Seatbelt Violation Detection in Road Surveillance Images Based on Edge Artificial Intelligence for Sustainable Road Safety

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Sustainable road safety continues to be a major concern worldwide, with failure to wear seatbelts representing a significant risk, particularly in developing countries. In this study, we propose a deep learning-based framework for the automatic detection of seatbelt violations, leveraging a custom dataset collected on Moroccan roads that captures different vehicle types, driver clothing color variations, and environmental conditions. We explore the YOLO family of object detection models, from YOLOv8 to YOLOv26, evaluating 23 variants ranging from nano to large architectures. Each model was trained to classify drivers as Person–Seatbelt or Person–No-Seatbelt. Experimental results on this diverse dataset demonstrate that the proposed approach achieves a validation mAP@0.5:0.95 of 63.5% and a test mAP@0.5:0.95 of 56.1% on three entirely unseen vehicles, highlighting its robustness across previously unseen driver and vehicle conditions. To assess real-world deployment, the best-performing models are further benchmarked on the NVIDIA Jetson Orin Nano embedded platform using PyTorch and TensorRT FP16 formats in MAXN mode, with the most efficient configuration sustaining 94.4 FPS inference at 6.40 W. This work contributes to sustainable road safety by developing traffic violation detection systems and provides a foundation for future automated monitoring of driver behavior in real-world conditions.