Research and Optimization Exploration of Multi-Target Tracking Reproduction Based on Long-Tail Trajectory Distribution

The performance of Multi-Object Tracking (MOT) in complex scenes is often limited by the long-tail distribution characteristics of the data. This study systematically verified the negative impact of long-tail distribution on tracking performance by reproducing the static view enhancement (SVA), dynamic view enhancement (DVA), and group Softmax (GS) modules proposed by Chen et al. Experiments have shown that on the MOT20 dataset, the replication method improves the comprehensive tracking accuracy (MOTA) and identity retention rate (IDF1) of tail categories by 9.1% and 5.3%, respectively, significantly reducing the problems of ID switching (IDS) and missed detection (FN). In addition, this article proposes innovative directions for Dynamic Trajectory Enhancement (DTA) and Adaptive Group Classification (AGC), exploring the feasibility of further improving tail sample diversity and classifier generalization ability through optical flow guided interpolation and dynamic threshold adjustment, respectively. This study provides a reproducible technical path and theoretical extension space for multi-target tracking in long-tail distribution scenarios.