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

net/publication/385377436

Adaptive Traffic Control Framework for Urban Intersections

Conference Paper · October 2024

CITATIONS READS

0 248

1 author:

Isuru Lakmal
NSBM Green University
54 PUBLICATIONS 80 CITATIONS

SEE PROFILE

All content following this page was uploaded by Isuru Lakmal on 30 October 2024.

The user has requested enhancement of the downloaded file.

Adaptive Traffic Control Framework for Urban
Intersections
K.K.S.S.Wijayarathna H.K.I.S.Lakmal
Dept. of Electrical, Electronic and Systems Dept. of Mechatronic and Industrial Engineering
Engineering NSBM Green University Sri Lanka.
NSBM Green University [email protected]
Sri Lanka.
[email protected]

Abstract — Traffic congestion in Colombo, Existing approaches in Colombo have not

particularly at heavily used intersections, has been a sufficiently addressed the need for real-time
persistent challenge for decades, largely due to the adaptability in traffic management. Although some
increasing number of vehicles and the limitations of intersections are monitored by CCTV cameras, these
current traffic management systems. The existing fixed- systems are primarily for recording footage rather than
time traffic signals and manual interventions fail to adapt active traffic management. Moreover, the high costs
to the complex and evolving traffic patterns, leading to associated with installing multiple cameras to monitor
extended vehicle queues, prolonged waiting times, and a single intersection make widespread deployment
significant commuter dissatisfaction. This research
impractical, especially in a developing country like Sri
addresses these issues by proposing an adaptive traffic
signal control framework for a major intersection in
Lanka during economic challenges. This research seeks
Colombo. The system utilizes real-time data collection to fill this gap by developing an adaptive traffic signal
from a CCTV camera installed near the intersection to control system specifically tailored to a major
monitor traffic patterns continuously. An LSTM intersection in Colombo, proposing a cost-effective and
network is employed to predict traffic volumes based on scalable framework for implementation across Sri
the identified patterns, while a Deep Deterministic Policy Lanka.
Gradient (DDPG)-based reinforcement learning
The proposed solution integrates real-time data
algorithm is used to optimize traffic light timings. The
collection, advanced neural networks, and
proposed solution aims to maximize vehicle throughput,
reduce waiting times, and minimize vehicle queues, while
reinforcement learning algorithms to dynamically
also being easily integrated into existing infrastructure. optimize traffic signal timings. A CCTV camera
The framework is designed to be both economically installed at the intersection continuously monitors
feasible, requiring reduced processing power and traffic patterns, with the data feeding into a Long Short-
offering a robust solution for urban traffic management. Term Memory (LSTM) network designed to predict
traffic volumes based on both historical and real-time
Keywords - Adaptive traffic control, Reinforcement inputs. These predictions are used by a Deep
Learning (RL), Deep Neural Networks Deterministic Policy Gradient (DDPG)-based
reinforcement learning model to adjust traffic light
I. INTRODUCTION sequences, maximizing vehicle throughput, reducing
Traffic congestion is an escalating problem in urban waiting times, and minimizing queues. A key feature of
areas worldwide, adversely impacting daily life, this approach is the ability to test the framework in a
economic productivity, and environmental SUMO simulation environment before real-world
sustainability. As urban populations grow and vehicle implementation which will be able to cut-off
numbers increase, traffic bottlenecks become more unexpected cost that have to be bared with prototypes
frequent and severe. Traditional traffic management development.
systems, which rely on static, fixed-time traffic signals, The novelty of this approach lies in its adaptability
are often inadequate for addressing the dynamic and to real-time conditions and its economic feasibility for
complex traffic patterns observed in modern urban a developing country like Sri Lanka. The system is
environments. designed to require minimal infrastructure and
In Colombo, Sri Lanka’s capital, traffic congestion processing power, making it cost-effective and easily
has long been a persistent issue, particularly at key integral into existing traffic management frameworks.
urban intersections that are vital nodes in the city’s By addressing the limitations of current traffic
transportation network. These intersections, where management strategies and offering a scalable, cost-
vehicles from multiple directions converge, are often effective solution, this research has the potential to
the most congested points, leading to complex traffic significantly improve traffic flow and reduce
scenarios and longer waiting times. The current traffic congestion in Colombo’s urban environment.
control system in Colombo, which combines fixed- This study aims to establish a robust framework for
time signal controllers with manual interventions, is adaptive traffic signal control that can be implemented
inadequate for handling the growing traffic volumes in other intersections across Colombo and similar urban
and complex flows, necessitating frequent manual areas facing congestion challenges. Expected outcomes
adjustments during peak hours. This inefficiency include enhanced traffic management efficiency,
underscores the need for an adaptive traffic control reduced environmental impact, and improved
system that can intelligently adjust signal timings in commuter satisfaction, contributing to sustainable
real-time based on actual traffic conditions. urban development.
 II. RELATED WORK complex urban environments. Traditional traffic signal
Traffic signal control has undergone significant control systems, such as Webster’s Formula and
evolution over the years, driven by the growing need to SCATS, relied on static models and struggled with
manage increasing congestion and optimize traffic flow unpredictable traffic patterns. RL provides a dynamic,
in urban environments. This section reviews various data-driven alternative, allowing systems to learn
approaches to traffic signal control, focusing on static optimal strategies by interacting with their
systems, adaptive methods, and the application of environment in real time. The foundational framework
reinforcement learning (RL) techniques. of RL is rooted in the Markov Decision Process
(MDP), where agents learn by taking actions based on
observed states and maximizing cumulative rewards.
A. Static Traffic Signal Control Q-learning is a popular RL technique for estimating
Static traffic signal control remains prevalent in long-term value, and the integration of Deep
many regions, including Colombo, Sri Lanka. In these Reinforcement Learning (DRL) has further enhanced
systems, traffic signal parameters—such as phase its capabilities. DRL enables traffic control systems to
durations and cycle times—are predetermined based on handle complex environments and process high-
historical traffic data. While effective under stable dimensional data, such as real-time sensor feeds,
traffic conditions, but struggle with dynamic and without manual feature extraction.
unpredictable scenarios, such as adverse weather, high
traffic conditions. These limitations often lead to RL-based methods have shown great potential in
inefficiencies, as the fixed timings cannot adapt to improving adaptive traffic control by dynamically
changing traffic volumes and conditions. adjusting to real-time traffic conditions. [11]
demonstrated how Q-learning could outperform
traditional methods by learning from real-time traffic
B. Adaptive Traffic Signal Control patterns at isolated intersections, with the potential for
To address the limitations of static traffic systems, expansion to networked systems. Similarly, [12]
adaptive traffic signal control has emerged as a surveyed the use of multi-agent reinforcement learning
promising solution. Various adaptive methods have (MARL), where intersections act as independent
been proposed: agents that cooperate to optimize traffic across urban
Genetic Algorithms (GAs): GAs optimize traffic signal networks. This decentralized approach is crucial for
settings by treating phases and timings as managing interconnected urban intersections.
chromosomes, aiming to minimize queue lengths and
travel times through a fitness function [3][4]. However, DRL's ability to process complex data, such as visual
GAs can be computationally intensive and slow to inputs from traffic cameras, allows adaptive systems
converge. to adjust signal timings based on traffic patterns. Real-
world implementations, such as the Surtrac system in
Fuzzy Logic: This method uses linguistic variables and
Pittsburgh, have shown success in reducing wait times
fuzzy inference rules to model uncertainty and
determine signal timings based on imprecise data and emissions. Additionally, [11] reported success in
[5][6]. While useful for handling uncertainty, fuzzy applying RL to an isolated traffic signal in Toronto,
logic systems can become complex and less adaptable with real-time adaptation resulting in improved traffic
to highly dynamic traffic conditions. flow. [12] emphasized the importance of sophisticated
simulation environments in validating RL models
Reinforcement Learning (RL): Q-learning, a popular before deployment, ensuring scalability and
RL approach, learns optimal signal timings based on adaptability across various traffic conditions.
vehicle queues and actions that adjust green times [7].
Despite its advantages, Q-learning is limited by its
discrete action space, restricting its ability to manage
continuous traffic variations effectively. D. DDPG Reinforcement Learning Approach
Swarm Intelligence: Techniques like Ant Colony The Deep Deterministic Policy Gradient
Optimization (ACO) mimic social insect behaviour to (DDPG) algorithm offers significant advantages for
solve combinatorial problems in traffic control [1]. real-time traffic signal control, as highlighted in Wu's
While effective in exploring complex search spaces, research. One of the key strengths of DDPG is its
ACO may struggle with real-time applications due to capability to handle continuous action spaces, which is
high computational demands. essential for making fine-tuned decisions regarding
traffic signal timings (e.g., adjusting green, yellow,
and red light durations). In contrast, traditional
C. Advancements in Reinforcement Learning for reinforcement learning (RL) algorithms, such as Q-
Adaptive Traffic Control learning and Deep Q-Networks (DQN), are limited to
Recent advancements in adaptive traffic control, discrete action spaces, which can hinder their
particularly through reinforcement learning (RL), have effectiveness in complex traffic scenarios.
significantly improved vehicle flow optimization in
 DDPG's actor-critic architecture allows it to temporal dependencies in traffic data, resulting in
operate in high-dimensional state and action spaces, superior forecasting accuracy compared to traditional
effectively processing multi-dimensional inputs like methods like random walk and support vector
vehicle counts, waiting times, and road occupancy. regression. The unique architecture of LSTMs, which
This enables the development of sophisticated policies incorporates memory cells and gating mechanisms,
that can adapt to various traffic patterns, unlike simpler allows them to capture intricate traffic patterns over
RL algorithms that may struggle in such environments. time.

Wu's research also indicates that DDPG Additionally, LSTM implementation can lead
achieves faster convergence and more stable to significant cost savings by reducing the need for
performance compared to DQN and Normalized multiple camera installations at intersections. Instead,
Advantage Function (NAF) algorithms in simulations. a single camera, combined with date and time data, can
The use of experience replay in DDPG further be used to estimate traffic volumes on other roads. This
enhances learning efficiency, allowing the agent to approach not only lowers installation and maintenance
leverage past experiences for improved policy costs but also simplifies data collection, making it an
robustness, which is crucial for effective traffic efficient solution for traffic management. The
management. DDPG's ability to manage continuous successful application of LSTM networks in traffic
action spaces and multi-dimensional state variables flow prediction underscores their reliability and cost-
positions it as an ideal candidate for real-time traffic effectiveness, enabling traffic authorities to optimize
signal control, demonstrating its potential to mitigate resource allocation and enhance overall traffic flow
congestion and enhance overall traffic flow. efficiency.

E. Camera Based Traffic Detection Systems

III. PROPOSED SYSTEM DESIGN
The increasing complexity of urban traffic
necessitates effective solutions for real-time vehicle
detection and identification. Camera-based systems, A. Proposed Framework
particularly those using advanced deep learning The proposed adaptive traffic signal control system
models, have proven reliable for traffic is designed to mitigate traffic congestion at urban
monitoring.[10] emphasize the vital role of Closed intersections by utilizing advanced machine learning
Circuit Television (CCTV) cameras in Intelligent and reinforcement learning techniques. The system is
Transportation Systems (ITS), enhancing incident composed of three primary components: a YOLO
detection and vehicle monitoring while improving model for vision-based real-time vehicle detection, a
traffic flow management. Long Short-Term Memory (LSTM) network for traffic
flow prediction, and a Deep Deterministic Policy
The YOLOv8 model stands out for its Gradient (DDPG)-based reinforcement learning
algorithm for optimizing traffic signal timings.
accuracy, achieving a training accuracy of 77% and a
testing accuracy of 96% on a diverse vehicle dataset.
Its robust architecture effectively handles challenges
B. Data Flow in the Proposed Framework
such as occlusion and overlapping vehicles.
Combining YOLOv8 with CCTV technology shows The data flow in the proposed system initiates with
great potential for real-time vehicle detection, as its the acquisition of a real-time video feed from a single
rapid image processing capabilities facilitate timely CCTV camera installed at the intersection. This camera
responses to traffic conditions. Data augmentation is placed to monitor traffic conditions on one of the
techniques further improve detection accuracy, roads leading to the intersection. The captured video
feed is processed in real time by the YOLO model.
particularly for small objects. In this study, utilizing a
YOLO model extracts essential information. This
CCTV camera positioned at a distance to monitor processed data, along with supplementary contextual
intersection vehicle activity effectively leverages information such as the date and time, is subsequently
YOLOv8's strengths, offering a reliable solution to input into the LSTM network.
real-time vehicle detection challenges and ensuring
accurate urban traffic management. The LSTM network is tasked with predicting traffic
flow on the other three roads that converge at the
intersection. By analyzing the processed data from the
YOLO model, the LSTM network generates accurate
F. Traffic Prediction systems traffic predictions, leveraging historical traffic patterns
and real-time data. This prediction capability
Long Short-Term Memory (LSTM) networks significantly reduces the need for multiple cameras and
have proven to be a highly effective tool for traffic the associated processing power, as it enables the
flow prediction, as highlighted by [9]. Their research system to predict traffic conditions on all roads of the
intersection using data from a single camera.
demonstrates that LSTMs excel at modeling complex
 Following the traffic predictions, the output data is B. Data Collection
transferred to the DDPG-based reinforcement learning
(RL) model. The RL model employs these predictions The data collection for this study was carried out in
to determine the optimal traffic light sequence for the two distinct phases to ensure the accuracy and
intersection. The primary objectives are to maximize reliability of the dataset. The initial phase involved the
vehicle throughput, minimize waiting times, and use of CCTV footage provided by the Colombo City
reduce vehicle queues by dynamically adjusting the Police’s CCTV division. However, inconsistencies in
signal timings based on real-time traffic conditions. the quality and coverage of the footage made it
This integration of the YOLO model, LSTM network, necessary to conduct a second phase of data collection
and DDPG-based RL model into a unified framework to obtain a more reliable dataset.
creates a cost-effective and adaptive traffic In the second phase, a V380 Pro WiFi outdoor
management system capable of responding to the ever- camera was strategically installed at a construction site
changing demands of urban traffic. near the Borella intersection for continuous monitoring
over a three-week period. Prior to installation,
necessary permissions were obtained from the building
owners and both the Borella and Colombo North Police
Stations to ensure compliance with local regulations.
This study adopted a novel approach by using a single-
camera setup, chosen for its cost-effectiveness and
feasibility. The V380 Pro camera, equipped with a
wide-angle dual-lens, provided comprehensive
coverage of the entire intersection from a distance,
reducing the need for multiple cameras. The camera
was fitted with a 128GB micro SD card, offering ample
Fig. 1. Flow of proposed framework storage capacity for extended recording. Additionally,
the camera’s real-time monitoring capabilities allowed
remote access to check its status and manage SD card
IV. METHODOLOGY usage via a dedicated mobile application, ensuring
continuous oversight throughout the data collection
process.
A. Traffic Perception Survey The V380 Pro camera's night vision functionality
also provided clear footage during nighttime, which
Prior to selecting the Borella intersection for this study, was a significant improvement over the police
a comprehensive traffic perception survey was surveillance cameras used in the first phase. The
conducted to ensure the site accurately reflected typical camera setup required only a power source and a SIM
traffic conditions and challenges. The survey, card with sufficient data capacity to maintain
distributed online to a random sample of intersection connectivity, making it a cost-effective and practical
users, collected preliminary data on traffic patterns, solution for capturing traffic data at the intersection.
peak hours, and common congestion issues. The recorded video files were saved in AVI format and
later converted to MP4 for further analysis. VLC Media
Additionally, data from the Borella Police Station
Player was used for efficient transcoding, ensuring the
provided critical insights into traffic volume and
video files were in a more accessible format for
complexity at the intersection. The combined findings processing and analysis.
confirmed Borella as a high-demand area with intricate
traffic flows, making it an optimal focus for this
research. Fig 2 illustrates key concerns and traffic
patterns from the survey.

Fig. 3. Camera installed at the junction

Fig. 2. Insights from traffic demand analysis

C. Vehicle Detection model E. Traffic Flow Prediction model
The collected CCTV data will be employed to train The proposed system’s traffic volume prediction
a vehicle detection and counting model based on module leverages a Long Short-Term Memory (LSTM)
YOLOv8l. This proposed adaptive traffic signal control network, a specialized type of recurrent neural network
system integrates the YOLO (You Only Look Once) (RNN) optimized for time-series data analysis. This
model for real-time vehicle detection and tracking, LSTM model predicts traffic volumes at an intersection
which is essential for effectively managing traffic flow by utilizing real-time data from one road and drawing
at urban intersections. Utilizing the ultralytics/yolo on patterns learned from historical traffic data. This
framework, the YOLOv8m model is optimized for approach minimizes the need for multiple cameras,
high-speed processing, enabling real-time thus reducing monitoring costs. Developed using
classification of detected objects, including cars, buses, TensorFlow and Keras, the model's architecture is fine-
and motorcycles. tuned through Keras Tuner’s RandomSearch algorithm
for optimal hyperparameter settings.
To enhance detection capabilities, the YOLO
model is combined with the Deep SORT (Simple The input data, derived from the YOLO model
Online and Realtime Tracking with a Deep Association outputs (date, time, and traffic volumes), is processed
Metric) algorithm. This integration allows the system through six traffic-related metrics and standardized
not only to detect vehicles but also to track them across using a StandardScaler to enhance training efficiency.
multiple frames, assigning a unique identifier to each These inputs are then reshaped to match the LSTM’s
vehicle that is maintained until the vehicle exits the required format. Extensive hyperparameter tuning
scene. minimizes the mean absolute error (MAE) on a
validation set. To prevent overfitting and improve
Additionally, the system incorporates vehicle performance, strategies like early stopping and model
counting and speed estimation by defining reference checkpointing are employed during training. Once
lines at the intersection. These lines are used to count trained, the network predicts traffic volumes on other
vehicles as they enter and exit different roads. Vehicle roads at the intersection based on real-time input and
speed is estimated by analyzing movement across learned historical patterns.
frames in relation to these reference lines. The system
also captures and utilizes contextual data, including
date, time, and estimated vehicle speed. To optimize F. Traffic Control Algorithm
the YOLO model for the Borella intersection, video
data covering both daytime and nighttime traffic, The core of the decision-making process in the
including peak and off-peak hours, as well as proposed system is powered by the Deep Deterministic
challenging conditions such as adverse weather, will be Policy Gradient (DDPG) algorithm, a model-free, off-
extracted and annotated. A custom labelling system policy reinforcement learning approach well-suited for
was developed to differentiate between vehicle continuous action spaces. Unlike Q-learning, which is
appearances during day and night. This approach designed for discrete actions, DDPG excels in
enhances the model’s capability to detect and track environments where precise control over actions is
vehicles under varying lighting conditions. required, such as determining optimal traffic light
durations. DDPG integrates elements from Deep Q-
The annotated dataset will be used to train the Learning (DQL) and Policy Gradient methods,
YOLO model, optimizing its performance for detecting utilizing two neural networks: an actor network, which
and tracking vehicles across diverse lighting and traffic outputs a deterministic action (e.g., the duration of a
conditions. A key feature of this YOLO model is its green light), and a critic network, which evaluates this
customization for the Borella intersection, ensuring action by predicting the expected future rewards. The
efficient vehicle counting and tracking during both DDPG algorithm follows key steps:
daytime and nighttime. This specificity enhances the
model’s robustness and accuracy in various lighting  Experience Replay: The agent stores
scenarios, which is critical for maintaining consistent experiences (state, action, reward, next state)
performance throughout the day and night. in a replay buffer.
 Actor-Critic Framework: The actor network
outputs continuous actions based on the
D. Traffic Demand Analysis current state, while the critic network
Following the collection of traffic data via camera evaluates these actions by predicting the Q-
footage, a comprehensive traffic demand analysis was value, which reflects the expected cumulative
performed to evaluate patterns at the Borella reward, such as reduced congestion and
intersection. The YOLO model was used to accurately waiting times.
detect and classify vehicles, enabling a detailed
assessment of traffic volumes, peak periods, and daily  Target Networks: To stabilize training, target
flow variations. This analysis identified high and low networks are employed for both actor and
traffic periods, providing crucial insights into critic networks. These target networks are
congestion levels throughout the day. slowly updated to follow the main networks,
ensuring gradual and stable updates to the Q-
value.
  Policy Update: The actor network's policy is minimizing delays and congestion. This approach
updated by optimizing the expected harnesses the power of reinforcement learning in
cumulative reward, encouraging actions that continuous action spaces to provide an adaptive and
maximize future rewards. The critic network sophisticated traffic management solution.
minimizes the loss between the predicted and
target Q-values, calculated using the Bellman
equation:
Q(st ,at) = rt+1+γ⋅Q′(st+1,μ′(st+1))
where Q(st,at) is the estimated Q-value for the
current state-action pair, rt+1 is the reward and
Q′(st+1,μ′(st+1)) is the Q-value predicted by the
target networks.
 Exploration vs. Exploitation: To ensure that
the agent explores a wide range of actions
during training, noise is added to the actions
generated by the actor network.

The proposed adaptive traffic signal control

framework leverages the Deep Deterministic Policy Fig. 4. RL model training process
Gradient (DDPG) algorithm to optimize traffic light
sequences at the intersection. The primary objective is
to enhance traffic flow, minimize vehicle waiting
times, and reduce queue lengths. G. Simulation Model
The Simulation of Urban Mobility (SUMO)
 State: In the DDPG model, the state is a environment played a crucial role in developing and
multidimensional vector representing real- testing the proposed adaptive traffic signal control
time traffic conditions at the intersection. It system. SUMO served two primary functions: it acted
includes the number of vehicles on each as a platform for training the Deep Deterministic Policy
incoming lane, waiting time and queue Gradient (DDPG) reinforcement learning algorithm
lengths. and as a testing ground for evaluating the RL model
using real-world traffic data. In reinforcement learning,
 Action: The action involves the agent
an environment is essential for the agent to learn from
determining the traffic light time sequences, the consequences of its actions. SUMO offers a
specifically the duration for which each light controlled, risk-free setting to simulate various traffic
remains green, yellow, or red. Given that these scenarios, enabling the RL model to experiment with
durations are continuous variables. different traffic light sequences and assess their
 Reward: The reward function is crafted to effectiveness prior to real-world deployment.
balance various aspects of traffic
management, encouraging the agent to The system's performance will be compared against
optimize the overall traffic flow. The reward traditional static pre-timed traffic light controllers,
focusing on metrics such as vehicle throughput,
is calculated as:
average waiting time, and queue lengths. SUMO's
capability to accurately simulate and measure these
R=w1⋅(Number of vehicles passing through t metrics, especially queue lengths—which are
he intersection)−w2⋅(Average waiting time)− challenging to capture via cameras due to extensive
w3⋅(Average queue length) queuing distances at Borella—provides valuable
insights. This simulation-based approach ensures that
where w1, w2, and w3 are the weights the RL model is rigorously tested, refined, and prepared
assigned to each component. The goal is to to manage real-world traffic conditions effectively,
maximize vehicle throughput while thereby minimizing the risks associated with direct
minimizing waiting times and queue lengths. implementation.
V. RESULTS
The DDPG model's output is a priority-based
green light time sequence, dynamically adjusted based Initial progress has been made in several key areas
on real-time traffic volumes. Roads with higher traffic of the project. The data collection phase was
volumes receive extended green light durations, successfully completed, with 24-hour monitoring using
ensuring that congestion is effectively mitigated. The a Wi-Fi CCTV camera installed at the Borella
intersection over three continuous weeks, providing
final output is an optimized sequence of green light
real-world data records. The footage obtained indicated
timings for each road at the Borella intersection,
by Fig.6.
designed to maximize vehicle throughput while
 Fig. 6. Frame Obtained from the CCTV video Fig. 9. Annotated Frame

Figure 7 presents the results of the traffic demand

analysis, highlighting the identified traffic patterns,
peak hours, and flow variations throughout the day.

Fig. 7. Traffic Demand Analysis

Fig. 10. Confusion Matrix of YOLO mod

The YOLO model was successfully developed and

tested, demonstrating its ability to detect vehicles, An LSTM network was initially developed using a
count them, and record timestamps from the captured dataset sourced from Kaggle until the completion of
footage. Further fine-tuning was conducted using real-world data collection. While the model showed
footage from the camera installed at the Borella promising initial results, it was subsequently fine-tuned
intersection to account for its specific conditions. After with extracted data from real-world camera footage
annotating over 300-500 frames, the model was specific to the Borella junction to enhance its accuracy
retrained to handle varying traffic volumes, weather and applicability. The results obtained from this
conditions such as rain, and different lighting scenarios, refinement are illustrated in Fig 10.
including daylight and night conditions, until the
desired performance was achieved.

Fig. 11. Performance of LSTM network

Fig. 8. Processed frame from YOLO model
 The SUMO network representing the Borella [4] H. Wu, “Control method of traffic signal lights based on
intersection has been successfully constructed using DDPG reinforcement learning,” J. Phys. Conf. Ser., vol. 1646,
no. 1, p. 012077, 2020.
OpenStreetMap data. A Deep Deterministic Policy
[5] T. Mahmood, M. E. M. Ali, A. Durdu, and Electrical and
Gradient (DDPG) based reinforcement learning Electronic Engineering, Selçuk University, Konya, Turkey, “A
algorithm has been developed; however, the model two stage fuzzy logic adaptive traffic signal control for an
remains in the exploration phase. Further fine-tuning is isolated intersection based on real data using SUMO
necessary to optimize the algorithm for effectively simulator,” International Journal of Trend in Scientific
exploiting actions to achieve the desired outcomes. Research and Development, vol. 3, no. 4, pp. 656–659, 2019.
[6] W. A. C. J. K. Chandrasekara, R. M. K. T. Rathnayaka, and L.
L. G. Chathuranga, “A real-time density-based traffic signal
control system,” in 2020 5th International Conference on
Information Technology Research (ICITR), 2020.
[7] Y.-J. Lee and O. Min, “Long short-term memory recurrent
neural network for urban traffic prediction: A case study of
Seoul,” in 2018 21st International Conference on Intelligent
Transportation Systems (ITSC), 2018.
[8] R. T. Wakkumbura, B. Hettige, and A. Edirisuriya, “Real-time
traffic controlling system using Multi-Agent technology,” J.
Eur. Syst. Autom., vol. 54, no. 4, pp. 633–640, 2021.
[9] Shao, H., & Soong, B.-H. Traffic flow prediction with Long
Short-Term Memory Networks (LSTMs). Interdisciplinary
Graduate School, Nanyang Technological University.
[10] Hati Telaumbanua, A. P., Larosa, T. P., Pratama, P. D., Fauza,
R. H., & Husein, A. M. (2024). Vehicle Detection and
Identification Using Computer Vision Technology with the
Utilization of the YOLOv8 Deep Learning Method. Universitas
Fig. 12. SUMO network of Borella junction Prima Indonesia, Fakultas Sains dan Teknologi, Kota Medan.
[11] Abdulhai, B., Pringle, R., & Karakoulas, G. J. (2003).
Reinforcement Learning for True Adaptive Traffic Signal
VI. CONCLUSION Control. Journal of Transportation Engineering, 129(3), 278-
This study proposes an innovative adaptive traffic 285.
signal framework for urban intersections and conducts [12] Wei, H., Zheng, G., Gayah, V., & Li, Z. (2021). Recent
a quantitative evaluation to assess its feasibility for Advances in Reinforcement Learning for Traffic Signal
Control: A Survey of Models and Evaluation.
real-world implementation. Future research should
focus on real-world deployment of the proposed
framework, improving its capability to handle complex
scenarios, and exploring its scalability across multiple
intersections for wider application.

ACKNOWLEDGMENT
I would like to express my sincere gratitude to Mr.
Janaka Rathnayaka for his invaluable support in
providing the building space and power facilities
necessary for the successful installation of the camera
used in this research. His assistance was instrumental
in enabling the data collection process, and I am deeply
appreciative of his contributions to the successful
execution of this study.

REFERENCES

[1] J. Withanawasam and A. Karunananda, “Multi-agent based

road traffic control optimization,” in 2017 IEEE 20th
International Conference on Intelligent Transportation
Systems (ITSC), 2017.
[2] M. Sutharsan, S. Rajakaruna, S. Y. Jayaweera, J. A. C. M.
Jayaweera, and S. Thayaparan, “Vision-based adaptive traffic
light controller for single intersection,” in 2020 5th
International Conference on Information Technology
Research (ICITR), 2020.
[3] K. Shingate, K. Jagdale, Y. Dias, and Fr. c Rodrigues Institute
of Technology, “Adaptive traffic control system using
reinforcement learning,” Int. J. Eng. Res. Technol.
(Ahmedabad), vol. V9, no. 02, 2020

View publication stats