PDTrack: Progressive Distance Association for Multiple Object Tracking

Yeqiang Liu, Weiran Li, Yanhao Ding, Zhenbo Li

Abstract: Multiple Object Tracking (MOT) has made significant progress in recent years. However, it still faces challenges such as frequent ID switches, trajectory fragmentation, and tracking losses in high-density pedestrian scenarios. To address these issues, we optimized the association algorithm based on the tracking by detection paradigm and proposed a progressive distance association matching algorithm (PDTrack). By using trajectory boxes gradually and orderly for association, we reduce computational costs. Additionally, we introduced the Disambiguation Model and Reassociation Model to reduce erroneous tracking results and avoid missed tracking objects, ensuring the integrity of the trajectories. Experimental results demonstrate that our proposed method achieves higher accuracy and fewer identity switches compared to other mainstream tracking algorithms, effectively enhancing multiple object tracking performance.

External IDs:dblp:conf/icassp/0001LDL25