Multi-Object Tracking with Adaptive Cost Matrix
Mingyan Wang, Bozheng Lit, Haoran Jiang, Junjie Zhang · 2022 IEEE 24th International Workshop on Multimedia Signal Processing (MMSP) · 2022
Multi-object tracking (MOT) aims at detecting and assigning identities for objects in videos. Complicated scenes, severe occlusions, irregular motions, and ambiguous appearances of objects hinder the further advance, which occurs frequently in pedestrian tracking. To tackle these challenges, we present a simple yet effective MOT framework focusing on two-fold: a more robust motion feature and a proper association paradigm. The Hybrid Motion Feature (HMF) integrates Intersection over Union (IoU), Euclidean distance metric, and area ratio information, and the latter two resort to the historical average variations to improve the cost matrix construction. Moreover, the new association paradigm, namely the Adaptive Calculation Method (ACM) performs better by avoiding the manual weighting of the motion and appearance-based cost matrix. In addition, the new correction method, i.e., reinitializing the state of the Kalman Filter when a severe mismatch occurs between the ground truth and the predicted trajectory, mitigates the effect of irregular motions. We achieved 80.7 MOTA, 78.5 IDF1, and 64.0 HOTA, outperforming the state-of-the-art model ByteTrack on the public MOT17 benchmark.