MEDIATrack: Advanced Matching Strategy for Detection-Based Multi-Object Tracking

Wei-Shan Chang, Jun-Wei Hsieh, Chuan-Wang Chang, Kuo-Chin Fan

Abstract: Multi-object tracking (MOT) technology is widely applied to traffic flow monitoring, human flow monitoring, pedestrian tracking, or tactical analysis of players on the courts. It associates the detection boxes with tracklets for each frame in the video. The challenges of MOT include long-term occlusions, missing detections, and complex scenes. Although many trackers have proposed to solve these problems, the tracking results still have room for improvement. In this paper, we propose a solution named MEDIATrack (Matching Embedding Distance & IOU Association Track), a two-stage online multi-object tracking method based on ByteTrack. We replace the Kalman Filter with the NSA Kalman Filter, introduce appearance features for track association, and design a punishment mechanism to alleviate errors in complex scenes. In addition, we remove the nonactivated strategy, and the high-score unmatched detection boxes are directly added to the tracklets. On MOT17, we achieve 79.3 MOTA, 76.5 IDF1, and state-of-the-art performance.