Research on Target Detection and Tracking Methods in Complex Road Scenes

Abstract

This paper investigates the issues of small object omission and poor tracking stability in complex road scenarios for autonomous driving. To address the problem of small object omission, we developed the ECF-YOLO model by integrating a multi-scale fusion feature enhancement module and a dilated convolution context enhancement module into the classical YOLOv5 architecture. Additionally, incorporating the NWD positioning loss, derived from a Gaussian distribution, significantly improves detection accuracy. Furthermore, lightweight models are achieved through DepGraph pruning and knowledge distillation techniques. Moreover, the matching strategy of the ByteTrack algorithm is optimized through weak clue adjustment and low-score box reuse based on the improved detection model. Experimental results demonstrate that the ECF-YOLO model achieves a 4.3% improvement in mAP performance on the self-made road target dataset RSTO. The lightweight model's parameter size and computational cost are reduced by 48.3% and 39.6% respectively. The improved ByteTrack algorithm shows fewer ID switches in real-world driving scenarios. © 2025 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.