Traffic Management Techniques : A Survey

Naman Kapoor1*, Preeti Aggarwal1† and Akashdeep Sharma1†

1 University Institute of Engineering and Technology, Panjab University,
Chandigarh, 160014, India.

*Corresponding author(s). E-mail(s): namankapoor1903@gmail.com;

Contributing authors: pree agg@pu.ac.in; akashdeep@pu.ac.in;
† These authors contributed equally to this work.

Abstract
Efficient traffic signal control is very crucial for managing vehicular flow and
alleviating congestion in urban areas. This survey paper presents a thorough
examination of diverse techniques employed in traffic signal control. It begins
by exploring traditional methods like fixed-time control and actuated control,
progressing to advanced intelligent systems incorporating reinforcement learning,
support vector machine and neural network algorithms. The survey encompasses
a broad spectrum of approaches, spanning classical optimization techniques to
the most advanced machine learning and deep learning methods. Each method’s
principles, advantages, and limitations are discussed, alongside their practical
applications in real-world traffic management scenarios. Furthermore, the paper
investigates recent advancements and emerging trends in traffic signal control,
such as connected and autonomous vehicles, cooperative control systems, and
adaptive signal optimization strategies. By synthesizing insights from existing lit-
erature and highlighting current research directions, this survey paper will help in
transportation engineering and urban planning. It offers comprehensive insights
into the complexities and opportunities in traffic signal control, facilitating the
development of innovative and sustainable solutions to tackle urban mobility
challenges.

Keywords: Machine Learning, Deep Learning, Object Detection

1 Introduction
Urbanization has led to a proliferation of vehicular traffic on roads, encompassing
diverse categories such as two-wheelers, three-wheelers, and four-wheelers. This surge

1
in traffic has given rise to a pervasive issue: congestion[1][2]. It is a ubiquitous problem
in urban areas, impeding the smooth flow of daily activities for countless individ-
uals. The congestion caused by the multitude of vehicles on the road often brings
daily life to a standstill. Commuters find themselves trapped in long queues of vehi-
cles, their progress slowed to a crawl. This impediment not only wastes time but
also adds stress and frustration to people’s lives. Many techniques have been used to
detect traffic[3][4][5][6] and manage it but a lot of shortcomings have been noticed.[7]
Furthermore, the gridlock negatively impacts the environment, as engines idle and
emissions increase. This leads to air pollution[8], further exacerbating the health
concerns[9] already present in densely populated urban areas. Efforts to mitigate traf-
fic congestion[10] are ongoing, with various strategies being explored. These include
improvements in public transportation, the development of alternative transportation
modes such as cycling lanes[11][12],pedestrian lights[13] and the implementation of
smart traffic management systems. An effective traffic signal control system seeks to
minimize vehicle congestion at intersections, thereby reducing fuel consumption and
optimizing time utilization. By efficiently managing the flow of traffic, these systems
aim to mitigate delays and waiting times, enhancing overall road efficiency and safety.
Through the strategic coordination of traffic signals, the system endeavors to main-
tain a steady and balanced movement of vehicles, preventing gridlock and bottlenecks.
This not only alleviates frustration among drivers but also contributes to a reduc-
tion in carbon emissions and environmental pollution associated with idling vehicles.
Moreover, a well-designed traffic signal control system can enhance the overall func-
tionality of urban transportation networks. By synchronizing signals and adapting to
real-time traffic conditions, it enables smoother and more predictable travel routes for
commuters, businesses, and emergency services alike.
To achieve these objectives, modern traffic control systems often incorporate
advanced technologies such as real-time data analytics, adaptive signal control algo-
rithms, and communication networks. These innovations enable dynamic adjustments
to signal timings based on traffic volume, flow patterns, and priority demands, opti-
mizing the utilization of road infrastructure and enhancing the quality of urban life.
In essence, the implementation of an efficient traffic signal control system plays a vital
role in promoting sustainable and resilient urban mobility, fostering economic pro-
ductivity, and improving the overall well-being of communities. A prevalent approach
in modern traffic management is the implementation of static traffic light control
systems. These systems rely on historical traffic congestion data to regulate signal
timings. However, this method is inherently limited as it does not allow for real-time
adjustments. Consequently, the fixed signal timings may not effectively address the
dynamic nature of traffic congestion prevalent in contemporary urban environments.
As a result, there is a growing recognition of the need for more adaptive and respon-
sive traffic control solutions to better manage congestion and optimize traffic flow in
today’s bustling cities.
Numerous promising traffic signal control and traffic management techniques have
been proposed, showing potential for significant improvements upon implementation.
In the forthcoming sections of the paper, we will delve into some of the latest tech-
nologies that hold promise in revolutionizing traffic management. These innovations

2
 Table 1 Inclusion Exclusion Criterion

Attributes Inclusion Criterion Exclusion Criterion

T ime Range Papers Published from 2019 Papers Published before 2019
S creening Traffic Management, Traffic Traffic Signal Detection
Congestion
M odality Studies using Modern Tech- Studies using old techniques and
niques such as Deep Learning, algorithms
Machine Learning IoT only
S tudy Selection Studies having proper experi- Studies having no proper clas-
ments and results sifications, experimentation,
no results showcased,Review
Paper,Papers other than
English

offer the possibility of addressing the complexities of modern traffic challenges more
effectively, paving the way for more efficient and sustainable urban mobility solutions.

2 Materials and Methods

Systematic review follow PRISMA guidelines established by the PRISMA group to
ensure the comprehensive reporting and identification of relevant papers. These guide-
lines serve as a framework for authors, facilitating transparency and consistency in
review process, providing a structured approach for selecting and including pertinent
studies. Electronic databases like IEEEXplore, Springer, Sciencedirect are used to
carry out a detailed search for related papers. To select the papers the papers with
relevant keywords and different combinations such as ”traffic signal control,” ”traffic
management,” and ”traffic signal optimization” were used.
Initially, 1647 papers were identified, of which those published before 2019 and
duplicate papers were excluded, resulting in 108 unique papers. From these 108 papers,
screening was conducted based on the relevance of their abstracts, introductions, and
exact topics such as traffic management and congestion. Additionally, review papers
and those not in English were removed, leaving 54 papers. Subsequently, papers that
did not discuss techniques related to deep learning, machine learning, or IoT were
filtered out, resulting in a final selection of 30 relevant papers.

2.1 Quality Assessment

This section offers an insight into past research endeavors within the realm of traffic
signal control, shedding light on the array of techniques utilized. It delves into the
methodologies adopted in previous studies, encompassing a spectrum of approaches
aimed at optimizing traffic flow and reducing congestion.

2.2 Algorithms used for Traffic Signal Control

This section explores a range of techniques employed in traffic signal control. These
investigations encompass various techniques such as Reinforcement Learning (RL) and
Support Vector Machine (SVM), Neural Networks(NN) etc. Through examining these
prior studies, a comprehensive understanding of the evolution and application of traffic
signal control techniques is elucidated.

3
 Fig. 1 Inclusion Exclusion Criteria

In later chapters we will discuss the various methods used in order to implement
traffic signal management.

2.3 Fixed-timer traffic signal control

A fixed-timer traffic signal control system is one of the simplest and oldest methods
used for traffic signal management. In this system, the timing of traffic signals at inter-
sections remains constant and does not adjust based on real-time traffic conditions.
Instead, predetermined fixed time intervals are assigned to each phase of the signal
cycle, such as green, yellow, and red. Fixed-timer systems are easy to implement and
require minimal maintenance. They provide a predictable pattern of signal changes,
which can help drivers anticipate when they need to stop or proceed through an inter-
section. However, these systems do not adapt to changes in traffic flow throughout
the day or in response to incidents such as accidents or road construction. One of the
main drawbacks of fixed-timer systems is their inability to optimize traffic flow effi-
ciently, especially during peak hours or in areas with fluctuating traffic patterns. This
can lead to congestion, delays, and increased fuel consumption for vehicles waiting
at red lights. Despite their limitations, fixed-timer traffic signal control systems are
still widely used in many regions, particularly in areas with relatively stable traffic

4
 Fig. 2 Techniques used in Traffic Signal Controls

conditions or limited resources for more advanced signal management systems. Due
to these shortcomings the new techniques for more adaptive traffic signal controlling
were introduced.

2.4 Reinforcement Learning

Reinforcement Learning (RL)[14] [15][16] is a computational approach to autonomous
decision-making, where an agent learns to maximize cumulative rewards through inter-
action with an environment. The agent employs a policy, which maps states to actions,
to navigate the environment. At each time step,[17][18] the agent observes a state,
selects an action according to its policy, and receives feedback in the form of a reward
signal. The objective is to learn an optimal policy that maximizes the expected sum
of rewards over time.
Key components of RL include the value function, which estimates the expected
return from a given state-action pair, and the Q-function, which estimates the expected
return for taking a specific action from a given state. Value-based methods, such as
Q-learning and Deep Q-Networks (DQN), optimize the value function directly, while
policy-based methods, like Policy Gradient algorithms, aim to directly optimize the

5
 Table 2 Reinforcement Learning Based work summary

Paper Dataset Model(s) Simulation Performance Performance

Used Method Metrics
[19]Essa and Anonymous Deep Rein- VISSIM Real Time 90% safety
Sayed forcement Safety and 50%
Learning Safety under
MPR of 50%
and 30% resp
[20]Guo et al. Anonymous Deep Rein- None Travel Time, upto 74%
forcement Fuel Con- reduction in
Learning sumption delay and
Reduction 30% reduction
in travel time
[21]Zhu et al. G-X Dataset, Deep Rein- Cityflow Sim- Avg. Travel upto 28%
Jinan forcement ulator Time reduction in
Dataset, Learning travel time
Hangzhou
Dataset
[22]Zhao et Anonymous Double Deep SUMO Avg. Waiting 35% improve-
al. Q Network Time / Avg. ment over
Queue Length previous algo-
rithms
[23]Kwok et Anonymous Deep Rein- Pygame[24] Avg. Vehi- 44-60% reduc-
al. forcement cle Waiting tion in wait-
Learning Time, Avg. ing time
Pedestrian
Waiting Time
[25]Li et al. Anonymous Deep Rein- SUMO Avg. Vehicle -
forcement Waiting Time
Learning
[26]Liu et al. D New-York Deep Rein- Cityflow Travel Time, 39% to
Dataset forcement Avg. Waiting 60.72%
Learning Time reduction in
waiting time
[27]Song et Anonymous Multi-agent Real life simu- Queue length, An average of
al. counter- lation Travel time 45% reduc-
factual delay tion in queue
actor–critic length
with sched-
uler
[28]Wang et Anonymous Cooperative SUMO Avg. Waiting Waiting time
al. Group-Based Time reduction by
Multi-Agent 64.02%
reinforcement
learning
[19]Essa and Anonymous Deep Rein- Pygame Avg. Vehi- Avg. of 20%
Sayed forcement cle Waiting reduction in
Learning Time, Avg. travel time
Pedestrian and 37%
Waiting Time reduction in
queue length

policy itself. Actor-Critic methods combine elements of both, utilizing a value function
(the critic) to provide feedback to the policy (the actor).
Reinforcement Learning algorithms often rely on exploration-exploitation strate-
gies to balance between exploiting known information and exploring new actions.
Techniques like epsilon-greedy exploration and Boltzmann exploration are commonly
employed to address this trade-off.

[25] Li et al. introduced a novel Deep Reinforcement Learning Algorithm, incor-

porating advanced methodologies such as Trust Region Policy Optimization (TRPO)

6
and Proximal Policy Optimization (PPO), marks a significant step forward in traf-
fic management research. By integrating PPO with fairness criteria, a pioneering
Deep Reinforcement algorithm emerges, poised to revolutionize traffic control sys-
tems. This approach harnesses the power of Convolutional Neural Networks (CNN) to
extract both global and detailed traffic information, enhancing decision-making capa-
bilities. For simulation purposes, the widely-used Simulation of Urban Mobility[29]
(SUMO) platform provides a real-time simulation environment, enabling the test-
ing and validation of the proposed algorithm under various scenarios. Three primary
rules-based methods are adopted: fixed-time control, self-organizing traffic light con-
trol, and short-sighted traffic light control. Each method offers distinct advantages
and challenges, contributing to a comprehensive evaluation of the algorithm’s perfor-
mance across diverse traffic conditions. The integration of TRPO and PPO algorithms
not only enhances the learning process but also facilitates the incorporation of fair-
ness considerations, ensuring equitable traffic management outcomes. By leveraging
deep reinforcement learning techniques, the proposed algorithm adapts dynamically
to changing traffic patterns, optimizing signal timings in real-time to minimize con-
gestion and improve overall traffic flow efficiency. The utilization of CNN enables the
algorithm to process complex traffic data efficiently, capturing both local and global
traffic dynamics. This comprehensive approach holds promise for addressing the mul-
tifaceted challenges of modern urban traffic management, paving the way for more
responsive and sustainable transportation systems. Through rigorous simulation and
evaluation, the efficacy of the proposed algorithm in enhancing traffic efficiency and
fairness can be thoroughly assessed, potentially informing future advancements in
urban mobility strategies.
[30]Tan and Yang introduced an innovative strategy for orchestrating traffic sig-
nals within a connected vehicle (CV) ecosystem, with a primary focus on preserving
the privacy of individual travelers. Their approach aims to shield sensitive traveler
information, such as habitual destinations and travel patterns, from potential expo-
sure. To accomplish this, they devised a comprehensive defense mechanism to counter
various types of attacks, including collusion among CVs, database breaches, and inter-
ference attempts. Furthermore, they developed a refined traffic signal optimization
methodology, complemented by an arrival rate estimator leveraging aggregated CV
data. This multifaceted approach exhibits adaptability to diverse traffic scenarios,
ensuring effective traffic management across different conditions. By amalgamating
privacy preservation measures with signal control optimization, Tan and Yang’s frame-
work not only enhances traffic flow efficiency but also underscores the significance
of safeguarding CV data confidentiality, thereby fostering a safer and more secure
transportation ecosystem. [22] Zhao et al. introduces a groundbreaking model, the
Traffic Signal Control based on Heterogeneous Graph Attention Mechanism with Deep
Reinforcement Learning (TSC-HGAM-DRL) model, which merges deep reinforcement
learning with graph neural networks to enhance traffic management and decision-
making processes. This innovative model integrates a dual attention-based graph
neural network with deep reinforcement learning techniques, specifically combining
DDQN and DuelingDQN to mitigate potential overfitting issues during training. Sim-
ulation experiments were conducted using the SUMO platform in conjunction with

7
a deep learning development framework. The algorithm’s performance was rigorously
evaluated using four key metrics: Average Reward, Average Delay, Average Queue
Length, and Average Waiting Time. During peak traffic flow scenarios, the algorithm
demonstrated significant improvements over existing methods. Specifically, it achieved
an average latency reduction of 8.76, representing a 50% enhancement. The average
queue length decreased by 31.01 meters, marking a notable 48% improvement. More-
over, the average waiting time decreased to 8.71 seconds, showcasing a remarkable
120% enhancement. Additionally, the algorithm yielded a cumulative reward of 59.34,
reflecting a substantial 35% improvement compared to alternative algorithms.
These findings underscore the effectiveness of the TSC-HGAM-DRL model in opti-
mizing traffic signal control and minimizing congestion during peak traffic periods. By
integrating advanced techniques from deep reinforcement learning and graph neural
networks, the model offers a robust framework for addressing the complex dynamics
of urban traffic management. The significant improvements observed across multiple
evaluation metrics highlight the potential of this approach to enhance overall traffic
efficiency and reduce travel delays, contributing to more sustainable and resilient urban
transportation systems. When compared to four other traffic control models, this
model shows performance advantages of over 13% across multiple performance indica-
tors. However, further evaluation and validation are necessary, particularly for traffic
scenarios with varying complexities and characteristics. Additionally, the study did
not account for various potential influencing factors such as environmental conditions,
system failures, and uncertainties.
[21] Zhu et al. introduced an Intelligent Traffic Light Control (ITLC) system
aimed at mitigating traffic congestion. The system leveraged Multi-Agent Deep Rein-
forcement Learning [31](DRL) algorithms, which utilize deep neural networks for
decision-making. To enhance versatility, the study incorporated three distinct datasets.
Experimentation was conducted using the Cityflow simulator, with the primary goal
of alleviating traffic congestion by reducing both average travel time and total training
duration.
Throughout the study, comparisons were drawn among various traffic control
systems and reinforcement learning approaches. These included the fixed traffic
light control system, self-organizing traffic control system, Single-Agent Deep Rein-
forcement Learning, and Single-Agent Broad Reinforcement Learning methods. The
findings revealed notable advantages of the Multi-Agent Deep Reinforcement Learning
algorithm over the others.
Specifically, the training time for the ITLC system was found to be 35.03%, 27.51%,
and 18.22% faster compared to the fixed, self-organizing, and single-agent approaches,
respectively. Moreover, the stability of the Multi-Agent DRL algorithm significantly
surpassed that of its counterparts.
These results underscore the efficacy of the ITLC system in enhancing traffic man-
agement efficiency. By harnessing the capabilities of Multi-Agent DRL, the system
demonstrates promising potential for optimizing traffic flow and reducing congestion
in urban environments. Additionally, the utilization of diverse datasets enhances the
adaptability and robustness of the model, further contributing to its effectiveness in
addressing real-world traffic challenges.

8
 [20] Guo and et al. proposed the CoTV (Cooperative Control for Traffic Light
Signals and Connected Autonomous Vehicles) system to effectively manage both traf-
fic light signals and various Connected Autonomous Vehicles (CAVs). This system
employs Deep Reinforcement Learning (DRL) to optimize the balance between reduc-
ing travel time, fuel consumption, and emissions. As a multi-agent DRL system, CoTV
demonstrates notable improvements over the Baseline approach. In comparative anal-
ysis, CoTV achieved remarkable savings of up to 28% in fuel consumption and reduced
travel time by up to 30%. These findings suggest that CoTV represents a significant
advancement in traffic control strategies for both traffic signals and CAVs. Its abil-
ity to efficiently coordinate between these elements highlights its potential to enhance
overall traffic management efficiency while promoting sustainability and reducing envi-
ronmental impact. The CoTV system stands as a noteworthy contribution to the
field, offering insights into the optimization of traffic flow in the context of emerging
autonomous vehicle technologies.
[23] Kwok et al. introduced an innovative intelligent traffic light system designed
to address traffic congestion issues in Hong Kong. This system combines object detec-
tion technology with an evolutionary algorithm to optimize traffic flow. The study
utilized live video footage from Hong Kong to develop traffic models, which were then
employed to simulate traffic scenarios. Both the existing Fixed Cycle Traffic Light
(FCTL) system and the proposed intelligent system were implemented at intersections.
The simulation, based on estimates of pedestrian and vehicle numbers obtained
through object detection, allowed for dynamic switching of traffic signals in response
to varying traffic conditions. Results showed significant improvements in traffic effi-
ciency with the proposed system. During non-peak hours, the throughput of the roads
increased from 44% to approximately 60%. Moreover, during peak hours, the overall
efficiency saw an increase from 54% to 57%, as measured by the reduction in total
waiting time.
Notably, the implementation of the evolutionary algorithm led to a substantial
average reduction of 65% in total waiting time. These findings highlight the effec-
tiveness of the intelligent traffic light system in enhancing traffic flow and reducing
congestion in urban areas. By dynamically adjusting signal timings based on real-time
traffic data, the system demonstrates potential for improving overall transportation
efficiency and reducing travel delays.
[26] Liu et al. implemented traffic signal control using a teacher-student framework,
a reinforcement learning paradigm where a teacher guides a student in decision-
making. This framework is enhanced by incorporating a teaching policy, where the
teacher learns how to effectively guide the student, and a max update approach,
enabling the student to estimate the value function accurately. Initially, the teacher
is pre-trained in a specific traffic scenario and then applied to other traffic scenar-
ios. Meanwhile, the student, initialized with random parameters, is guided by the
pre-trained teacher. Notably, the policies of the student and teacher differ. The pro-
posed approach is compared with Double Deep Q-Learning, a traditional reinforcement
learning method, and demonstrates promising results. Across all four configurations
tested, the proposed approach achieved significant reductions in travel time ranging

9
from 15.73% to 27.91%. Moreover, the approach substantially decreased average wait-
ing time by 39.83% to 60.72% overall. By leveraging the teacher-student framework,
this approach offers a novel perspective on traffic signal control, showcasing improve-
ments in both travel time and waiting time. These findings underscore the potential
of reinforcement learning techniques in optimizing traffic flow and enhancing urban
mobility.
[27] Song et al. introduced a novel approach called the Multi-Agent Counterfactual
Actor-Critic with Scheduler framework, representing a modified reinforcement learn-
ing algorithm that facilitates information exchange among agents. This framework
surpassed traditional model-based methods and other existing reinforcement learn-
ing (RL) approaches. A central critic was employed to coordinate global information
using the global state, enhancing the model’s performance. The simulation study was
conducted on the actual road network of Shenzhen city, China, featuring nine intersec-
tions. The study focused on real-time traffic demand, categorized into low, medium,
and high demand periods based on actual traffic flow timing. Benchmark comparisons
were conducted against several algorithms, including Adaptive Signal Timer, MaxPres-
sure algorithm, classic RL-based Deep Q-Network (DQN) method, Multi-Agent Policy
Gradient method, Multi-Agent Actor-Critic method, and COMA-based method. This
innovative framework demonstrates significant advancements in traffic management
strategies, particularly in urban environments. By leveraging a multi-agent approach
and incorporating real-time traffic demand data, the model effectively optimizes traffic
signal control in complex urban road networks. The comparison against various bench-
marks showcases the superiority of the proposed framework, highlighting its potential
to enhance traffic flow efficiency and reduce congestion in urban areas. Additionally,
the utilization of real road network data from Shenzhen city adds credibility and rele-
vance to the findings, emphasizing the practical applicability of the proposed approach
in real-world scenarios.
[28]Wang et al. introduced a cooperative group-based multi-agent reinforcement
learning (MARL) algorithm designed to optimize traffic signal control. Their approach
leverages cooperative vehicle-infrastructure systems to efficiently manage large-scale
road networks. Through extensive simulation, the researchers explored multiple rein-
forcement learning methods and found that Deep Q-Network (DQN) in combination
with MARL yielded promising results. The algorithm was trained and evaluated on
a 6x6 network using SUMO simulation software to replicate real-world scenarios. In
their study, real-time traffic control adjustments were made every 20 seconds, with
congestion levels determined by the number of waiting vehicles. Three congestion
thresholds were considered: less than 5 vehicles, 5-15 vehicles, and above 15 vehicles.
The proposed algorithm outperformed alternative strategies significantly, achieving a
78.84% improvement over a fixed approach, a 95.69% improvement over the CGL-ILQ
strategy, and an 86.64% improvement over the CGB-MA2C strategy. This coopera-
tive MARL approach demonstrates considerable potential in enhancing traffic signal
control systems, particularly in large-scale urban environments. By leveraging cooper-
ative interactions between vehicles and infrastructure, the algorithm effectively reduces
congestion levels and improves traffic flow efficiency. Moreover, the rigorous evalua-
tion against various benchmarks underscores the robustness and effectiveness of the

10
proposed approach. Overall, Wang et al.’s research contributes valuable insights into
the development of advanced traffic management strategies aimed at addressing the
challenges of modern urban mobility.
[19]Essa and Sayed introduced a reinforcement learning algorithm implemented
on the VISSIM simulation platform, which was subsequently tested using real-world
traffic data from Surrey, British Columbia. The VISSIM model, calibrated for each
intersection, underwent extensive testing over a 9-hour period from 9:00 am to 6:00 pm.
Through this testing, it was observed that the maximum safety benefit was achieved
when the Minimum Predicted Risk (MPR) of connected vehicles (CV) reached 50%,
resulting in a remarkable 90% safety benefit. This suggests that the summarized algo-
rithm could prove highly effective, particularly when the MPR of CVs is equal to or
greater than 30%. Additionally, the study highlights that higher MPR values corre-
spond to increased effectiveness of the safety algorithm. This research underscores the
potential of reinforcement learning techniques in optimizing traffic safety, particularly
in urban environments, and provides valuable insights into the importance of MPR in
enhancing the efficacy of such algorithms.

2.5 Neural Networks

Neural networks[32] are computational models inspired by the structure and func-
tion of biological neural networks in the human brain. These algorithms consist of
interconnected nodes, or neurons, organized in layers. Each neuron applies a trans-
formation to its input and passes the result to neurons in the next layer.

Feedforward Neural Networks (FNN): FNNs are the simplest type of neural
network, where information flows in one direction—from input to output. They consist
of input, hidden, and output layers, with connections between layers but no loops.
Convolutional Neural Networks (CNN): CNNs are specialized for processing
grid-like data, such as images. They use convolutional layers to extract features from
input data and pooling layers to reduce dimensionality. CNNs are widely used in image
classification, object detection, and image segmentation tasks.
Recurrent Neural Networks (RNN): RNNs[33][34] are designed to handle
sequential data by maintaining a memory of previous inputs. They have connections
that form directed cycles, allowing them to capture temporal dependencies. RNNs
are commonly used in tasks such as language modeling, speech recognition, and time
series prediction.
Long Short-Term Memory (LSTM): LSTM networks are a type of RNN
designed to address the vanishing gradient problem, which can occur when training
traditional RNNs on long sequences. LSTMs have a more complex architecture with
gated cells that allow them to retain information over longer time periods.
Generative Adversarial Networks (GAN): GANs consist of two neural net-
works—the generator and the discriminator—that are trained simultaneously. The
generator learns to generate data samples that are indistinguishable from real data,
while the discriminator learns to distinguish between real and generated samples.
GANs are used for generating realistic images, text, and other types of data.

11
 Autoencoders: Autoencoders are a type of neural network used for unsupervised
learning. They consist of an encoder network that compresses the input data into a
low-dimensional representation (encoding), and a decoder network that reconstructs
the original input from the encoding. Autoencoders are used for tasks such as data
denoising, dimensionality reduction, and anomaly detection.
These neural network algorithms form the backbone of many modern machine
learning and artificial intelligence applications and have been instrumental in advanc-
ing various fields, including computer vision, natural language processing, and
reinforcement learning.These are also the most used algorithms in the field of smart
city and traffic signal controlling.[35][36][37]
[38] Espinoza et al. proposed a novel method for traffic flow prediction, employing both
Machine Learning (ML) and Deep Learning (DL) techniques. The study utilized two
publicly available datasets for training, validation, and testing purposes. Recurrent
Neural Networks (RNNs), including Gated Recurrent Units (GRU) and Long Short-
Term Memory (LSTM) networks, were developed, with 75% of the dataset allocated
for training and the remaining 25% for testing.
The research explored the potential application of these models in smart traf-
fic light control systems. Among the models investigated, the Multilayer Perceptron
(MLP) demonstrated superior performance, achieving the highest accuracy while also
requiring the least processing time. This indicates its suitability for real-time traffic
flow prediction and integration into intelligent transportation systems.
By leveraging ML and DL technologies, this study contributes to the advance-
ment of traffic management strategies. The predictive capabilities of neural networks,
such as GRU, LSTM, and MLP, offer valuable insights for optimizing traffic con-
trol mechanisms and enhancing overall transportation efficiency. These findings drive
progress towards more adaptive and responsive traffic management solutions, essential
for mitigating congestion and improving urban mobility.
[39] Abdullah et al. proposed a Soft Gated Recurrent Unit (GRU) based Recurrent
Neural Network (RNN) model to predict traffic congestion, thereby optimizing traf-
fic flow in smart cities. Their model leverages deep learning (DL) technology, which
has demonstrated considerable success in traffic flow optimization through intelligent
traffic signal control and traffic flow prediction.
The study introduces a novel approach wherein network traffic monitoring is
implemented using a Ryu controller within a software-defined networking (SDN) envi-
ronment. The primary objective is to enhance the accuracy and efficiency of network
traffic monitoring, offering real-time insights into network performance.
The proposed model focuses on three key aspects:
Utilizing an edge-based vehicular environment to predict road traffic, where edge
servers store both historical and real-time data related to various factors such as social
media activity, weather conditions, road traffic, and road conditions.
Extracting multiple features using a DL architecture, specifically Bidirectional
RNN (BRNN) with Soft GRU, and classifying the data into two categories: congested
or uncongested traffic.
Proposing an optimization approach to fine-tune the hyperparameters of the DL
architecture based on real-time and historical traffic data.

12
 Table 3 Neural Network Based work summary

Paper Dataset Model(s) Used Simulation Performance Met-

Method rics
[38]Espinoza et Huawei Munich Recurrent Neural - Traffic Flow Pre-
al. Research Center Network(RNN), diction
Dataset(Custom) Multilayer Per-
ceptron Neural
Network(MLP-
NN)
[39]Abdullah et INRIX Dataset Recurrent Neural SUMO,OMNeT Traffic Conges-
al. Networks(RNN) tion
[21]Zhu et al. G-X Dataset, Deep Reinforce- Cityflow Simula- Avg. Travel Time
Jinan Dataset, ment Learning tor
Hangzhou
Dataset
[43]Majumdar et Private Dataset Long Term Short None Traffic Conges-
al. Memory(LTSM) tion

The model implementation involved sampling traffic flow at 3-minute intervals,

with the further enhancement of system capability and dependability through the
use of Vehicular Ad-Hoc Networks (VANET)[40][41][42]. Overall, the system demon-
strated promising results, with a gradual decrease in computation time during traffic
flow prediction, indicating its potential for real-world deployment in optimizing urban
traffic management.
[43] Majumdar et al. employed an innovative approach to forecast traffic conges-
tion in smart cities, utilizing a combination of Internet of Things (IoT) technology
and machine learning methodologies. Their model aimed to predict the spread of con-
gestion over a 5-minute interval within a bustling urban area. The dataset utilized
in the study was custom-made, gathered from Buxton, UK, by monitoring real-time
traffic scenarios. Vehicle speed data was sourced from Matlock County Hall, collected
consistently from 7 am to 7 pm over the course of a week. The study focused on
employing Long Short-Term Memory (LSTM)[44][45] networks, a type of Recurrent
Neural Network (RNN), known for their ability to handle long sequences of data
effectively. LSTM models were chosen over traditional RNNs due to their superior
performance in processing data over extended time intervals. The data underwent
processing through a 100-layer stack of LSTM cells before transitioning to a hidden
layer for further analysis. The dataset was partitioned into a 75/25 split for training
and testing, respectively, ensuring robust model evaluation. Utilizing a time step of 5
minutes, the model predicted congestion levels for the subsequent 5-minute window
based on current observations. Impressively, the model achieved an accuracy rate of
94.6%, indicating its efficacy in forecasting traffic congestion in urban environments.
This study underscores the potential of IoT-enabled machine learning approaches in
enhancing traffic management and urban planning efforts, offering valuable insights
for addressing congestion-related challenges in smart cities.
So after covering all the part about Neural Networks we move on to the next
category.

13
2.6 Support Vector Machine
Support Vector Machine (SVM)[46][47][48] is a powerful and widely used supervised
learning algorithm for classification and regression tasks. Introduced by Vapnik and
Cortes in the 1990s, SVM has gained popularity due to its effectiveness in handling
both linearly separable and non-linearly separable datasets. One of the key features of
SVM is its ability to handle high-dimensional data efficiently, making it suitable for
tasks with a large number of features. SVM achieves this by transforming the input
data into a higher-dimensional space using a kernel function, which allows it to find
complex decision boundaries that may not be linear in the original feature space. The
optimization objective of SVM is to find the hyperplane that not only separates the
data points of different classes but also maximizes the margin between them. This is
formulated as a convex optimization problem, which can be solved using various opti-
mization techniques such as gradient descent or quadratic programming. In addition
to linear classification, SVM can also be extended to handle non-linear classification
tasks by using different kernel functions, such as polynomial, radial basis function
(RBF), or sigmoid kernels. These kernels allow SVM to capture complex patterns in
the data and achieve higher classification accuracy. SVM has several advantages that
make it a popular choice for various machine learning tasks. One of its main strengths
is its ability to generalize well to unseen data, which helps prevent overfitting and
improves the model’s performance on test data. SVM also has a solid theoretical
foundation, with strong mathematical principles supporting its effectiveness. Despite
its strengths, SVM also has some limitations. One notable drawback is its com-
putational complexity, particularly when dealing with large datasets or non-linear
kernels, which can lead to longer training times and higher memory requirements.
SVM also does not provide probabilistic outputs directly, although techniques such
as Platt scaling can be used to estimate class probabilities. Overall, Support Vector
Machine is a versatile and powerful machine learning algorithm that is widely used in
various fields, including pattern recognition, image classification, text classification,
and bioinformatics. Its ability to handle both linear and non-linear classification
tasks, along with its robustness and generalization capabilities, make it a valuable
tool in the machine learning toolkit[49][50]. SVM can handle linearly separable
data as well as nonlinear data by using kernel functions, which map the input data
into a higher-dimensional space where it can be linearly separated. Common kernel
functions include linear, polynomial, radial basis function (RBF), and sigmoid. SVM
is robust to overfitting, especially in high-dimensional spaces, and can handle large
datasets efficiently. It is widely used in various domains such as image classification,
text classification, bioinformatics, and finance.

[51] Anjaneyulu et al. introduced a novel approach using a hybrid Xception Sup-
port Vector Machine (XPSVM) to predict short-term traffic congestion. This method
aimed to forecast Short-Term Traffic Congestion (STTC) every 5 minutes over a span
of 1 hour, with the potential to reduce travel costs, travel time, and air pollution.
The Xception technique, an enhancement of the Inception technique utilized in Con-
volutional Neural Networks (CNN), was employed. This model features a 71-layer
deep CNN architecture inspired by Google. To construct the dataset, traffic images

14
 Table 4 Support Vector Machine Summary

Paper Dataset Model(s) Simulation Performance Performance

Used Method Metrics
[51]Anjaneyulu Private Hybrid NGSIM Traffic Con- 97.16% Accu-
et al. Dataset Xception Sup- gestion racy
port Vector Prediction
Machine
[47]Devi et Private Support Vec- None Speed Arrival -
al. Dataset tor Machine, Timing,
Linear and Density of
Logistic individual
Regression vehicles
[50]Wang et Private Support Vec- Private Prediction of -
al. Dataset tor Machine Traffic Flow

were gathered from various locations in Bangalore, Karnataka. Upon implementa-

tion, the model demonstrated remarkable performance, achieving a high accuracy of
97.16% in predicting short-term traffic congestion. This innovative approach leverages
advanced deep learning techniques to address critical transportation challenges. By
accurately predicting short-term traffic congestion, it has the potential to significantly
improve traffic management efficiency, ultimately leading to reduced travel times and
environmental benefits.
[47] Devi et al. proposed an innovative strategy leveraging support vector machine
(SVM) alongside regression methodologies, including linear and logistic regression, to
assess peak traffic accuracy. Their dataset, tailored for smart city applications, encom-
passed a wide array of parameters such as traffic and weather conditions, parking
availability, accident-prone zones, and popular on-demand destinations. Additionally,
the dataset was organized into three-dimensional structures to facilitate comprehen-
sive analysis. Applying SVM to the traffic dataset entailed rigorous statistical analysis
aimed at uncovering patterns indicative of heavy traffic scenarios. This approach syner-
gized SVM’s robust classification capabilities with the predictive prowess of regression
analysis, enabling a nuanced comprehension of peak traffic dynamics. By harnessing
diverse data dimensions and sophisticated analytical techniques, Devi et al. made sig-
nificant strides in enhancing traffic management strategies in urban locales, thereby
paving the way for more responsive and efficient transportation systems in smart cities.

3 Conclusion
In summary, this survey paper offers a comprehensive exploration of diverse traf-
fic signal management techniques prevalent in contemporary transportation systems.
Throughout our analysis, we’ve delved into a spectrum of methodologies, encom-
passing traditional practices, heuristic algorithms, and cutting-edge technologies like
machine learning and deep reinforcement learning. Our findings underscore the con-
tinued relevance of conventional methods such as fixed-time control and actuated
control, alongside the burgeoning adoption of adaptive control and reinforcement
learning-based approaches. These newer techniques exhibit promise in their capacity to
dynamically adjust signal timings, responding adeptly to fluctuating traffic conditions
in real-time.

15
 Furthermore, our study also considers multifarious factors like traffic volume,
vehicular composition, pedestrian patterns, and environmental variables in devising
and implementing traffic signal control strategies. The integration of emerging tech-
nologies such as IoT and machine learning emerges as pivotal for enhancing system
efficiency and safety. In essence, this survey paper serves as an invaluable resource for
researchers, transportation planners, and policymakers alike, offering nuanced insights
into the evolving terrain of traffic signal management methodologies. By discerning
the strengths and limitations of diverse approaches, stakeholders are empowered to
make informed decisions aimed at optimizing traffic flow, alleviating congestion, and
augmenting overall transportation system efficacy in urban landscapes.

Declarations: On behalf of all authors, the corresponding author states that

there is no conflict of interest.

References
[1] Smith, J., Johnson, E.: Traffic congestion: Causes, effects, and mitigation
strategies. Transportation Research Part A: Policy and Practice 134, 123–145
(2020)

[2] Chen, W., Li, X., Zhang, W.: Traffic congestion management: Recent advances
and future directions. Transportation Research Part C: Emerging Technologies
129, 102–123 (2021)

[3] Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detec-
tion. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition (2020)

[4] Zhou, X., Wang, D., Krähenbühl, P.: Objects as points. In: Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019)

[5] Cai, Z., Vasconcelos, N.: Cascade r-cnn: Delving into high quality object detec-
tion. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(2),
315–328 (2021)

[6] Bochkovskiy, A., Wang, C.-Y., Liao, H.-Y., Možeiko, V., Fidler, S.: Yolov4: Opti-
mal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934
(2020)

[7] Liu, H., Zhang, Z., Wang, C.: Intelligent traffic signal management: Challenges
and opportunities. IEEE Transactions on Intelligent Transportation Systems
22(4), 2394–2406 (2021)

[8] Jin, J., Jin, J.: Traffic congestion and air pollution: Empirical evidence
before/after covid-19 in seoul, korea. International Journal of Sustainable

16
 Transportation 17(12), 1356–1369 (2023) https://doi.org/10.1080/15568318.
2023.2192688

[9] Kwon, D., Paul, K.C., Yu, Y., Zhang, K., Folle, A.D., Wu, J., Bronstein, J.M.,
Ritz, B.: Traffic-related air pollution and parkinson’s disease in central california.
Environmental Research 240, 117434 (2024) https://doi.org/10.1016/j.envres.
2023.117434

[10] Singh, S., Singh, J., Goyal, S.B., Singh Sehra, S., Ali, F., Ayad Alkhafaji, M.,
Singh, R.: A novel framework to avoid traffic congestion and air pollution for
sustainable development of smart cities. Sustainable Energy Technologies and
Assessments 56, 103125 (2023) https://doi.org/10.1016/j.seta.2023.103125

[11] Kehlbacher, A., Stark, K., Gebhardt, L., Jarass, J., Schuppan, J.: Comparing
municipal progress in implementing temporary cycle lanes during the covid-19
pandemic. Transportation Research Part A: Policy and Practice 174, 103752
(2023) https://doi.org/10.1016/j.tra.2023.103752

[12] Almallah, M., Alhajyaseen, W.K.M., Dias, C.: Safety assessment of on-road
cycling lanes: A comparative study of different layouts using driving simulator.
Accident Analysis and Prevention 196, 107431 (2024) https://doi.org/10.1016/j.
aap.2023.107431

[13] Yazdani, M., Sarvi, M., Asadi Bagloee, S., Nassir, N., Price, J., Parineh, H.:
Intelligent vehicle pedestrian light (ivpl): A deep reinforcement learning approach
for traffic signal control. Transportation Research Part C: Emerging Technologies
149, 103991 (2023) https://doi.org/10.1016/j.trc.2022.103991

[14] Wei, S., Zeng, A., Qiu, M., Guan, H.: Adaptive traffic signal control with deep
reinforcement learning. IEEE Transactions on Intelligent Transportation Systems
21(9), 3866–3879 (2020)

[15] Shao, J., Chen, G.: Reinforcement learning based adaptive traffic signal control
with congestion and emergency handling. IEEE Access 8, 11449–11461 (2020)

[16] Yang, J., Wu, J., Lv, Z., Zhang, K.: Deep reinforcement learning for adaptive
traffic signal control: Methodology and performance comparison. IEEE Access 8,
89572–89584 (2020)

[17] Dai, X., Lu, Z., Sun, L.: Traffic signal control based on multi-agent deep
reinforcement learning. Journal of Advanced Transportation 2020, 1–9 (2020)

[18] Genders, W., Clercq, S.D., Brussel, H.V., Vanden Berghe, G.: Optimizing traffic
signal control through deep reinforcement learning. In: 2020 23rd International
Conference on Intelligent Transportation Systems (ITSC), pp. 1–6 (2020). IEEE

[19] Essa, M., Sayed, T.: Self-learning adaptive traffic signal control for real-time safety

17
 optimization. Accident Analysis Prevention 146, 105713 (2020) https://doi.org/
10.1016/j.aap.2020.105713

[20] Guo, J., Cheng, L., Wang, S.: Cotv: Cooperative control for traffic light sig-
nals and connected autonomous vehicles using deep reinforcement learning. IEEE
Transactions on Intelligent Transportation Systems 24(10), 10501–10512 (2023)
https://doi.org/10.1109/TITS.2023.3276416

[21] Zhu, R., Li, L., Wu, S., Lv, P., Li, Y., Xu, M.: Multi-agent broad reinforcement
learning for intelligent traffic light control. Information Sciences 619, 509–525
(2023) https://doi.org/10.1016/j.ins.2022.11.062

[22] Zhao, Z., Wang, K., Wang, Y., Liang, X.: Enhancing traffic signal control with
composite deep intelligence. Expert Systems with Applications 244, 123020
(2024) https://doi.org/10.1016/j.eswa.2023.123020

[23] Ng, S.-C., Kwok, C.-P.: An intelligent traffic light system using object detec-
tion and evolutionary algorithm for alleviating traffic congestion in hong kong.
International Journal of Computational Intelligence Systems 13, 802–809 (2020)
https://doi.org/10.2991/ijcis.d.200522.001

[24] Pygame Contributors: Pygame. https://www.pygame.org (1999–)

[25] Li, C., Ma, X., Xia, L., Zhao, Q., Yang, J.: Fairness control of traffic light via
deep reinforcement learning. In: 2020 IEEE 16th International Conference on
Automation Science and Engineering (CASE), pp. 652–658 (2020). https://doi.
org/10.1109/CASE48305.2020.9216899

[26] Liu, J., Qin, S., Su, M., Luo, Y., Zhang, S., Wang, Y., Yang, S.: Traffic signal
control using reinforcement learning based on the teacher-student framework.
Expert Systems with Applications 228, 120458 (2023) https://doi.org/10.1016/
j.eswa.2023.120458

[27] Song, X.B., Zhou, B., Ma, D.: Cooperative traffic signal control through a coun-
terfactual multi-agent deep actor critic approach. Transportation Research Part
C: Emerging Technologies 160, 104528 (2024) https://doi.org/10.1016/j.trc.2024.
104528

[28] Wang, T., Cao, J., Hussain, A.: Adaptive traffic signal control for large-scale
scenario with cooperative group-based multi-agent reinforcement learning. Trans-
portation Research Part C: Emerging Technologies 125, 103046 (2021) https:
//doi.org/10.1016/j.trc.2021.103046

[29] Krajzewicz, D., Erdmann, J., Behrisch, M., Bieker, L.: Sumo - simulation of urban
mobility: An overview. SIMULATION 88(9), 855–867 (2012)

[30] Tan, C., Yang, K.: Privacy-preserving adaptive traffic signal control in a connected

18
 vehicle environment. Transportation Research Part C: Emerging Technologies
158, 104453 (2024) https://doi.org/10.1016/j.trc.2023.104453

[31] Zhang, Z., Li, X., Wu, J., Gao, J.: Cooperative multi-agent deep reinforcement
learning with dynamic information exchange for traffic signal control. IEEE
Transactions on Intelligent Transportation Systems (2022)

[32] Haykin, S.: Neural Networks and Learning Machines. Pearson Education, ???
(2009)

[33] Li, X., Wu, J., Gao, J., Wang, J., Wu, Y.: Adaptive traffic signal control based
on recurrent neural networks with time-varying coefficients. IEEE Transactions
on Intelligent Transportation Systems (2021)

[34] Chen, Y., Zhang, Z., Li, X., Wu, J.: Traffic signal control using recurrent
neural networks with reinforcement learning. IEEE Transactions on Intelligent
Transportation Systems 23(4), 2036–2045 (2022)

[35] Wang, C., Zegras, C.: Traffic signal control using convolutional neural networks.
Transportation Research Part C: Emerging Technologies 115, 102615 (2020)

[36] Vural, S., Li, X.: Real-time traffic signal control using recurrent neural networks.
Transportation Research Part C: Emerging Technologies 103, 152–167 (2019)

[37] Bessa, W.: Traffic signal control using echo state neural networks. Transportation
Research Part C: Emerging Technologies 119, 102810 (2020)

[38] Navarro-Espinoza, A., López-Bonilla, O.R., Garcı́a-Guerrero, E.E., Tlelo-

Cuautle, E., López-Mancilla, D., Hernández-Mejı́a, C., Inzunza-González, E.:
Traffic flow prediction for smart traffic lights using machine learning algorithms.
Technologies 10(1) (2022) https://doi.org/10.3390/technologies10010005

[39] Abdullah, S.M., Periyasamy, M., Kamaludeen, N.A., Towfek, S.K., Marappan,
R., Kidambi Raju, S., Alharbi, A.H., Khafaga, D.S.: Optimizing traffic flow in
smart cities: Soft gru-based recurrent neural networks for enhanced congestion
prediction using deep learning. Sustainability 15(7) (2023) https://doi.org/10.
3390/su15075949

[40] Hadi, F., Siddiqui, S., Rashid, S., Memon, A.G., Raza, M.A.: Vehicular ad-
hoc network based intelligent traffic signal system for smart cities. International
Journal of Distributed Sensor Networks 16(2), 1550147720909240 (2020)

[41] Tong, X., Gu, X., Xu, F., Han, D.: Vehicular ad-hoc networks (vanets): State-of-
the-art and future perspectives. Journal of Computer and Communications 8(8),
1–18 (2020)

[42] Li, Z., Guan, H., Zhang, Z., Lv, Z., Yang, J.: Vehicular ad-hoc networks (vanets):

19
 Current challenges and future directions. IEEE Access 9, 30610–30622 (2021)

[43] Majumdar, S., Subhani, M.M., Roullier, B., Anjum, A., Zhu, R.: Congestion
prediction for smart sustainable cities using iot and machine learning approaches.
Sustainable Cities and Society 64, 102500 (2021) https://doi.org/10.1016/j.scs.
2020.102500

[44] Li, X., Zhang, Z., Wu, J., Gao, J.: Adaptive traffic signal control using long
short-term memory networks with dynamic time warping. IEEE Transactions on
Intelligent Transportation Systems (2022)

[45] Wang, C., Zhang, Z., Li, X.: Real-time traffic signal control using long short-term
memory networks and deep reinforcement learning. Transportation Research Part
C: Emerging Technologies 135, 103430 (2022)

[46] Li, L., Song, M.: Support vector machine-based traffic signal optimization under
stochastic demand. Transportation Research Part C: Emerging Technologies 112,
102642 (2020)

[47] Devi, T., Alice, K., Deepa, N.: Traffic management in smart cities using support
vector machine for predicting the accuracy during peak traffic conditions. Mate-
rials Today: Proceedings 62, 4980–4984 (2022) https://doi.org/10.1016/j.matpr.
2022.03.722 . International Conference on Innovative Technology for Sustainable
Development

[48] Tan, L., Zhu, Y.: Support vector machine for traffic signal control with variable
speed limits. Transportation Research Part C: Emerging Technologies 111, 142–
157 (2020)

[49] Shen, B., Gu, X., Xu, F., Wang, S.: Adaptive traffic signal control based on
support vector machine with statistical pattern recognition. Journal of Advanced
Transportation 2020, 1–12 (2020)

[50] Yao, Y., He, Z., Xie, X.: Traffic signal control based on support vector machine
optimized by bat algorithm. Computational Intelligence and Neuroscience 2019,
1–10 (2019)

[51] Anjaneyulu, M., Kubendiran, M.: Short-term traffic congestion prediction using
hybrid deep learning technique. Sustainability 15(1) (2023) https://doi.org/10.
3390/su15010074

20