International Journal of Research Publication and Reviews Vol ( ) Issue ( ) (2021) Page 000
International Journal of Research Publication and Reviews
Journal homepage: [Link] ISSN 2582-7421
A Comprehensive Review of Artificial Intelligence-Based Helmet and
Seatbelt Detection Systems for Vehicle Safety Enforcement
Gaurav Ingle, Himanshu Shahare, Vaibhav Vaidya, Nishant Wankhede
Dr. Yogesh Golhar, Computer Engineering Department, St. Vincent Pallotti Collage of Engineering and Technology
ABSTRACT
The global burden of road traffic injuries and fatalities remains alarmingly high, particularly in developing nations where non-compliance with helmet and
seatbelt laws is rampant. As urban populations increase and vehicle ownership rises, manual law enforcement methods have proven insufficient in ensuring
adherence to safety protocols. With the advent of artificial intelligence (AI) and deep learning, automated systems for real-time helmet and seatbelt detection
have emerged as promising tools for improving road safety. This review presents an in-depth analysis of current research on AI-driven safety compliance
systems, focusing on models such as YOLO, CNNs, FCOS, and traditional machine learning approaches. We explore the methodologies, datasets, challenges,
and deployment strategies, offering a comprehensive understanding of the current landscape and identifying future research directions to advance these
technologies toward real-world implementation.
Keywords: Seatbelt Detection, Road Safety, YOLO, Computer Vision, Real-Time Monitoring, Vehicle Safety Compliance, Automated Traffic
Enforcement
1. Introduction
Road traffic injuries are a major public health issue, with approximately 1.3 million deaths occurring globally each year due to vehicle-related incidents,
according to the World Health Organization (WHO). A significant proportion of these fatalities could be prevented through the use of personal protective
equipment such as helmets and seatbelts. Helmets reduce the risk of death by 42% and head injuries by 69%, while seatbelts cut the risk of death among
front-seat passengers by 45%. Despite widespread awareness of these facts, compliance remains low, particularly in regions where traffic laws are weakly
enforced. Manual enforcement, though necessary, is not scalable and cannot guarantee consistent monitoring across vast transportation networks. In this
context, AI and computer vision technologies offer an efficient alternative, enabling real time, non-intrusive surveillance of vehicle occupants and
motorcyclists. This review critically examines the evolution, methodologies, and performance of AI-based helmet and seatbelt detection systems. We analyze
the key models and techniques employed in recent studies and highlight their strengths, limitations, and areas for improvement.
2. Background and Motivation
The motivation for automating traffic safety compliance stems from several factors:
Manual Enforcement Limitations:
Law enforcement officers cannot continuously monitor every road, intersection, or highway lane. Human error and subjectivity further reduce reliability.
Increasing Vehicle Density:
As vehicle ownership increases, particularly in urban areas, the scale of monitoring required outpaces manual capacity.
Technological Readiness:
The availability of powerful GPUs, embedded platforms (e.g., Jetson Nano, Raspberry Pi), and open-source libraries like TensorFlow, PyTorch, and OpenCV
has accelerated the development of real-time vision based applications.
* Corresponding author. Tel.: +0-000-000-0000 ; fax: +0-000-000-0000.
E-mail address: author@[Link]
2 International Journal of Research Publication and Reviews Vol ( ) Issue ( ) (2021) Page 000
In addition to traffic safety, automated detection systems can reduce confrontations between drivers and enforcement personnel, ensure unbiased rule
enforcement, and support post-incident investigations through video evidence.
3. Literature Review
Over the past five years, a substantial body of literature has emerged focusing on helmet and seatbelt detection. We summarize the contributions of major
studies in this domain below:
3.1 Helmet Detection Systems
• YOLO-Based Approaches:
YOLO (You Only Look Once) models are renowned for their balance between accuracy and real-time performance. YOLOv3 and YOLOv5 have widely
used for helmet detection in both urban and industrial settings. Aboah et al. (2023) proposed a YOLOv8-based model enhanced by few-shot sampling to
overcome data scarcity. Their system was effective under varied lighting and weather conditions and was recognized at the AI City Challenge 2023.
• CNN-Based Methods:
Lin et al. (2020) introduced a multi-task CNN framework capable of simultaneously detecting helmets, tracking motorcycles, and distinguishing between
riders and passengers. Their HELMET dataset, comprising over 91,000 annotated frames, provided a strong benchmark for training and evaluation.
• Traditional Image Processing:
Earlier studies employed color segmentation, HOG feature extraction, and SVM classifiers. Although computationally efficient, these methods suffer from
poor generalization under occlusions, lighting changes, or unconventional helmet designs.
3.2 Seatbelt Detection Systems
• YOLOv7 and YOLOv5 Systems:
Nkuzo et al. (2023) proposed a YOLOv7-based in cabin system that achieved 99.6% mAP and demonstrated high robustness under occlusions. Feng et al.
(2022) combined YOLOv5 with the Hough transform to detect seatbelt usage using external cameras positioned near windshields.
• Tiny-YOLO and Embedded Systems:
Lightweight versions of YOLO have been used in embedded platforms like Jetson Nano and Raspberry Pi. These systems can detect seatbelt misuse such
as wearing belts behind the back, which is harder to identify through external surveillance.
• Thermal and NIR Imaging:
The use of Fully Connected One-Stage (FCOS) detectors with thermal cameras enables accurate detection during night-time or low-light scenarios. NIR
imaging is also used in mobile enforcement units, such as police vehicles, for day-and-night monitoring.
• Classical Machine Learning:
Adaboost classifiers, trained on Haar or HOG features, have been used in earlier systems. Though less accurate than modern deep learning models, they
are suitable for edge environments with limited computational resources.
[Link] Methodologies
4.1 Object Detection Models
• YOLO Family (v3-v8):
These models detect multiple objects in a single pass, making them suitable for real-time video streams. Improvements in each version include better feature
pyramids (YOLOv5), backbone efficiency (YOLOv7), and support for limited data scenarios (YOLOv8).
• FCOS:
Unlike anchor-based detectors like YOLO and SSD, FCOS uses a per-pixel prediction approach. It is especially effective in object localization under poor
lighting when paired with thermal imaging.
CNNs and Multi-Task Learning:
CNNs extract hierarchical features and are adaptable to various classification tasks. Multi-task networks enable helmet detection and motorcycle tracking
within a single model, reducing computational cost.
4.2 Imaging Modalities
International Journal of Research Publication and Reviews Vol ( ) Issue ( ) (2021) Page 000 3
• RGB Cameras:
Most systems use high-resolution RGB feeds from CCTV or dashboard-mounted cameras.
• Thermal and NIR Cameras:
These sensors extend detection capabilities to low-light environments, providing thermal signatures and near-infrared contrast even in the dark.
• In-Cabin Imaging:
Seatbelt detection benefits from internal cabin cameras, which provide closer and more detailed views of seatbelt usage.
5. Deployment Scenarios and Applications
• Urban Monitoring Systems:
Integrated with traffic surveillance networks to monitor compliance at intersections and highways.
• Industrial Sites:
Used in construction zones to ensure workers wear helmets and seatbelts when operating vehicles or machinery.
• Mobile Law Enforcement:
Mounted on police vehicles to scan moving traffic and automatically issue violation notices.
• Access Control:
Deployed at gates of corporate offices or shopping malls to permit only compliant riders.
6. Challenges and Limitations
While current research shows promising accuracy, several limitations persist:
• Occlusion and Ambiguity:
Helmets disguised by turbans or hoods, and belts worn under clothing, pose recognition challenges.
• Data Diversity:
Most training datasets lack diversity in weather, geography, and ethnic attire, affecting model generalization.
• Hardware Constraints:
Deploying high-performance models on low-power embedded systems requires compression and pruning.
• Regulatory Acceptance:
Legal enforcement based on automated systems requires policy adaptation and standardization.
• False Positives:
Minor visual obstructions, like hair strands or shadows, often lead to misclassification.
7. Future Research Directions
To achieve real-world deployment at scale, future research should focus on:
• Dataset Expansion and Standardization:
Building large, annotated, multi-class datasets with global diversity.
• Cross-Domain Transfer Learning:
Leveraging pre-trained models across multiple cities or countries with minimal re-training.
• Edge Computing Optimization:
Using quantization, model distillation, and lightweight architectures for embedded platforms.
• Sensor Fusion:
Combining RGB, thermal, and NIR data for robust detection across all lighting and weather conditions.
• Regulatory Frameworks:
Developing legal protocols and government-backed frameworks to support AI-based enforcement.
8. Conclusion
4 International Journal of Research Publication and Reviews Vol ( ) Issue ( ) (2021) Page 000
AI-driven helmet and seatbelt detection systems have emerged as critical tools in modern traffic safety enforcement. From YOLO-based deep learning
models to hybrid systems combining classical ML with advanced sensors, significant progress has been made in achieving real-time, automated compliance
monitoring. However, to transition from research prototypes to large-scale implementations, these systems must overcome limitations in robustness,
generalization, and edge deployment.
Future innovations will rely on better datasets, smarter models, multi-modal sensing, and close collaboration between technologists, city planners, and
policymakers. These intelligent systems, when deployed effectively, can play a pivotal role in reducing traffic fatalities and making roads safer for all.
REFERENCES
1. Aboah, A., Zhang, X., & Chen, L. (2023). Real-time helmet detection using YOLOv8 with few-shot data sampling. Proceedings of the AI City Challenge 2023.
2. Lin, M., Chen, Y., & Ho, J. (2020). Helmet detection for tracked motorcyclists using CNN-based multi-task learning. Journal of Intelligent Transportation
Systems, 24(6), 527–540.
3. Nkuzo, T., Raman, A., & Singh, P. (2023). YOLOv7-based in-cabin seatbelt detection system for real-time vehicle safety compliance. International Conference on
Smart Transportation Systems.
4. Feng, J., Lee, C., & Kumar, R. (2022). Seatbelt detection using YOLOv5 and Hough transform techniques in challenging traffic conditions. IEEE Transactions on
Intelligent Vehicles.
5. Elihos, T., Giannopoulos, S., & Charalambous, C. (2018). Comparative analysis of SSD and CNN for seatbelt violation detection using NIR and RGB imaging.
IET Intelligent Transport Systems, 12(5), 378–385.
6. Jain, V., & Rajawat, A. (2021). Helmet violation detection using TensorFlow and Keras. International Journal of Computer Applications, 183(31), 22–28.
7. Kumar, S., & Shah, M. (2020). Seatbelt detection using cascade Adaboost and probabilistic Hough transforms. International Journal of Image Processing and
Vision Science, 5(2), 60–67.
8. Patel, A., & Raval, M. S. (2021). Seatbelt fastness detection based on in-cabin camera using Tiny-YOLO. International Conference on Embedded Systems and
Applications.
9. Bose, A., & Iyer, A. (2022). NIR camera-based mobile seatbelt enforcement system for day and night surveillance. Proceedings of the National Conference on
Embedded Vision Systems.
10. Sharma, D., & Khanna, R. (2020). Image-based seatbelt detection using traditional machine learning algorithms. Indian Journal of Computer Science and
Engineering, 11(3), 315–322.
11. Gupta, S., & Mehta, K. (2021). Dual detection system for seatbelt and helmet compliance using Raspberry Pi and OpenCV. International Journal of Emerging
Technologies in Engineering Research, 9(5), 67–73.
