One-shot pedestrian multi-object tracking with multi-level matching mechanism

In multi-object tracking, the one-shot method combines detection and appearance embedding, simultaneously locating the target and extracting appearance features in a network, which improves the efficiency of the multi-object tracking algorithm. However, the previous method did not fully utilize the detection information. We propose multi-level matching mechanism is proposed to improve tracking performance by processing targets hierarchically. The results on the MOT17 dataset show that our method achieves competitive results.