YOLOv8-TMS: spatiotemporal attention networks for real-time occlusion-resilient urban traffic monitoring

Abstract

Traffic monitoring from roadside cameras benefits from fast object detection, yet real street scenes remain difficult because occlusions, small targets, and adverse weather conditions reduce visual reliability. This study presents YOLOv8 for traffic management system (TMS), which enhances YOLOv8 using hybrid attention refinement, temporal coherence modeling, and adaptive occlusion handling to improve stability in crowded frames. Experiments on the traffic management enhanced dataset from the Roboflow universe street view project use 5,805 training images and 279 testing images across five road-user categories. The model achieves 95.2% mAP@0.50 in sunny scenes and 90.0% mAP@0.50inrainyscenes, whilesustaining 50msinference time and30frames per second throughput with 8 GB graphics processing unit memory. The results support reliable deployment for near real-time traffic analytics under varying conditions.