Research on intelligent analysis of traffic scene target detection and tracking based on YOLOv8 and DeepSORT

To address the challenges of identity loss and switching in complex scene tracking, this paper presents an enhanced multi-object tracking framework integrating YOLOv8 with DeepSORT. In the detection phase, we introduce a parameter-free attention mechanism to strengthen vehicle feature extraction capability, thereby optimizing detector performance. For the tracking phase, we implement a multi-view fusion strategy through computer vision techniques involving homography matrix computation and image stitching, which facilitates robust target matching via the DeepSORT algorithm's association of single response relationships between targets. The proposed method undergoes comprehensive evaluation on standard MOT16 and MOT17 benchmarks. Experimental results demonstrate significant improvements: the enhanced framework achieves 63.1% MOTA (Multiple Object Tracking Accuracy), representing a 3.3% absolute increase over the baseline. Furthermore, it effectively reduces identity switches by 6.2% compared to the original implementation. These quantitative improvements verify the effectiveness of our technical enhancements in handling complex tracking scenarios, particularly in maintaining target identity consistency under challenging conditions. The experimental validation confirms that our approach substantially mitigates common tracking failures in crowded environments while maintaining real-time processing capabilities.