Multi-Pedestrian Tracking by Decoupling Confidence Weighting and Hierarchical Appearance Embedding Updates

The advent of powerful object detectors has led to the widespread adoption of the “tracking-by-detection” paradigm in multi-object tracking (MOT). However, challenges such as association errors during occlusions continue to hinder tracking performance. To address this issue, we propose three key improvements. First, we leverage the confidence scores generated by a detector with decoupled heads to weight the cost matrix. Second, we introduce a height-constrained intersection-over-union (IoU) distance metric. Finally, we classify the appearance features of objects in the current frame into two levels—high and low—based on their respective confidence scores. These features are then matched and updated against stored appearance embeddings from previous frames. Extensive comparative and ablation experiments conducted on the MOT17 dataset demonstrate that the proposed method significantly improves tracking performance in multi-object tracking scenarios.