VeTrack leverages transformer-based appearance modeling and confidence-guided multi-level association for robust multi-vehicle tracking

Multi-Object Tracking (MOT), a critical computer vision task, is essential for dynamic scene understanding in applications like autonomous driving and intelligent surveillance. Road vehicle tracking faces key challenges: identity switches (ID switches) due to highly similar vehicle appearances and tracking interruptions from frequent occlusions in dense traffic. This paper proposes VeTrack, a novel vehicle MOT framework addressing these issues. First, the TransformerReID module leverages Transformer self-attention to model global context and extract fine-grained discriminative appearance features, significantly improving distinction of similar targets and reducing ID switches. Second, a Multi-level Association Strategy adaptively selects matching methods based on detection box confidence, mitigating occlusion-induced interruptions in complex traffic. Integrating YOLOX detection, TransformerReID, and this strategy yields an efficient VeTrack tracker. Extensive experiments on mainstream public datasets, including MOT17, MOT20, and BDD100K demonstrate the superior performance of VeTrack. It achieves state-of-the-art (SOTA) levels, particularly in handling target occlusion and appearance similarity. For instance, VeTrack achieves an MOTA score of 82.2% on the MOT17 and ranks among the top three in FPS performance, significantly outperforming existing approaches. Its strong BDD100K vehicle tracking performance further validates effectiveness and generalization, highlighting real-world deployment potential for intelligent transportation systems.