WDTtrack: tracking multiple objects with indistinguishable appearance and irregular motion

With the surge in object detection, Multi-Object Tracking (MOT) research has recently witnessed significant advancements. However, most previous studies have primarily focused on benchmarks involving distinguishing appearances and linear motion. In scenarios involving non-linear motion and similar appearances, these methods exhibit a drastic drop in performance. To address this issue, we propose WDTtrack, which incorporates spatiotemporal proximity, velocity orientation, and appearance similarity simultaneously. Firstly, we employ the Centroid Triplet Loss ReID (CTL) model to extract high-quality appearance embeddings. Second, we introduce Wider Bounding Box (W-BBox) and Direction Bank (DB) to capture abundant credible, and discriminative motion cues. Finally, we devise the Tracklet Recovery Mechanism (TRM) to facilitate long-term tracking maintenance. Extensive empirical results demonstrate that WDTtrack outperforms other trackers on the DanceTrack and SportsMOT dataset, highlighting its effectiveness and potential for further development. Specifically, WDTtrack achieves a 66.8 HOTA score, a 72.8 IDF1 score and a 55.9 AssA score on DanceTrack, and a 73.8 HOTA score, a 80.5 IDF1 score and a 64.3 AssA score on SportsMOT, substantially surpassing other non-Transformer algorithms.