efficiency of traffic

signaling systems.

 Real-World Implementation: Implement the proposed system in the

real traffic environment.

116
 Finding the EVs car positions: Applying another technique for

improving the traffic signaling based on the EVs positions. In this

case, the lane with EV must be preempted depending on the number

of cars until the traffic signal turn from yellow to green.

117
REFERENCES
AGGARWAL, V. J. A. T. O. I. P. & VISION, C. 2018. A review: deep learning
technique for image classification. 4, 21.
ALHUDHAIF, A., SAEED, A., IMRAN, T., KAMRAN, M., ALGHAMDI, A.
S., ASEERI, A. O. & ALSUBAI, S. J. C. S. S. E. 2022. A Particle Swarm
Optimization Based Deep Learning Model for Vehicle Classification.
40, 223-235.
ARIFFIN, W. N. S. F. W., KEAT, C. S., PRASATH, T., SURIYAN, L., NORE,
N. A. M., HASHIM, N. B. M. & ZAIN, A. S. M. Real-time Dynamic
Traffic Light ControlSystem with Emergency Vehicle Priority. Journal
of Physics: Conference Series, 2021. IOP Publishing, 012063.
ASHA, C. & NARASIMHADHAN, A. Vehicle counting for traffic
management system using YOLO and correlation filter. 2018 IEEE
International Conference on Electronics, Computing and
Communication Technologies (CONECCT), 2018. IEEE, 1-6.
BIN CHE MANSOR, M. A. H., KAMAL, N. A. M., BIN BAHAROM, M. H.
& BIN ZAINOL, M. A. Emergency Vehicle Type Classification using
Convolutional Neural Network. 2021 IEEE International Conference on
Automatic Control & Intelligent Systems (I2CACIS), 2021. IEEE, 126-
129.
BISWAS, D., SU, H., WANG, C., BLANKENSHIP, J. & STEVANOVIC, A.
J. S. 2017. An automatic car counting system using OverFeat framework.
17, 1535.
BISWAS, D., SU, H., WANG, C., STEVANOVIC, A. & WANG, W. 2019.
An automatic traffic density estimation using Single Shot Detection
(SSD) and MobileNet-SSD. Physics and Chemistry of the Earth, Parts
A/B/C, 110, 176-184.
BORISAGAR, P., AGRAWAL, Y. & PAREKH, R. Efficient Vehicle Accident
Detection System using Tensorflow and Transfer Learning. 2018

118
 International Conference on Networking, Embedded and Wireless
Systems (ICNEWS), 27-28 Dec. 2018 2018. 1-6.
BUI, K.-H. N., YI, H., JUNG, H. & CHO, J. Video-Based Traffic Flow
Analysis for Turning Volume Estimation at Signalized Intersections. In:
NGUYEN, N. T., JEARANAITANAKIJ, K., SELAMAT, A.,
TRAWIŃSKI, B. & CHITTAYASOTHORN, S., eds. Intelligent
Information and Database Systems, 2020// 2020 Cham. Springer
International Publishing, 152-162.
CAI, L., WANG, Z., KULATHINAL, R., KUMAR, S., JI, S. J. I. T. O. N. N.
& SYSTEMS, L. 2021. Deep low-shot learning for biological image
classification and visualization from limited training samples. 34, 2528-
2538.
CHAMIE, J. J. Y. G. O. 2020. World population: 2020 overview.
CHEN, L., LEI, C. J. D. L. & MINDSPORE, P. W. 2021. Deep learning basics.
17-28.
CHEN, Z., WU, R., LIN, Y., LI, C., CHEN, S., YUAN, Z., CHEN, S. & ZOU,
X. J. A. 2022. Plant disease recognition model based on improved
YOLOv5. 12, 365.
CHOUDEKAR, P., BANERJEE, S. & MUJU, M. Implementation of image
processing in real time traffic light control. 2011 3rd International
Conference on Electronics Computer Technology, 2011. IEEE, 94-98.
COELHO, M. C., FARIAS, T. L., ROUPHAIL, N. M. J. T. R. P. D. T. &
ENVIRONMENT 2005. Impact of speed control traffic signals on
pollutant emissions. 10, 323-340.
DAI, J., LI, Y., HE, K. & SUN, J. J. A. P. A. 2016. R-FCN: object detection
via regionbased fully convolutional networks, CoRR abs/1605.06409.
DOGRU, N. & SUBASI, A. Traffic accident detection using random forest
classifier. 2018 15th Learning and Technology Conference (L&T), 25-
26 Feb. 2018 2018. 40-45.

119
DONG, S., WANG, P. & ABBAS, K. J. C. S. R. 2021. A survey on deep
learning and its applications. 40, 100379.
DU, X., LI, Y., GUO, Y. & XIONG, H. Vision-Based Traffic Light Detection
for Intelligent Vehicles. 2017 4th International Conference on
Information Science and Control Engineering (ICISCE), 21-23 July
2017 2017. 1323-1326.
EAMTHANAKUL, B., KETCHAM, M. & CHUMUANG, N. The traffic
congestion investigating system by image processing from cctv camera.
2017 International Conference on Digital Arts, Media and Technology
(ICDAMT), 2017. IEEE, 240-245.
ELLAHYANI, A., ANSARI, M. E. & JAAFARI, I. E. 2016. Traffic sign
detection and recognition based on random forests. Applied Soft
Computing, 46, 805-815.
FARAJ, M. A., BOSKANY, N. W. J. U. J. O. S. & TECHNOLOGY 2020.
Intelligent Traffic Congestion Control System using Machine Learning
and Wireless Network. 4, 123-131.
FATIMA, N. J. A. A. I. D. C. & JOURNAL, A. I. 2020. Enhancing
performance of a deep neural network: A comparative analysis of
optimization algorithms. 9, 79-90.
GARG, K., LAM, S.-K., SRIKANTHAN, T. & AGARWAL, V. Real-time
road traffic density estimation using block variance. 2016 IEEE Winter
Conference on Applications of Computer Vision (WACV), 2016. IEEE,
1-9.
GHAHREMANNEZHAD, H., SHI, H. & LIU, C. Real-Time Accident
Detection in Traffic Surveillance Using Deep Learning. 2022 IEEE
International Conference on Imaging Systems and Techniques (IST), 21-
23 June 2022 2022. 1-6.

120
GHAZALI, W., ZULKIFLI, C. & PONRAHONO, Z. The effect of traffic
congestion on quality of community life. 4th International Conference
on Rebuilding Place, 2019.
GIRSHICK, R. 2015. Fast r-cnn. IEEE International Conference on Computer
Vision. IEEE.
GIRSHICK, R., DONAHUE, J., DARRELL, T. & MALIK, J. Rich feature
hierarchies for accurate object detection and semantic segmentation.
Proceedings of the IEEE conference on computer vision and pattern
recognition, 2014. 580-587.
GLUČINA, M., ANĐELIĆ, N., LORENCIN, I. & CAR, Z. J. E. 2023.
Detection and Classification of Printed Circuit Boards Using YOLO
Algorithm. 12, 667.
GONG, M. J. I. J. O. M. I. T. V. 2021. A novel performance measure for
machine learning classification. 13.
GOUTTE, C. & GAUSSIER, E. A probabilistic interpretation of precision,
recall and F-score, with implication for evaluation. European conference
on information retrieval, 2005. Springer, 345-359.
GUO, G. & ZHANG, Z. 2022. Road damage detection algorithm for improved
YOLOv5. Scientific Reports, 12, 15523.
GUO, Q., LI, L. & BAN, X. 2019. Urban traffic signal control with connected
and automated vehicles: A survey. Transportation Research Part C:
Emerging Technologies, 101, 313-334.
HANLEY, J. A. & MCNEIL, B. J. J. R. 1982. The meaning and use of the area
under a receiver operating characteristic (ROC) curve. 143, 29-36.
HAQUE, M. F., LIM, H.-Y. & KANG, D.-S. Object detection based on VGG
with ResNet network. 2019 International Conference on Electronics,
Information, and Communication (ICEIC), 2019. IEEE, 1-3.
HAQUE, S., SHARMIN, S. & DEB, K. Emergency vehicle detection using
deep convolutional neural network. Proceedings of International Joint

121
 Conference on Advances in Computational Intelligence: IJCACI 2021,
2022. Springer, 535-547.
HE, K. & SUN, J. Convolutional neural networks at constrained time cost.
Proceedings of the IEEE conference on computer vision and pattern
recognition, 2015. 5353-5360.
HE, K., ZHANG, X., REN, S. & SUN, J. Deep residual learning for image
recognition. Proceedings of the IEEE conference on computer vision
and pattern recognition, 2016. 770-778.
HINTON, G. E., OSINDERO, S. & TEH, Y.-W. J. N. C. 2006. A fast learning
algorithm for deep belief nets. 18, 1527-1554.
HINTON, G. E. & SALAKHUTDINOV, R. R. J. S. 2006. Reducing the
dimensionality of data with neural networks. 313, 504-507.
HOWARD, A. G., ZHU, M., CHEN, B., KALENICHENKO, D., WANG, W.,
WEYAND, T., ANDREETTO, M. & ADAM, H. J. A. P. A. 2017.
Mobilenets: Efficient convolutional neural networks for mobile vision
applications.
HU, W., TAN, T., WANG, L., MAYBANK, S. J. I. T. O. S., MAN, &
CYBERNETICS, P. C. 2004. A survey on visual surveillance of object
motion and behaviors. 34, 334-352.
HUANG, Y.-S., CHUNG, T.-H. & CHEN, C.-T. Modeling traffic signal
control systems using timed colour Petri nets. 2005 IEEE International
Conference on Systems, Man and Cybernetics, 2005. IEEE, 1759-1764.
HUANG, Y. S. & CHUNG, T. H. 2009a. Modeling and analysis of urban traffic
light control systems. Journal of the Chinese Institute of Engineers, 32,
85-95.
HUANG, Y. S. & CHUNG, T. H. J. J. O. T. C. I. O. E. 2009b. Modeling and
analysis of urban traffic light control systems. 32, 85-95.

122
HUSSAIN, T. M., BAIG, A. M., SAADAWI, T. N. & AHMED, S. A. J. I. T.
O. V. T. 1995. Infrared pyroelectric sensor for detection of vehicular
traffic using digital signal processing techniques. 44, 683-689.
IJJINA, E. P., CHAND, D., GUPTA, S. & GOUTHAM, K. Computer Vision-
based Accident Detection in Traffic Surveillance. 2019 10th
International Conference on Computing, Communication and
Networking Technologies (ICCCNT), 6-8 July 2019 2019. 1-6.
IKIRIWATTE, A., PERERA, D., SAMARAKOON, S., DISSANAYAKE, D.
& RUPASIGNHE, P. Traffic density estimation and traffic control using
convolutional neural network. 2019 International Conference on
Advancements in Computing (ICAC), 2019. IEEE, 323-328.
ISMAILOV, A. S., JO‘RAYEV, Z. B. J. S. & EDUCATION 2022. Study of
arduino microcontroller board. 3, 172-179.
JACCARD, P. J. B. S. V. S. N. 1901. Étude comparative de la distribution
florale dans une portion des Alpes et des Jura. 37, 547-579.
JAIN, N. K., SAINI, R., MITTAL, P. J. S. C. T. & SOCTA, A. P. O. 2019. A
review on traffic monitoring system techniques. 569-577.
JAISWAL, A., GIANCHANDANI, N., SINGH, D., KUMAR, V., KAUR, M.
J. J. O. B. S. & DYNAMICS 2021. Classification of the COVID-19
infected patients using DenseNet201 based deep transfer learning. 39,
5682-5689.
JASWAL, D., VISHVANATHAN, S., KP, S. J. I. J. O. S. & RESEARCH, E.
2014. Image classification using convolutional neural networks. 5, 1661-
1668.
JEON, H., LEE, J. & SOHN, K. J. J. O. I. T. S. 2018. Artificial intelligence for
traffic signal control based solely on video images. 22, 433-445.
JING, P., HUANG, H. & CHEN, L. J. I. 2017. An adaptive traffic signal control
in a connected vehicle environment: A systematic review. 8, 101.

123
JONNADULA, E. P. & KHILAR, P. M. A New Hybrid Architecture for Real-
Time Detection of Emergency Vehicles. In: NAIN, N., VIPPARTHI, S.
K. & RAMAN, B., eds. Computer Vision and Image Processing, 2020//
2020a Singapore. Springer Singapore, 413-422.
JONNADULA, E. P. & KHILAR, P. M. A New Hybrid Architecture for Real-
Time Detection of Emergency Vehicles. Computer Vision and Image
Processing: 4th International Conference, CVIP 2019, Jaipur, India,
September 27–29, 2019, Revised Selected Papers, Part II 4, 2020b.
Springer, 413-422.
JOO, H., AHMED, S. H. & LIM, Y. J. C. C. 2020. Traffic signal control for
smart cities using reinforcement learning. 154, 324-330.
KAMAL, K. & HAMID, E.-Z. 2023. A comparison between the VGG16,
VGG19 and ResNet50 architecture frameworks for classification of
normal and CLAHE processed medical images.
KAPLAN BERKAYA, S., GUNDUZ, H., OZSEN, O., AKINLAR, C. &
GUNAL, S. 2016. On circular traffic sign detection and recognition.
Expert Systems with Applications, 48, 67-75.
KATEB, F. A., MONOWAR, M. M., HAMID, M. A., OHI, A. Q. & MRIDHA,
M. F. J. A. 2021. FruitDet: Attentive feature aggregation for real-time
fruit detection in orchards. 11, 2440.
KAUSHIK, S., RAMAN, A. & RAO, K. V. S. R. Leveraging Computer Vision
for Emergency Vehicle Detection-Implementation and Analysis. 2020
11th International Conference on Computing, Communication and
Networking Technologies (ICCCNT), 1-3 July 2020 2020. 1-6.
KE, X., SHI, L., GUO, W. & CHEN, D. J. I. T. O. I. T. S. 2018. Multi-
dimensional traffic congestion detection based on fusion of visual
features and convolutional neural network. 20, 2157-2170.
KELLEHER, J. D. 2019. Deep learning, MIT press.

124
KHAN, S. D. & ULLAH, H. 2019. A survey of advances in vision-based
vehicle re-identification. Computer Vision and Image Understanding,
182, 50-63.
KHASAWNEH, N., FRAIWAN, M. & FRAIWAN, L. J. C. C. 2023. Detection
of K-complexes in EEG signals using deep transfer learning and
YOLOv3. 26, 3985-3995.
KRIZHEVSKY, A., SUTSKEVER, I. & HINTON, G. E. J. A. I. N. I. P. S.
2012. Imagenet classification with deep convolutional neural networks.
25.
KUMARAN, S. K., MOHAPATRA, S., DOGRA, D. P., ROY, P. P. & KIM,
B.-G. 2019a. Computer vision-guided intelligent traffic signaling for
isolated intersections. Expert Systems with Applications, 134, 267-278.
KUMARAN, S. K., MOHAPATRA, S., DOGRA, D. P., ROY, P. P. & KIM,
B.-G. J. E. S. W. A. 2019b. Computer vision-guided intelligent traffic
signaling for isolated intersections. 134, 267-278.
KURNIAWAN, F., SAJATI, H., DINARYANTO, O. J. I. J. O. E. &
TECHNOLOGY 2017. Image processing technique for traffic density
estimation. 9, 1496-1503.
KUUTTI, S., BOWDEN, R., JIN, Y., BARBER, P. & FALLAH, S. J. I. T. O.
I. T. S. 2020. A survey of deep learning applications to autonomous
vehicle control. 22, 712-733.
LAKSHMI, C. J. & KALPANA, S. Intelligent traffic signaling system. 2017
International Conference on Inventive Communication and
Computational Technologies (ICICCT), 10-11 March 2017 2017. 247-
251.
LECUN, Y., BENGIO, Y. & HINTON, G. J. N. 2015. Deep learning. 521, 436-
444.

125
LECUN, Y., BOTTOU, L., BENGIO, Y. & HAFFNER, P. J. P. O. T. I. 1998.
Gradient-based learning applied to document recognition. 86, 2278-
2324.
LEE, W.-H. & CHIU, C.-Y. J. S. 2020. Design and implementation of a smart
traffic signal control system for smart city applications. 20, 508.
LILLESAND, T., KIEFER, R. W. & CHIPMAN, J. 2015. Remote sensing and
image interpretation, John Wiley & Sons.
LIN, A. X., HO, A. F. W., CHEONG, K. H., LI, Z., CAI, W., CHEE, M. L.,
NG, Y. Y., XIAO, X., ONG, M. E. H. J. I. J. O. E. R. & HEALTH, P.
2020. Leveraging Machine Learning Techniques and Engineering of
Multi-Nature Features for National Daily Regional Ambulance Demand
Prediction. 17, 4179.
LIU, P.-H., SU, S.-F., CHEN, M.-C. & HSIAO, C.-C. Deep learning and its
application to general image classification. 2015 international
conference on informative and cybernetics for computational social
systems (ICCSS), 2015. IEEE, 7-10.
LIU, W., ANGUELOV, D., ERHAN, D., SZEGEDY, C., REED, S., FU, C.-Y.
& BERG, A. C. Ssd: Single shot multibox detector. Computer Vision–
ECCV 2016: 14th European Conference, Amsterdam, The Netherlands,
October 11–14, 2016, Proceedings, Part I 14, 2016. Springer, 21-37.
LIU, X., YAN, W. Q. J. M. T. & APPLICATIONS 2021. Traffic-light sign
recognition using Capsule network. 80, 15161-15171.
MALHI, M. H., ASLAM, M. H., SAEED, F., JAVED, O. & FRAZ, M. Vision
based intelligent traffic management system. 2011 Frontiers of
Information Technology, 2011. IEEE, 137-141.
MANGURI, K. H. K. 2016. Traffıc Sıgnalıng Control At Hıghway
Intersectıons Usıng Morphologıcal Image Processıng Technıque. Hasan
Kalyoncu Üniversitesi.

126
MANGURI, K. H. K., MOHAMMED, A. A. J. U. J. O. S. & TECHNOLOGY
2023. A Review of Computer Vision–Based Traffic Controlling and
Monitoring. 7, 6-15.
MINH, T. N., SINN, M., LAM, H. T. & WISTUBA, M. J. A. P. A. 2018.
Automated image data preprocessing with deep reinforcement learning.
NARKHEDE, S. 2018. Understanding Confusion Matrix [Online]. Available:
https://towardsdatascience.com/understanding-confusion-matrix-
a9ad42dcfd62 [Accessed July 22 2024].
NIE, C., WEI, H., SHI, J. & ZHANG, M. J. T. R. I. P. 2021. Optimizing
actuated traffic signal control using license plate recognition data:
Methods for modeling and algorithm development. 9, 100319.
OHN-BAR, E. & TRIVEDI, M. M. To boost or not to boost? on the limits of
boosted trees for object detection. 2016 23rd international conference
on pattern recognition (ICPR), 2016. IEEE, 3350-3355.
OZTURK, S. & FTHENAKIS, V. J. E. 2020. Predicting frequency, time-to-
repair and costs of wind turbine failures. 13, 1149.
PADILLA, R., COSTA FILHO, C., COSTA, M. J. W. A. O. S.,
ENGINEERING & TECHNOLOGY 2012. Evaluation of haar cascade
classifiers designed for face detection. 64, 362-365.
PADILLA, R., NETTO, S. L. & DA SILVA, E. A. A survey on performance
metrics for object-detection algorithms. 2020 international conference
on systems, signals and image processing (IWSSIP), 2020. IEEE, 237-
242.
PON, M. Z. A., KK, K. P. J. S. T. O. A. I. & COMPUTING, Q. 2021.
Hyperparameter tuning of deep learning models in keras. 1, 36-40.
QADRI, S. S. S. M., GÖKÇE, M. A. & ÖNER, E. 2020. State-of-art review of
traffic signal control methods: challenges and opportunities. European
Transport Research Review, 12, 55.

127
RAJI, C., VP, F. F. & KT, S. S. Emergency Vehicles Detection during Traffic
Congestion. 2022 6th International Conference on Trends in Electronics
and Informatics (ICOEI), 2022. IEEE, 32-37.
REDMON, J., DIVVALA, S., GIRSHICK, R. & FARHADI, A. You only look
once: Unified, real-time object detection. Proceedings of the IEEE
conference on computer vision and pattern recognition, 2016. 779-788.
REN, S., HE, K., GIRSHICK, R. & SUN, J. J. A. I. N. I. P. S. 2015. Faster r-
cnn: Towards real-time object detection with region proposal networks.
28.
ROY, S. & RAHMAN, M. S. Emergency vehicle detection on heavy traffic
road from cctv footage using deep convolutional neural network. 2019
international conference on electrical, computer and communication
engineering (ECCE), 2019. IEEE, 1-6.
SAINI, A., SUREGAONKAR, S., GUPTA, N., KARAR, V. & PODDAR, S.
Region and feature matching based vehicle tracking for accident
detection. 2017 Tenth International Conference on Contemporary
Computing (IC3), 10-12 Aug. 2017 2017. 1-6.
SALTI, S., PETRELLI, A., TOMBARI, F., FIORAIO, N. & DI STEFANO, L.
2015. Traffic sign detection via interest region extraction. Pattern
Recognition, 48, 1039-1049.
SAMADI, S., RAD, A. P., KAZEMI, F. M. & JAFARIAN, H. J. J. O. T. T.
2012. Performance evaluation of intelligent adaptive traffic control
systems: A case study. 2, 248.
SERMANET, P., EIGEN, D., ZHANG, X., MATHIEU, M., FERGUS, R. &
LECUN, Y. J. A. P. A. 2013. Overfeat: Integrated recognition,
localization and detection using convolutional networks.
SHAO, F., WANG, X., MENG, F., RUI, T., WANG, D. & TANG, J. J. S. 2018.
Real-time traffic sign detection and recognition method based on
simplified Gabor wavelets and CNNs. 18, 3192.

128
SHARMA, M., BANSAL, A., KASHYAP, V., GOYAL, P. & SHEIKH, T. H.
Intelligent traffic light control system based on traffic environment using
deep learning. IOP Conference Series: Materials Science and
Engineering, 2021. IOP Publishing, 012122.
SHIRVANI SHIRI, M. & MALEKI, H. R. J. I. J. O. F. S. 2017. Maximum
green time settings for traffic-actuated signal control at isolated
intersections using fuzzy logic. 19, 247-256.
SHOBEIRI, S., AAJAMI, M. J. A. J. O. S. & ENGINEERING 2021. Shapley
value in convolutional neural networks (CNNs): A Comparative Study.
2, 9-14.
SHUSTANOV, A. & YAKIMOV, P. 2017. CNN Design for Real-Time Traffic
Sign Recognition. Procedia Engineering, 201, 718-725.
SIMONYAN, K. & ZISSERMAN, A. J. A. P. A. 2014. Very deep
convolutional networks for large-scale image recognition.
SUHAIMY, M. A., HALIM, I. S. A., HASSAN, S. L. M. & SAPARON, A.
Classification of ambulance siren sound with MFCC-SVM. AIP
Conference Proceedings, 2020. AIP Publishing LLC, 020032.
SUN, T., HUANG, Z., ZHU, H., HUANG, Y. & ZHENG, P. J. I. A. 2020.
Congestion pattern prediction for a busy traffic zone based on the Hidden
Markov Model. 9, 2390-2400.
SUN, Z., BEBIS, G., MILLER, R. J. I. T. O. P. A. & INTELLIGENCE, M.
2006. On-road vehicle detection: A review. 28, 694-711.
SZEGEDY, C., LIU, W., JIA, Y., SERMANET, P., REED, S., ANGUELOV,
D., ERHAN, D., VANHOUCKE, V. & RABINOVICH, A. Going
deeper with convolutions. Proceedings of the IEEE conference on
computer vision and pattern recognition, 2015. 1-9.
TAMMINA, S. J. I. J. O. S. & PUBLICATIONS, R. 2019. Transfer learning
using vgg-16 with deep convolutional neural network for classifying
images. 9, 143-150.

129
TARANTO, D. J. P., FACULTY OF TECHNICAL SCIENCES, BITOLA,
MACEDONIA 2012. UTOPIA Urban Traffic Control Overview.
TIAN, D., ZHANG, C., DUAN, X. & WANG, X. J. I. A. 2019. An automatic
car accident detection method based on cooperative vehicle
infrastructure systems. 7, 127453-127463.
VERES, M. & MOUSSA, M. J. I. T. O. I. T. S. 2019. Deep learning for
intelligent transportation systems: A survey of emerging trends. 21,
3152-3168.
VIOLA, P. & JONES, M. Rapid object detection using a boosted cascade of
simple features. Proceedings of the 2001 IEEE computer society
conference on computer vision and pattern recognition. CVPR 2001,
2001. Ieee, I-I.
VITI, F. & VAN ZUYLEN, H. J. J. T. R. P. C. E. T. 2010. A probabilistic
model for traffic at actuated control signals. 18, 299-310.
WAHLSTEDT, J. Evaluation of the two self-optimising traffic signal systems
Utopia/Spot and ImFlow, and comparison with existing signal control in
Stockholm, Sweden. 16th International IEEE Conference on Intelligent
Transportation Systems (ITSC 2013), 2013. IEEE, 1541-1546.
WANG, Y., YANG, X., LIANG, H. & LIU, Y. J. J. O. A. T. 2018. A review
of the self-adaptive traffic signal control system based on future traffic
environment. 2018.
WEI, H., ZHENG, G., GAYAH, V. & LI, Z. J. A. P. A. 2019. A survey on
traffic signal control methods.
WENQI, L., DONGYU, L. & MENGHUA, Y. A model of traffic accident
prediction based on convolutional neural network. 2017 2nd IEEE
International Conference on Intelligent Transportation Engineering
(ICITE), 1-3 Sept. 2017 2017. 198-202.

130
WU, T.-H., WANG, T.-W. & LIU, Y.-Q. Real-time vehicle and distance
detection based on improved yolo v5 network. 2021 3rd World
Symposium on Artificial Intelligence (WSAI), 2021. IEEE, 24-28.
YANG, Y., LUO, H., XU, H. & WU, F. 2016. Towards Real-Time Traffic Sign
Detection and Classification. IEEE Transactions on Intelligent
Transportation Systems, 17, 2022-2031.
YU, Y., XU, M. & GU, J. J. I. I. T. S. 2019. Vision-based traffic accident
detection using sparse spatio-temporal features and weighted extreme
learning machine. 13, 1417-1428.
YUAN, Y., XIONG, Z. & WANG, Q. ACM: Adaptive cross-modal graph
convolutional neural networks for RGB-D scene recognition.
Proceedings of the AAAI conference on artificial intelligence, 2019.
9176-9184.
ZHANG, G. & WANG, Y. J. I. T. O. I. T. S. 2010. Optimizing minimum and
maximum green time settings for traffic actuated control at isolated
intersections. 12, 164-173.
ZHANG, J., WANG, F.-Y., WANG, K., LIN, W.-H., XU, X. & CHEN, C. J. I.
T. O. I. T. S. 2011. Data-driven intelligent transportation systems: A
survey. 12, 1624-1639.
ZHENG, X., RECKER, W. & CHU, L. J. J. O. I. T. S. 2010. Optimization of
control parameters for adaptive traffic-actuated signal control. 14, 95-
108.
ZHONG, Z., YAN, J., WU, W., SHAO, J. & LIU, C.-L. Practical block-wise
neural network architecture generation. Proceedings of the IEEE
conference on computer vision and pattern recognition, 2018. 2423-
2432.
ZHOU, F., ZHAO, H. & NIE, Z. Safety helmet detection based on YOLOv5.
2021 IEEE International conference on power electronics, computer
applications (ICPECA), 2021. IEEE, 6-11.

131
APPENDIX A: RESULTS OF CLASSIFICATION PROCESS
Image Size: 64*64
No of epochs: 15

Model Optimizer accuracy precision recall f1-score

Adam 0.713580247 0.7194 0.7136 0.7146
Adamax 0.57037037 0.5674 0.5704 0.5665
MobileNet
Nadam 0.716049383 0.7198 0.716 0.7164
RMSprop 0.728395062 0.7334 0.7284 0.7255
Adam 0.81399177 0.8272 0.814 0.8187
Adamax 0.670781893 0.6913 0.6708 0.6793
ResNet50
Nadam 0.823045267 0.8289 0.823 0.8255
RMSprop 0.83127572 0.8455 0.8313 0.8363
Adam 0.863374486 0.9096 0.8634 0.8742
Adamax 0.902880658 0.909 0.9029 0.9051
VGG16
Nadam 0.902057613 0.9045 0.9021 0.903
RMSprop 0.902880658 0.9094 0.9029 0.9051
Adam 0.903703704 0.9083 0.9037 0.9048
Adamax 0.863374486 0.8751 0.8634 0.8671
VGG19
Nadam 0.890534979 0.9077 0.8905 0.8952
RMSprop 0.86255144 0.8972 0.8626 0.872
Adam 0.824261275 0.8278 0.8243 0.8258
Adamax 0.702954899 0.7082 0.703 0.7015
DenseNet201 Freeze 0
Nadam 0.828926905 0.8302 0.8289 0.8292
RMSprop 0.841368585 0.8457 0.8414 0.843
Adam 0.846034215 0.8499 0.846 0.8472
Adamax 0.712286159 0.7175 0.7123 0.7131
DenseNet201 Freeze 0-30
Nadam 0.833592535 0.8335 0.8336 0.8333
RMSprop 0.836702955 0.8382 0.8367 0.8371
Adam 0.842923795 0.8482 0.8429 0.8444
Adamax 0.727838258 0.7296 0.7278 0.725
DenseNet201 Freeze 0-60
Nadam 0.810264386 0.8121 0.8103 0.8107
RMSprop 0.822706065 0.8211 0.8227 0.8211
Adam 0.839813375 0.8372 0.8398 0.8381
Adamax 0.724727838 0.7278 0.7247 0.7238
DenseNet201 Freeze 0-90
Nadam 0.822706065 0.8249 0.8227 0.8235
RMSprop 0.816485226 0.8157 0.8165 0.8157
Adam 0.786936236 0.7898 0.7869 0.7877
Adamax 0.702954899 0.7091 0.703 0.7042
DenseNet201 Freeze 0-120
Nadam 0.804043546 0.802 0.804 0.8028
RMSprop 0.828926905 0.8271 0.8289 0.8276
Adam 0.800933126 0.7967 0.8009 0.7984
DenseNet201 Freeze 0-150
Adamax 0.707620529 0.6987 0.7076 0.699

132
 Nadam 0.804043546 0.8024 0.804 0.8026
RMSprop 0.808709176 0.8132 0.8087 0.809

Image Size: 64*64

No of epochs: 20

Model Optimizer accuracy precision recall f1-score

Adam 0.734979424 0.7437 0.735 0.7368
Adamax 0.595884774 0.5957 0.5959 0.5893
MobileNet
Nadam 0.753909465 0.7592 0.7539 0.7552
RMSprop 0.748971193 0.7601 0.749 0.7525
Adam 0.810699588 0.8247 0.8107 0.8142
Adamax 0.719341564 0.7305 0.7193 0.724
ResNet50
Nadam 0.825514403 0.8281 0.8255 0.8265
RMSprop 0.826337449 0.8425 0.8263 0.832
Adam 0.883950617 0.8917 0.884 0.8857
Adamax 0.869958848 0.8928 0.87 0.8758
VGG16
Nadam 0.891358025 0.8977 0.8914 0.8935
RMSprop 0.882304527 0.9079 0.8823 0.8885
Adam 0.900411523 0.9044 0.9004 0.9017
Adamax 0.891358025 0.9031 0.8914 0.8953
VGG19
Nadam 0.885596708 0.9054 0.8856 0.8917
RMSprop 0.89382716 0.9101 0.8938 0.898
Adam 0.872427984 0.8857 0.8724 0.8771
Adamax 0.795884774 0.8179 0.7959 0.8027
DenseNet201 Freeze 0
Nadam 0.795884774 0.8179 0.7959 0.8027
RMSprop 0.869958848 0.8766 0.87 0.872
Adam 0.865843621 0.8755 0.8658 0.8694
Adamax 0.803292181 0.8116 0.8033 0.8059
DenseNet201 Freeze 0-30
Nadam 0.86255144 0.8723 0.8626 0.8663
RMSprop 0.872427984 0.8885 0.8724 0.8779
Adam 0.868312757 0.8752 0.8683 0.871
Adamax 0.786831276 0.8016 0.7868 0.7901
DenseNet201 Freeze 0-60
Nadam 0.855967078 0.8632 0.856 0.8587
RMSprop 0.867489712 0.8727 0.8675 0.8688
Adam 0.874897119 0.8766 0.8749 0.8756
Adamax 0.799176955 0.8101 0.7992 0.8027
DenseNet201 Freeze 0-90
Nadam 0.852674897 0.863 0.8527 0.8566
RMSprop 0.86255144 0.8687 0.8626 0.8651
Adam 0.852674897 0.8659 0.8527 0.857
Adamax 0.790946502 0.8008 0.7909 0.7944
DenseNet201 Freeze 0-120
Nadam 0.857613169 0.8694 0.8576 0.8614
RMSprop 0.857613169 0.8702 0.8576 0.8618

133
 Adam 0.84691358 0.8544 0.8469 0.8499
Adamax 0.786831276 0.7909 0.7868 0.7876
DenseNet201 Freeze 0-150
Nadam 0.855144033 0.8596 0.8551 0.8565
RMSprop 0.854320988 0.8601 0.8543 0.8562

Image Size: 64*64

No of epochs: 25

Model Optimizer accuracy precision recall f1-score

Adam 0.769547325 0.7853 0.7695 0.7757
Adamax 0.635390947 0.6317 0.6354 0.6334
MobileNet
Nadam 0.771193416 0.7805 0.7712 0.7749
RMSprop 0.78600823 0.7867 0.786 0.7856
Adam 0.827983539 0.8341 0.828 0.8305
Adamax 0.827983539 0.8341 0.828 0.8305
ResNet50
Nadam 0.826337449 0.8307 0.8263 0.8275
RMSprop 0.832098765 0.8403 0.8321 0.835
Adam 0.903703704 0.9116 0.9037 0.9061
Adamax 0.897119342 0.901 0.8971 0.8983
VGG16
Nadam 0.879012346 0.9049 0.879 0.8867
RMSprop 0.90781893 0.9163 0.9078 0.9103
Adam 0.90617284 0.9069 0.9062 0.9059
Adamax 0.887242798 0.8911 0.8872 0.8886
VGG19
Nadam 0.887242798 0.8928 0.8872 0.8885
RMSprop 0.905349794 0.9024 0.9053 0.9023
Adam 0.865020576 0.8763 0.865 0.8692
Adamax 0.81399177 0.8179 0.814 0.815
DenseNet201 Freeze 0
Nadam 0.869135802 0.8802 0.8691 0.8723
RMSprop 0.867489712 0.8845 0.8675 0.8731
Adam 0.869135802 0.872 0.8691 0.8698
Adamax 0.817283951 0.8273 0.8173 0.8209
DenseNet201 Freeze 0-30
Nadam 0.875720165 0.8885 0.8757 0.8799
RMSprop 0.888065844 0.8903 0.8881 0.8887
Adam 0.855967078 0.8679 0.856 0.86
Adamax 0.811522634 0.8183 0.8115 0.8137
DenseNet201 Freeze 0-60
Nadam 0.865843621 0.8758 0.8658 0.8694
RMSprop 0.885596708 0.8905 0.8856 0.8876
Adam 0.864197531 0.8736 0.8642 0.8675
Adamax 0.809053498 0.8248 0.8091 0.8155
DenseNet201 Freeze 0-90
Nadam 0.851028807 0.8676 0.851 0.857
RMSprop 0.863374486 0.873 0.8634 0.866
Adam 0.86090535 0.8681 0.8609 0.8636
DenseNet201 Freeze 0-120 Adamax 0.791769547 0.8078 0.7918 0.7982
Nadam 0.849382716 0.8678 0.8494 0.8552

134
 RMSprop 0.856790123 0.8647 0.8568 0.8594
Adam 0.849382716 0.8581 0.8494 0.8521
Adamax 0.776954733 0.7913 0.777 0.7816
DenseNet201 Freeze 0-150
Nadam 0.860082305 0.8715 0.8601 0.8637
RMSprop 0.855967078 0.8646 0.856 0.8592

Image Size: 64*64

No of epochs: 30
Model Optimizer accuracy precision recall f1-score
Adam 0.775308642 0.775309 0.7753 0.77531
Adamax 0.667489712 0.671 0.6675 0.6684
MobileNet
Nadam 0.795884774 0.8067 0.7959 0.7999
RMSprop 0.800823045 0.8114 0.8008 0.8049
Adam 0.846090535 0.8579 0.8461 0.8495
Adamax 0.778600823 0.7893 0.7786 0.782
ResNet50
Nadam 0.84526749 0.851 0.8453 0.8472
RMSprop 0.82962963 0.8419 0.8296 0.8328
Adam 0.897942387 0.9044 0.8979 0.9
Adamax 0.895473251 0.8986 0.8955 0.8967
VGG16
Nadam 0.902880658 0.9105 0.9029 0.9048
RMSprop 0.898765432 0.9161 0.8988 0.9034
Adam 0.902057613 0.9042 0.9021 0.9011
Adamax 0.888065844 0.8965 0.8881 0.8911
VGG19
Nadam 0.906995885 0.9165 0.907 0.9101
RMSprop 0.880658436 0.8868 0.8807 0.8779
Adam 0.872427984 0.8795 0.8724 0.874
Adamax 0.824691358 0.8306 0.8247 0.827
DenseNet201 Freeze 0
Nadam 0.87654321 0.8875 0.8765 0.8799
RMSprop 0.880658436 0.8996 0.8807 0.8868
Adam 0.868312757 0.8835 0.8683 0.8729
Adamax 0.834567901 0.8406 0.8346 0.8364
DenseNet201 Freeze 0-30
Nadam 0.865843621 0.8794 0.8658 0.8706
RMSprop 0.879012346 0.8942 0.879 0.8839
Adam 0.874897119 0.8857 0.8749 0.8788
Adamax 0.819753086 0.8312 0.8198 0.824
DenseNet201 Freeze 0-60
Nadam 0.861728395 0.8734 0.8617 0.8661
RMSprop 0.873251029 0.8861 0.8733 0.8773
Adam 0.861728395 0.8643 0.8617 0.8626
Adamax 0.821399177 0.8305 0.8214 0.8248
DenseNet201 Freeze 0-90
Nadam 0.871604938 0.8756 0.8716 0.8733
RMSprop 0.874074074 0.8856 0.8741 0.8781
Adam 0.855144033 0.8692 0.8551 0.8603
DenseNet201 Freeze 0-120
Adamax 0.81399177 0.8257 0.814 0.8187

135
 Nadam 0.857613169 0.8676 0.8576 0.8612
RMSprop 0.847736626 0.8681 0.8477 0.8542
Adam 0.864197531 0.8709 0.8642 0.8665
Adamax 0.788477366 0.8079 0.7885 0.7956
DenseNet201 Freeze 0-150
Nadam 0.84526749 0.8579 0.8453 0.8487
RMSprop 0.863374486 0.8634 0.8634 0.8632

Image Size: 128*128

No of epochs: 15
Model Optimizer accuracy precision recall f1-score
Adam 0.950617284 0.9508 0.9506 0.9507
Adamax 0.930041152 0.9291 0.93 0.9292
MobileNet
Nadam 0.947325103 0.947 0.9473 0.947
RMSprop 0.950617284 0.9503 0.9506 0.9504
Adam 0.922633745 0.9261 0.9226 0.9236
Adamax 0.914403292 0.9151 0.9144 0.9146
ResNet50
Nadam 0.925925926 0.9275 0.9259 0.9266
RMSprop 0.924279835 0.9245 0.9243 0.9241
Adam 0.945679012 0.946 0.9457 0.9452
Adamax 0.942386831 0.9429 0.9424 0.9426
VGG16
Nadam 0.948148148 0.9504 0.9481 0.9484
RMSprop 0.9218107 0.9295 0.9218 0.9221
Adam 0.946502058 0.9461 0.9465 0.9461
Adamax 0.93744856 0.941 0.9374 0.9384
VGG19
Nadam 0.95473251 0.9552 0.9547 0.9549
RMSprop 0.940740741 0.9423 0.9407 0.9412
Adam 0.956378601 0.9574 0.9564 0.9567
Adamax 0.928395062 0.9293 0.9284 0.9285
DenseNet201 Freeze 0
Nadam 0.963786008 0.9637 0.9638 0.9637
RMSprop 0.959670782 0.9598 0.9597 0.9597
Adam 0.956378601 0.9573 0.9564 0.9567
Adamax 0.935802469 0.9356 0.9358 0.9356
DenseNet201 Freeze 0-30
Nadam 0.960493827 0.9604 0.9605 0.9604
RMSprop 0.953909465 0.955 0.9539 0.9541
Adam 0.950617284 0.9527 0.9506 0.9514
Adamax 0.935802469 0.9386 0.9358 0.9368
DenseNet201 Freeze 0-60
Nadam 0.958024691 0.9592 0.958 0.9584
RMSprop 0.95308642 0.9543 0.9531 0.9535
Adam 0.951440329 0.9517 0.9514 0.9515
Adamax 0.922633745 0.9223 0.9226 0.9222
DenseNet201 Freeze 0-90
Nadam 0.95473251 0.9557 0.9547 0.955
RMSprop 0.940740741 0.946 0.9407 0.9422
DenseNet201 Freeze 0-120 Adam 0.950617284 0.9518 0.9506 0.9509

136
 Adamax 0.926748971 0.9289 0.9267 0.9276
Nadam 0.944855967 0.9468 0.9449 0.9454
RMSprop 0.943209877 0.9454 0.9432 0.9438
Adam 0.951440329 0.9524 0.9514 0.9517
Adamax 0.930041152 0.9295 0.93 0.9295
DenseNet201 Freeze 0-150
Nadam 0.939917695 0.942 0.9399 0.9404
RMSprop 0.949794239 0.9511 0.9498 0.9502

Image Size: 128*128

No of epochs: 20
Model Optimizer accuracy precision recall f1-score
Adam 0.95473251 0.9551 0.9547 0.9549
Adamax 0.942386831 0.9417 0.9424 0.9419
MobileNet
Nadam 0.941563786 0.9417 0.9416 0.9416
RMSprop 0.95308642 0.9531 0.9531 0.9531
Adam 0.922633745 0.9286 0.9226 0.9243
Adamax 0.914403292 0.915 0.9144 0.9146
ResNet50
Nadam 0.928395062 0.9302 0.9284 0.929
RMSprop 0.926748971 0.9291 0.9267 0.9276
Adam 0.95308642 0.9545 0.9531 0.9531
Adamax 0.948148148 0.9511 0.9481 0.949
VGG16
Nadam 0.943209877 0.945 0.9432 0.9438
RMSprop 0.950617284 0.9532 0.9506 0.9515
Adam 0.955555556 0.9562 0.9556 0.9558
Adamax 0.936625514 0.9382 0.9366 0.9372
VGG19
Nadam 0.938271605 0.938 0.9383 0.9371
RMSprop 0.950617284 0.9526 0.9506 0.9513
Adam 0.956378601 0.9563 0.9564 0.9562
Adamax 0.941563786 0.944 0.9416 0.9425
DenseNet201 Freeze 0
Nadam 0.956378601 0.9566 0.9564 0.9565
RMSprop 0.967901235 0.9684 0.9679 0.9681
Adam 0.958024691 0.9586 0.958 0.9582
Adamax 0.943209877 0.9448 0.9432 0.9437
DenseNet201 Freeze 0-30
Nadam 0.955555556 0.9569 0.9556 0.956
RMSprop 0.956378601 0.9566 0.9564 0.9562
Adam 0.952263374 0.9534 0.9523 0.9527
Adamax 0.930864198 0.9312 0.9309 0.931
DenseNet201 Freeze 0-60
Nadam 0.948148148 0.949 0.9481 0.9483
RMSprop 0.956378601 0.9566 0.9564 0.9565
Adam 0.947325103 0.9487 0.9473 0.9477
Adamax 0.941563786 0.9427 0.9416 0.942
DenseNet201 Freeze 0-90
Nadam 0.953909465 0.955 0.9539 0.9543
RMSprop 0.949794239 0.9507 0.9498 0.95

137
 Adam 0.95308642 0.9538 0.9531 0.9533
Adamax 0.930864198 0.9353 0.9309 0.9324
DenseNet201 Freeze 0-120
Nadam 0.953909465 0.9569 0.9539 0.9548
RMSprop 0.955555556 0.9568 0.9556 0.956
Adam 0.948148148 0.9506 0.9481 0.9487
Adamax 0.938271605 0.9385 0.9383 0.9383
DenseNet201 Freeze 0-150
Nadam 0.953909465 0.9542 0.9539 0.9539
RMSprop 0.948148148 0.95 0.9481 0.9486

Image Size: 128*128

No of epochs: 25
Model Optimizer accuracy precision recall f1-score
Adam 0.953909465 0.954 0.9539 0.9539
Adamax 0.941563786 0.9412 0.9416 0.9413
MobileNet
Nadam 0.948148148 0.9481 0.9481 0.948
RMSprop 0.949794239 0.9513 0.9498 0.9503
Adam 0.926748971 0.9319 0.9267 0.9282
Adamax 0.919341564 0.9195 0.9193 0.9193
ResNet50
Nadam 0.930864198 0.9338 0.9309 0.9318
RMSprop 0.9218107 0.924 0.9218 0.9224
Adam 0.95473251 0.9552 0.9547 0.9548
Adamax 0.929218107 0.9335 0.9292 0.9299
VGG16
Nadam 0.948971193 0.9491 0.949 0.9488
RMSprop 0.95308642 0.9553 0.9531 0.9538
Adam 0.934156379 0.9358 0.9342 0.9343
Adamax 0.938271605 0.9385 0.9383 0.9384
VGG19
Nadam 0.952263374 0.9522 0.9523 0.952
RMSprop 0.947325103 0.9478 0.9473 0.947
Adam 0.960493827 0.961 0.9605 0.9607
Adamax 0.941563786 0.9427 0.9416 0.9419
DenseNet201 Freeze 0
Nadam 0.953909465 0.9552 0.9539 0.9541
RMSprop 0.947325103 0.9486 0.9473 0.9478
Adam 0.962139918 0.9621 0.9621 0.962
Adamax 0.948971193 0.9494 0.949 0.9488
DenseNet201 Freeze 0-30
Nadam 0.958024691 0.9582 0.958 0.9581
RMSprop 0.947325103 0.949 0.9473 0.9479
Adam 0.951440329 0.9527 0.9514 0.9519
Adamax 0.948148148 0.95 0.9481 0.9488
DenseNet201 Freeze 0-60
Nadam 0.949794239 0.9504 0.9498 0.9498
RMSprop 0.93744856 0.938 0.9374 0.9375
Adam 0.953909465 0.9546 0.9539 0.9541
DenseNet201 Freeze 0-90 Adamax 0.944855967 0.9455 0.9449 0.9451
Nadam 0.953909465 0.9547 0.9539 0.9542

138
 RMSprop 0.93744856 0.9386 0.9374 0.9378
Adam 0.949794239 0.9521 0.9498 0.9506
Adamax 0.936625514 0.9392 0.9366 0.9375
DenseNet201 Freeze 0-120
Nadam 0.953909465 0.955 0.9539 0.9543
RMSprop 0.941563786 0.9439 0.9416 0.9424
Adam 0.949794239 0.951 0.9498 0.9502
Adamax 0.942386831 0.9432 0.9424 0.9427
DenseNet201 Freeze 0-150
Nadam 0.95473251 0.9553 0.9547 0.9547
RMSprop 0.922633745 0.9292 0.9226 0.9248

Image Size: 128*128

No of epochs: 30
Model Optimizer accuracy precision recall f1-score
Adam 0.950617284 0.9505 0.9506 0.9504
Adamax 0.945679012 0.9458 0.9457 0.9456
MobileNet
Nadam 0.950617284 0.9522 0.9506 0.9512
RMSprop 0.944032922 0.9452 0.944 0.9444
Adam 0.938271605 0.9394 0.9383 0.9387
Adamax 0.904526749 0.9054 0.9045 0.9047
ResNet50
Nadam 0.934156379 0.9364 0.9342 0.9349
RMSprop 0.924279835 0.9254 0.9243 0.9247
Adam 0.934156379 0.9394 0.9342 0.935
Adamax 0.948148148 0.948 0.9481 0.9479
VGG16
Nadam 0.95308642 0.9526 0.9531 0.9526
RMSprop 0.955555556 0.9558 0.9556 0.9556
Adam 0.935802469 0.9466 0.9358 0.9384
Adamax 0.939917695 0.9404 0.9399 0.94
VGG19
Nadam 0.958847737 0.9605 0.9588 0.9593
RMSprop 0.944032922 0.9492 0.944 0.945
Adam 0.958847737 0.9595 0.9588 0.9591
Adamax 0.944855967 0.9456 0.9449 0.9451
DenseNet201 Freeze 0
Nadam 0.958847737 0.9594 0.9588 0.9589
RMSprop 0.950617284 0.9517 0.9506 0.9506
Adam 0.958024691 0.9586 0.958 0.958
Adamax 0.939917695 0.9417 0.9399 0.9404
DenseNet201 Freeze 0-30
Nadam 0.957201646 0.9588 0.9572 0.9577
RMSprop 0.952263374 0.9539 0.9523 0.9527
Adam 0.955555556 0.9576 0.9556 0.9563
Adamax 0.933333333 0.9366 0.9333 0.9344
DenseNet201 Freeze 0-60
Nadam 0.949794239 0.9501 0.9498 0.9497
RMSprop 0.948148148 0.9523 0.9481 0.9494
Adam 0.962139918 0.9626 0.9621 0.9623
DenseNet201 Freeze 0-90
Adamax 0.93744856 0.9392 0.9374 0.9379

139
 Nadam 0.957201646 0.9577 0.9572 0.9574
RMSprop 0.949794239 0.952 0.9498 0.9504
Adam 0.95308642 0.9547 0.9531 0.9533
Adamax 0.93909465 0.9393 0.9391 0.9391
DenseNet201 Freeze 0-120
Nadam 0.947325103 0.9491 0.9473 0.9479
RMSprop 0.948971193 0.9517 0.949 0.9495
Adam 0.95308642 0.9541 0.9531 0.9534
Adamax 0.93909465 0.9413 0.9391 0.9398
DenseNet201 Freeze 0-150
Nadam 0.949794239 0.9505 0.9498 0.9499
RMSprop 0.95308642 0.953 0.9531 0.9529

Image Size: 224*224

No of epochs: 15

Model Optimizer accuracy precision recall f1-score

Adam 0.960493827 0.9608 0.9605 0.9606
Adamax 0.953909465 0.9541 0.9539 0.954
MobileNet
Nadam 0.963786008 0.9636 0.9638 0.9635
RMSprop 0.957201646 0.9574 0.9572 0.9573
Adam 0.967078189 0.9676 0.9671 0.9672
Adamax 0.960493827 0.9604 0.9605 0.9604
ResNet50
Nadam 0.967901235 0.9683 0.9679 0.9681
RMSprop 0.961316872 0.9617 0.9613 0.961
Adam 0.960493827 0.9619 0.9605 0.961
Adamax 0.942386831 0.9423 0.9424 0.9423
VGG16
Nadam 0.952263374 0.9533 0.9523 0.951
RMSprop 0.944032922 0.955 0.944 0.9464
Adam 0.957201646 0.9591 0.9572 0.9569
Adamax 0.950617284 0.9504 0.9506 0.9502
VGG19
Nadam 0.972839506 0.9726 0.9728 0.9726
RMSprop 0.966255144 0.9665 0.9663 0.9657
Adam 0.973662551 0.9739 0.9737 0.9737
Adamax 0.977777778 0.9777 0.9778 0.9777
DenseNet201 Freeze 0
Nadam 0.978600823 0.9787 0.9786 0.9786
RMSprop 0.979423868 0.9794 0.9794 0.9793
Adam 0.98600823 0.9861 0.986 0.986
Adamax 0.976954733 0.9771 0.977 0.977
DenseNet201 Freeze 0-30
Nadam 0.983539095 0.9835 0.9835 0.9835
RMSprop 0.976954733 0.978 0.977 0.9773
Adam 0.980246914 0.9805 0.9802 0.9803
Adamax 0.972839506 0.9729 0.9728 0.9728
DenseNet201 Freeze 0-60
Nadam 0.977777778 0.9777 0.9778 0.9777
RMSprop 0.974485597 0.9747 0.9745 0.9746
DenseNet201 Freeze 0-90 Adam 0.980246914 0.9805 0.9802 0.9803

140
 Adamax 0.973662551 0.9738 0.9737 0.9737
Nadam 0.975308642 0.9752 0.9753 0.9752
RMSprop 0.976131687 0.9765 0.9761 0.9762
Adam 0.980246914 0.9802 0.9802 0.9802
Adamax 0.973662551 0.9741 0.9737 0.9738
DenseNet201 Freeze 0-120
Nadam 0.981069959 0.9811 0.9811 0.9811
RMSprop 0.977777778 0.978 0.9778 0.9777
Adam 0.977777778 0.9778 0.9778 0.9778
Adamax 0.96872428 0.9686 0.9687 0.9686
DenseNet201 Freeze 0-150
Nadam 0.980246914 0.9804 0.9802 0.9803
RMSprop 0.978600823 0.9789 0.9786 0.9787

Image Size: 224*224

No of epochs: 20

Model Optimizer accuracy precision recall f1-score

Adam 0.967901235 0.9677 0.9679 0.9677
Adamax 0.944855967 0.9453 0.9449 0.945
MobileNet
Nadam 0.96872428 0.9685 0.9687 0.9686
RMSprop 0.946502058 0.9467 0.9465 0.9465
Adam 0.97037037 0.9706 0.9704 0.9705
Adamax 0.961316872 0.9613 0.9613 0.9613
ResNet50
Nadam 0.974485597 0.9745 0.9745 0.9745
RMSprop 0.967901235 0.9679 0.9679 0.9679
Adam 0.967901235 0.9678 0.9679 0.9677
Adamax 0.944032922 0.9436 0.944 0.9436
VGG16
Nadam 0.969547325 0.9699 0.9695 0.9696
RMSprop 0.948971193 0.9592 0.949 0.9513
Adam 0.920164609 0.9254 0.9202 0.92
Adamax 0.942386831 0.9461 0.9424 0.9429
VGG19
Nadam 0.962962963 0.9633 0.963 0.9631
RMSprop 0.97037037 0.9718 0.9704 0.9707
Adam 0.979423868 0.9794 0.9794 0.9794
Adamax 0.975308642 0.9755 0.9753 0.9754
DenseNet201 Freeze 0
Nadam 0.980246914 0.9803 0.9802 0.9803
RMSprop 0.981893004 0.9818 0.9819 0.9819
Adam 0.967901235 0.9688 0.9679 0.9681
Adamax 0.972016461 0.9722 0.972 0.9721
DenseNet201 Freeze 0-30
Nadam 0.974485597 0.9746 0.9745 0.9745
RMSprop 0.976131687 0.9762 0.9761 0.9758
Adam 0.980246914 0.9803 0.9802 0.9802
Adamax 0.975308642 0.9754 0.9753 0.9753
DenseNet201 Freeze 0-60
Nadam 0.981069959 0.9813 0.9811 0.9812
RMSprop 0.982716049 0.9827 0.9827 0.9827

141
 Adam 0.982716049 0.9827 0.9827 0.9827
Adamax 0.97037037 0.9705 0.9704 0.9704
DenseNet201 Freeze 0-90
Nadam 0.980246914 0.9805 0.9802 0.9803
RMSprop 0.972839506 0.9743 0.9728 0.9732
Adam 0.975308642 0.9754 0.9753 0.975
Adamax 0.974485597 0.9746 0.9745 0.9745
DenseNet201 Freeze 0-120
Nadam 0.981069959 0.981 0.9811 0.981
RMSprop 0.974485597 0.9754 0.9745 0.9747
Adam 0.981069959 0.981 0.9811 0.981
Adamax 0.976131687 0.9762 0.9761 0.9762
DenseNet201 Freeze 0-150
Nadam 0.97037037 0.971 0.9704 0.9706
RMSprop 0.978600823 0.9787 0.9786 0.9786

Image Size: 224*224

No of epochs: 25
Model Optimizer accuracy precision recall f1-score
Adam 0.962139918 0.9621 0.9621 0.962
Adamax 0.953909465 0.9539 0.9539 0.9538
MobileNet
Nadam 0.958024691 0.9578 0.958 0.9576
RMSprop 0.959670782 0.9606 0.9597 0.96
Adam 0.967901235 0.9679 0.9679 0.9678
Adamax 0.961316872 0.9611 0.9613 0.9611
ResNet50
Nadam 0.967078189 0.9672 0.9671 0.9671
RMSprop 0.972839506 0.973 0.9728 0.9729
Adam 0.929218107 0.9376 0.9292 0.9311
Adamax 0.93744856 0.951 0.9374 0.9408
VGG16
Nadam 0.958024691 0.9648 0.958 0.9596
RMSprop 0.927572016 0.9366 0.9276 0.9211
Adam 0.967901235 0.9689 0.9679 0.9682
Adamax 0.952263374 0.9527 0.9523 0.952
VGG19
Nadam 0.963786008 0.9643 0.9638 0.9639
RMSprop 0.952263374 0.9555 0.9523 0.9527
Adam 0.981069959 0.981 0.9811 0.981
Adamax 0.976954733 0.9769 0.977 0.9769
DenseNet201 Freeze 0
Nadam 0.979423868 0.9795 0.9794 0.9795
RMSprop 0.978600823 0.9786 0.9786 0.9786
Adam 0.979423868 0.9794 0.9794 0.9793
Adamax 0.979423868 0.9796 0.9794 0.9795
DenseNet201 Freeze 0-30
Nadam 0.981893004 0.9819 0.9819 0.9818
RMSprop 0.981069959 0.981 0.9811 0.981
Adam 0.981893004 0.9819 0.9819 0.9819
DenseNet201 Freeze 0-60 Adamax 0.977777778 0.9779 0.9778 0.9778
Nadam 0.983539095 0.9835 0.9835 0.9835

142
 RMSprop 0.983539095 0.9835 0.9835 0.9835
Adam 0.98436214 0.9844 0.9844 0.9843
Adamax 0.966255144 0.9665 0.9663 0.9663
DenseNet201 Freeze 0-90
Nadam 0.982716049 0.9828 0.9827 0.9828
RMSprop 0.982716049 0.9828 0.9827 0.9827
Adam 0.976131687 0.9763 0.9761 0.9762
Adamax 0.978600823 0.9787 0.9786 0.9786
DenseNet201 Freeze 0-120
Nadam 0.976954733 0.9773 0.977 0.9771
RMSprop 0.975308642 0.9753 0.9753 0.9752
Adam 0.975308642 0.9752 0.9753 0.9753
Adamax 0.976954733 0.977 0.977 0.9769
DenseNet201 Freeze 0-150
Nadam 0.981069959 0.981 0.9811 0.981
RMSprop 0.986831276 0.9871 0.9868 0.9869

Image Size: 224*224

No of epochs: 30
Model Optimizer accuracy precision recall f1-score
Adam 0.967078189 0.9673 0.9671 0.9672
Adamax 0.961316872 0.9614 0.9613 0.9609
MobileNet
Nadam 0.96872428 0.9688 0.9687 0.9688
RMSprop 0.963786008 0.9638 0.9638 0.9636
Adam 0.967078189 0.9673 0.9671 0.9671
Adamax 0.967078189 0.9671 0.9671 0.9671
ResNet50
Nadam 0.960493827 0.9611 0.9605 0.9607
RMSprop 0.959670782 0.9615 0.9597 0.9602
Adam 0.967078189 0.9685 0.9671 0.9675
Adamax 0.95308642 0.9553 0.9531 0.9538
VGG16
Nadam 0.966255144 0.9678 0.9663 0.9667
RMSprop 0.967901235 0.9708 0.9679 0.9686
Adam 0.969547325 0.9694 0.9695 0.9688
Adamax 0.948971193 0.9484 0.949 0.9481
VGG19
Nadam 0.967901235 0.9688 0.9679 0.9681
RMSprop 0.963786008 0.9647 0.9638 0.9639
Adam 0.976131687 0.9763 0.9761 0.9762
Adamax 0.973662551 0.9737 0.9737 0.9737
DenseNet201 Freeze 0
Nadam 0.98436214 0.9844 0.9844 0.9844
RMSprop 0.976131687 0.9763 0.9761 0.9762
Adam 0.978600823 0.9786 0.9786 0.9783
Adamax 0.978600823 0.9786 0.9786 0.9786
DenseNet201 Freeze 0-30
Nadam 0.978600823 0.9785 0.9786 0.9785
RMSprop 0.977777778 0.9778 0.9778 0.9778
Adam 0.98436214 0.9844 0.9844 0.9843
DenseNet201 Freeze 0-60
Adamax 0.976131687 0.9765 0.9761 0.9762

143
 Nadam 0.97037037 0.9713 0.9704 0.9707
RMSprop 0.982716049 0.9828 0.9827 0.9827
Adam 0.979423868 0.9795 0.9794 0.9794
Adamax 0.976954733 0.9769 0.977 0.9769
DenseNet201 Freeze 0-90
Nadam 0.980246914 0.9803 0.9802 0.9803
RMSprop 0.985185185 0.9852 0.9852 0.9851
Adam 0.982716049 0.9827 0.9827 0.9827
Adamax 0.972839506 0.9732 0.9728 0.973
DenseNet201 Freeze 0-120
Nadam 0.983539095 0.9837 0.9835 0.9836
RMSprop 0.988477366 0.9885 0.9885 0.9885
Adam 0.976954733 0.977 0.977 0.9769
Adamax 0.969547325 0.97 0.9695 0.9697
DenseNet201 Freeze 0-150
Nadam 0.976954733 0.977 0.977 0.9767
RMSprop 0.975308642 0.9751 0.9753 0.9752

144
 ‫پوختە‬
‫ژمارەی دانیشتوانی جیهان بە شێوەیەکی بەرچاو زیادی کردووە‪ ،‬ئەمەش کاریگەری لەسەر‬
‫بەکارهێنانی ئۆتۆمبێل لەالیەن تاکەکانەوە هەیە و دەبێتە هۆی زیادبوونی ژمارەی ئۆتۆمبێل لە‬
‫شارەکاندا‪ .‬بەهۆی هەبونی پەیوەندی ڕاستەوخۆی نێوان دانیشتووان و بەکارهێنانی ئۆتۆمبێل‪ ،‬ئەمەش‬
‫وادەکات کە بەڕێوەبردنی هاتوچۆ بووەتە پرسێکی گرنگ کە پێویستە چارەسەر بکرێت‪ .‬بۆ ئەم‬
‫مەبەستە‪ ،‬ئاماژەدانی هاتوچۆی زیرەک بە ناو شارەکان خێرا پێویستە بۆ زاڵبوون بەسەر قەرەباڵغی‬
‫هاتوچۆ‪ ،‬و کەمکردنەوەی تێچوون و کاتی گەشتکردن‪ .‬بۆ زاڵبوون بەسەر ئەم کێشانەدا‪ ،‬بینینی‬
‫کۆمپیوتەر (‪ )Computer Vision‬و فێربوونی قووڵ (‪ )Deep Learning‬بژاردەی گرنگن بۆ‬
‫مامەڵەکردن لەگەڵ ئەم پرسەدا چونکە ڕۆڵێکی گرنگ دەگێڕن بۆ بەڕێوەبردن و کۆنترۆڵکردنی‬
‫سیگناڵەکانی هاتوچۆ‪ .‬سەرەڕای ئەوەش‪ ،‬دۆزینەوە و جیاکردنەوەی نێوان شتەکان(‪)Objects‬‬
‫یارمەتیدەرە بۆ ژماردنی ئۆتۆمبێلەکان و شتەکانی تر کە خۆیان لە قەرەباڵغی و کۆنترۆڵکردنی‬
‫سیگناڵەکان لە ناوچەکانی هاتوچۆدا بەدوور دەگرن‪ .‬جگە لەوەش دۆزینەوەی ئۆتۆمبێلی فریاگوزاری‬
‫و پێدانی ئەولەویەت بۆیان کە پێویستە بۆ سیستەمی زیرەکی ئاماژەدانی هاتوچۆ ‪.‬‬
‫ئامانجی سەرەکی ئەم توێژینەوەیە دیزاینکردن و جێبەجێکردنی سیستەمێکی کارامەیە بۆ سیستەمی‬
‫سیگناڵی هاتوچۆ لەسەر بنەمای دۆزینەوەی ئۆتۆمبێل لە ترافیکەکان‪ .‬جگە لەوەش‪ ،‬سیستەمی‬
‫پێشنیارکراوی چوار قۆناغ لەخۆدەگرێت؛ یەکەمیان گرتنی وێنە لە هەردوو کامێرای هاوشێوە و کاتی‬
‫ڕاستەقینە لە ڕێگاکانەوە‪ .‬لە قۆناغی دووەمدا ئەلگۆریتمەکانی جیاوازی پێش پرۆسێسکردنی وێنە بۆ‬
‫وێنە گیراوەکان وەک هەنگاوێکی پێش پرۆسێسکردن ئەنجام دەدرێن‪ .‬جگە لەوەش‪ ،‬تەکنیکەکانی‬
‫فێربوونی قووڵ بەکاردەهێنرێن بۆ دیاریکردنی شتەکانی وەک ئۆتۆمبێلەکانی (ئاسایی‪ ،‬پۆلیس‪،‬‬
‫فریاگوزاری‪ ،‬و ئاگرکوژێنەوە و هتد‪ )..‬لەکاتێکدا لە قۆناغی کۆتاییدا‪ ،‬سیستەمی پێشنیارکراوی‬
‫تاقیدەکرێتەوە بۆ هەڵسەنگاندنی وردی کارایی ئۆتۆمبێلە دۆزراوەکان ‪.‬‬
‫ڕێبازێکی فێربوونی گواستنەوەی دەستکاریکراو بۆ مۆدێلی ‪ DenseNet201‬بەکارهێنراوە بۆ‬
‫چەندین پۆلێنکردن‪ ،‬لەوانەش ئۆتۆمبێلەکانی (ئاسایی‪ ،‬فریاگوزاری‪ ،‬ئەمبوالنس‪ ،‬پۆلیس و‬
‫ئاگرکوژێنەوە)‪ .‬ڕێبازەکە بریتییە لە بەستنی هەندێ لە چینەکانی مۆدێلەکە‪ .‬ڕێژەی وردبینی بەرز بەم‬
‫مۆدێلە بەدەست دێت و دەگاتە ‪ .%98.6‬هەروەها‪ ،‬شێوازە جیاوازەکانی باشکردن‪ ،‬لەوانە ( ‪Adam,‬‬
‫‪ )Adamax, Nadam, and RMSprob‬بەکاردەهێنرێن بۆ باشترکردنی ئەدای دیاریکردن لەسەر‬
‫بنەمای باشترین هەڵبژاردنی باشترکەر و وردبینی ‪%98.84‬ی بەدەستهێنا‪ .‬سەرەڕای ئەوەش‪،‬‬
‫وەشانێکی دەستکاریکراوی ‪ YOLOv5‬پێشنیارکرا بۆ دۆزینەوەی ئۆتۆمبێل‪ ،‬کە ئامانجی‬
‫بەرزکردنەوەی مامناوەندی وردبینی مامناوەندە (‪ )mAP‬بە ڕێژەی ‪ .%3‬لە کۆتاییدا‪ ،‬سیستەمی‬

‫‪145‬‬
‫پێشنیارکراوی هاوشێوە کرا بۆ کەمکردنەوەی کاتی چاوەڕوانی لە سیگناڵی هاتوچۆ‪ .‬ئەنجامی‬
‫تاقیکردنەوەکان کەمبوونەوەی بەرچاو لە کاتی چاوەڕوانیدا نیشان دەدەن‪ ،‬کە لە نێوان ‪ ٣٠‬بۆ ‪١٠٠‬‬
‫چرکەدایە بەپێی دۆخەکان‪.‬‬

‫‪146‬‬