Real Time Traffic Signal Optimization and Vehicle Surveillance

Using Deep Learning

Mohammad Fathimunnisa, Vemula Deepthi, L. Sandhya Rekha,

Anumula Pavithra and Takkasheela Archana

Department of Computer Science and Engineering, Ravindra College of Engineering for Women, Venkayapalle, Pasupala,

NH 340C, Nandikotkur Rd, Kurnool, Andhra Pradesh, India

Keywords: Traffic Signal Control, Deep Learning, Reinforcement Learning, Convolutional Neural Network, You Only

Look once, Real‑Time Vehicle Tracking, Intelligent Traffic Management.

Abstract: In the modern scenario wherein, traffic congestion and poor signal management are the serious concerns of

cities, there is a demand for best-in-class solutions for its real-time optimization. This paper proposes a traffic

signal control and vehicle detection system based on deep learning with CNN and object detection. It

automatically adjusts signal times based on traffic density and reduces jams and optimizes road traffic. The

detection of traffic offenses and observation of road activity further augment surveillance. Its method offers

better accuracy, reduced waiting time and increased flexibility compared to the traditional methods.

Experimental results confirm the system's effectiveness for optimal urban traffic management with prompt

decision-making. Using AI-based strategies together can ensure a hyper-efficient, scalable solution to

Transportation needs’ in the new age.

1 INTRODUCTION

urban traffic congestion is considered one of the

biggest urban problems growing with the growing

populations, high volume of traffic and limited road

facilities. Traditional traffic signal control systems

operate on fixed time cycles with no ability to adapt

significantly to real-time changing traffic conditions.

This inefficient process leads to not only delays and

driver irritation, but also more fuel being used,

contributing to environmental pollution, and

economic losses. Therefore, demand for intelligent

traffic control systems, that dynamically change

signal timings, based on real-time information and

improve road safety, with the help of improved

surveillance technologies, is increasing.

Deep Learning, with the advent of Artificial

Intelligence and the exponential growth of

computing power, is yet another powerful tool to

address complex problems in traffic management.

Deep learning-based algorithms, particularly CNNs,

along with advanced object detection frameworks,

such as YOLO (You Only Look Once), have achieved

remarkable accuracy in vehicle detection,

classification, and tracking in real time. These models

are designed to analyse video streams from

surveillance cameras to estimate traffic volume,

identify vehicles and track movement patterns with

high accuracy and speed. Another innovation is the

addition of predicting future traffic flow through

RNNs or LSTMs, which allows the system to

anticipate needs and adjust the signal proactively

rather than reactively.

Along with adapting traffic signal timings, the

system also features a rich native vehicle monitoring

suite that automatically identifies traffic violations,

tracks Threatful movements, and aids law

enforcement agencies with a computer vision-based

automated analysis. The entire system is designed to

work in real-time for immediate decision-making

and traffic management. This paper aims to evolve a

solution that is economic, scalable, and future-ready

that bridge the collaboration between deep learning

and smart infrastructure to improve urban mobility

and build smarter, safer, and sustainable smart cities.

Fathimunnisa, M., Deepthi, V., Rekha, L. S., Pavithra, A. and Archana, T.

Real Time Trafﬁc Signal Optimization and Vehicle Surveillance Using Deep Learning.

DOI: 10.5220/0013922400004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 5, pages

51-55

ISBN: 978-989-758-777-1

2 RELATED WORK

The development of deep learning has significantly

improved the optimization of traffic signal and

monitoring of cars. Older traffic management systems

rely on pre-set signal cycles that fail to adapt to current

ground traffic, leading to congestion and inefficiency.

To address this, researchers have also explored AI-

led approaches.

Car detection and traffic density estimation have

previously been performed using Convolutional

Neural Networks (CNN) and object detection

algorithms, such as YOLO and SSD. Zhang et al.

(2020) used CNN-based approach to classify cars in

real-time, significantly improving accuracy while

avoiding adaptive signal control. Wang et al. (2021)

extended on this concept by integrating YOLOv4 with

smart traffic surveillance systems, delivering better

detection accuracy at the expense of less predictive

power.

Recurrent Neural Networks (RNNs) and Long

Short-Term Memory (LSTM) models are proposed to

improve traffic prediction. Li et al. Zhang et al. (2019)

applied LSTM networks to predict traffic congestion

patterns using historic data for the adjustment of signal

times ahead of time. However, their model was not

real-time responsive, which was improved by Chen et

al. (2021), which relies on hybrid CNN-LSTM models

for real-time traffic density estimation and prediction.

These algorithms are helpful in the area of vehicle

monitoring as deep SORT and vehicle re-

identification methods are widely used Zhang et al.,

2019. Huang et al. m tracking approach based on

Deep SORT to analyse and track vehicle movements,

which significantly reduced tracking errors. Liu et al.

(2021) further supported anomaly detection by

integrating spatiotemporal features to improve

detection for traffic safety violations such as light

traversing and illegal lane change.

This is complemented by the emergence of edge

computing, which, combined with the expansion of

IoT-empowered infrastructure, has increased the

scale of AI driven traffic systems. Sharma et al. (2020)

described an IoT-based smart traffic system with

edge AI and a reduced delay in decisions. Although

efficient, such system did not possess any vehicle

tracking module that was improved by Patel et al. The

novel one-tier INNOWAT with integrated cloud

based real time infringement detection (Tuberen,

2021).

Despite all these advancements, still challenges

remain such as real-time scalability, accuracy in dense

traffic settings, and deploy ability. The solution we

propose is adhering to these loopholes, through a

synergic conjunction of CNN-based object detector

(YOLOv5, SSD), LSTM on traffic flow prediction,

Deep SORT on vehicle tracking, and an edge-cloud

hybrid deployment strategy to

achieve real-time

adaptable traffic signal control and surveillance.

3 METHODOLOGY

3.1 Theoretical Structure

The current work improves traffic signal control and

vehicle measurement using AI, Deep Learning and

ITS. CNNs (YOLOv5, SSD) for real-time detection

of vehicles and RNNs, LSTM to predict traffic flows.

Through reinforcement learning, a dynamic

algorithm in terms of traffic density controls

durations in an adaptive signal control. Violations

like breaking a red-light and lane crossing are

detected, tracking is done using Deep SORT and by

vehicle Re-ID.

It combines edge computing and cloud integration

to make real-time decisions and has a traffic

authority dashboard. The AI system enhances urban

mobility, reduces congestion and supports the

development of smart cities. Figure 1 represent

Schematic Flow of Theoretical Structure.

Figure 1: Schematic Flow of Theoretical Structure.

Data Acquisition &Prepocessing

Traffic Density Analysis & Signal

Optimization

Predictive Traffic Flow Analysis

Vehicle Surveillance & Cloud

Integration

Edge Computing & Cloud

Integration

System Deployment & Evaluation

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

3.2 Perceived Features

3.2.1 Adaptive Traffic Signal Control

AI-optimized control tailors the signal length

dynamically in accordance with real-time traffic

flow, which decreases congestion and optimizes

traffic flow. No fixed signal cycle and makes the best

use of the road possible minimizes time that vehicles

are idle and burning fuel.

3.2.2 Real-Time Vehicle Detection &

Surveillance

Advanced deep learning architectures like YOLOv5

and Deep SORT can successfully detect, classify and

track vehicles. It monitors three kinds of traffic

offenses: jumping red light, lane drifting, and illegal

parking, to make automated law enforcement more

effective, and also enhance road safety.

3.2.3 Predictive Traffic Flow Analysis

The system uses LSTM and RNN to predict traffic

congestion patterns from historical and real-time

traffic data. Thisleaves stars to more actively patrol

moving traffic and optimize signals ahead of time to

avoid causing a traffic jam and wasting time for

travel.

3.3.4 Edge Computing & Smart City

Integration

Real-time processing is performed through edge

computing and cloud-based integration is leveraged

for centralized management. It supports IoT-enabled

smart city architecture, thereby allowing scalable

implementation across several intersections for a

smart and sustainable urban transportation system.

4 RESULTS AND EVALUATION

The proposed deep learning-based traffic signal

optimization and car monitoring system was

evaluated in a simulated smart city environment. In

this case, the performance was evaluated based on

key performance parameters such as enhancement in

traffic flow, vehicle detection rate, prediction

reliability, and system responsiveness. The results

validate the system's potential to enhance urban

traffic management and the efficiency of public

security enforcement.

4.1 Traffic Signal Optimization

With the detection and classification models based

mostly on YOLOv5 and CNN it achieved 92.5%

MAP which will assure accurate detection of vehicle

even for multiple lanes and various angles. Dynamic

signal adjustment mechanism, reduced the average

waiting time by 35% compared with traditional

fixed-signal cycle. Traffic signals were optimized to

increase vehicle throughput and eliminate traffic jams

in peak hours. Table 1 represent Performance Metrics

of the Adaptive Traffic Control System.

Table 1: Performance metrics of the adaptive traffic control

system

Metric

Value

Achieve

Impact

Mean Average

Precision

(MAP)

92.5%

High accuracy in

vehicle detection

Reduction in

Waitin

Time

35%

reduction

Improved traffic

flow efficienc

Adaptive Signal

ustmen

Real-time

Smooth transition

in si

nal timin

4.2 Predictive Traffic Flow Analysis

To be specific, the predictive model was RNNs based

or LSTM based, it was trained on historical and real-

time traffic data. Train congestion prediction model

prediction achieved an average Root Mean Square

Error (RMSE) of 2.8 which enabled more proactive

adjustments to the signal control. By preventing such

major congestion from building, the prediction

helped reduce congestion on the road, providing a

better experience for people and increasing the

likelihood that the vehicles would move in a

coordinated manner rather than in a stop-start driving

style. Table 2 represent Predictive Model Evaluation

and Traffic Optimization Impact.

Real Time Trafﬁc Signal Optimization and Vehicle Surveillance Using Deep Learning

Table 2: Predictive model evaluation and traffic

optimization impact.

Metric

Value

Achieve

Impact

Prediction

Accuracy

87%

Effective

congestion

forecastin

RMSE 2.8

Reliable

predictive

modelin

Reduction in

Abrupt

Stoppa

28%

Improved

traffic

smoothness

4.3 Vehicle Monitoring and Violation

Enforcement

The vehicle surveillance module, implementing Deep

SORT tracking system and vehicle re-identification,

was able to detect vehicles movement and identify

traffic offense. The system achieved an F1-score of

89.3% while accurately identifying red-light

offenses, lane dissolution, and illegal parking.

Automation of surveillance reduced reliance on

manual tracing by 40%, enhancing the efficiency of

law enforcement. Table 3 represent Evaluation

Metrics for Real-Time Traffic Violation Detection

System.

Table 3: Evaluation metrics for real-time traffic violation

detection system.

Metric

Value

Achieve

Impact

F1-score 89.3%

Accurate

vehicle

classification

Violation

Detection

Accurac

91%

Effective

traffic law

enforcemen

Real-time

Alert

Response

Time

<1 second

Instant

detection of

rule violations

4.4 System Deployment and Real-time

Performance

It provided real-time processing with almost no

latency by integrating the edge computing and cloud

by implementing this system. Traffic Signal

Adaptation mean response time was

0.8(seconds)+which promotes swift adaptability to

varying traffic flows. Experimental analysis of the

system on bench tests with various volumes of traffic

demonstrated the pre-states of the heavy load

intersection. Table 4 represent System Performance

and Environmental Impact Analysis.

Table 4: System performance and environmental impact

analysis.

Metric

Value

Achieve

Impact

System

Response

Time

0.8 seconds

Fast reaction

to traffic

conditions

Scalability

Assessment

Successfully

tested

Suitable for

high-traffic

intersections

Reduction in

Fuel

Consumption

12%

Supports eco-

friendly

transportation

5 DISCUSSION

The envisioned system manifests a considerable leap

forward in real-time traffic management by

combining deep learning models with adaptive signal

control and vehicle monitoring. The results reveal a

very noticeable decrease in congestion, decreasing

waiting time by 35% based on real-time traffic

volume and dynamic adjustments in signals.

Predictive analysis of traffic flow using RNN and

LSTM models provides better congestion forecast

accuracy (87%), preventing sudden stoppages and

enhancing general smoothness in traffic. The vehicle

surveillance system has a high accuracy (F1-score:

89.3%), which allows effective violation detection

and real-time generation of alerts in one second, thus

enhancing the efficiency of law enforcement. Also,

the rapid response time of the system is 0.8 seconds,

which makes the system practical for large-scale

deployment in urban areas. With the fuel

consumption lowered by 12%, the suggested solution

helps make the environment sustainable, thus offering

a practical approach to smart city traffic management.

The research points to the efficiency of AI-based

traffic optimization, supporting its viability for

boosting mobility, safety, and conformity in cities.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

6 CONCLUSIONS

The designed deep learning-oriented traffic

management system efficiently optimizes traffic

signal control and improves vehicle monitoring in

urban areas. With the incorporation of real-time video

analytics, object recognition, and predictive

modeling, the system heavily lowers congestion,

enhances traffic flow efficiency, and enhances law

enforcement through violation detection automation.

The experimental outcomes show a 35% decrease in

waiting time, a prediction of traffic flow accuracy of

87%, and a precision in vehicle tracking and

enforcing rules of 89.3%. Moreover, the system is

helpful in achieving environmental sustainability by

decreasing fuel consumption by 12%. The

combination of edge computing and cloud-based

deployment makes it scalable and real-time, thus

ideal for smart city usage. In general, this study offers

a novel, AI-based solution to urban traffic control,

opening the door to future developments in intelligent

transportation systems.

REFERENCES

Bochkovskiy, A., Wang, C. Y., & Liao, H. Y. M. (2020).

YOLOv4: Optimal Speed and Accuracy of Object

Detection. arXiv preprint arXiv:2004.10934.

Chen, L., Li, Z., & Zhang, Y. (2020). Real-Time Traffic

Signal Control Using Deep Reinforcement Learning.

Transportation Research Record: Journal of the

Transportation Research Board.

Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.,

Fu, C. Y., & Berg, A. C. (2016). SSD: Single Shot

MultiBox Detector. European Conference on Computer

Vision (ECCV).

Ma, X., Tao, Z., Wang, Y., Yu, H., & Wang, Y. (2015).

Long short-term memory neural network for traffic

speed prediction using remote microwave sensor data.

Transportation Research Part C: Emerging

Technologies, 54, 187–197.

Ministry of Road Transport & Highways (MoRTH),

Government of India – Traffic Management and Smart

City Planning Report.

Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016).

You Only Look Once: Unified, Real-Time Object

Detection. Proceedings of the IEEE Conference on

Computer Vision and Pattern Recognition (CVPR).

Sultana, S., & Akbar, M. (2021). Smart Traffic

Management System using IoT and Deep Learning.

International Journal of Computer Applications,

174(14), 20-25.

Tang, T., Deng, H., & Huang, Y. (2021). Vehicle Re-

identification Based on Deep Learning: Methods,

Datasets, and Challenges. IEEE Transactions on

Intelligent Transportation Systems.

Wojke, N., Bewley, A., & Paulus, D. (2017). Simple Online

and Realtime Tracking with a Deep Association Metric.

IEEE International Conference on Image Processing

(ICIP).

Zhang, Y., Qin, L., & Liu, Y. (2019). Urban Traffic Flow

Prediction Based on a Spatiotemporal Deep Learning

Framework. Sensors, 19(18), 3929.

Real Time Trafﬁc Signal Optimization and Vehicle Surveillance Using Deep Learning