Skip to content
Open access

A Multi-Stage Deep Learning Framework for Real-Time Seatbelt Detection in Traffic Surveillance Videos

2026 · International Journal of Advanced Computer Science and Applications · 0 citations · 18 references

Abstract

Seatbelt non-compliance is a major road safety concern. Automated road traffic monitoring using surveillance cameras can support large-scale detection. This study builds a multi-stage deep learning framework for real-time seatbelt detection in roadside traffic surveillance videos. YOLOv8s detects vehicle windshields and defines the region of interest; then it detects occupants within that region. ByteTrack then tracks occupants across consecutive video frames. The ResNet18 model classifies each occupant as wearing or not wearing a seatbelt in each frame. To improve the reliability of frame-level predictions, confidence scores are temporally aggregated for each tracked occupant, and high-confidence occupant crops are selected using a top-k strategy. This study constructs a dataset that combines publicly available seatbelt-related datasets (eight sources) with 1,200 locally collected and annotated images from different areas in Jordan. The dataset includes 9,516 windshields, 5,058 occupants, 3,043 seatbelt-on, and 1,695 seatbelt-off instances. YOLOv8s achieved 98.7% mAP@50 and 96.8% mAP@50 for windshield and occupant detection, while ResNet18 achieved 96.73% classification accuracy. The complete temporal framework achieved 85.92% precision, 93.40% recall, and an F1-score of 89.51%. The core pipeline achieved an average processing rate of 34.04 frames per second. The complete application achieved 29.00 frames per second, including displaying the results on the screen and storing them on disk. These results demonstrate the feasibility of the proposed framework for real-time traffic surveillance applications.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.