Appearance Awared Detector for MOT: An Enhanced ReID Branch for Tracking Memorize
Hongyu Chen · Academic Journal of Science and Technology · 2023
The traditional ByteTrack approach to multi-target tracking, focusing on simple, effective algorithms. It performs well in short-term multi-target tracking tasks with 80.3 MOTA, 77.3 IDF1 and 63.1 HOTA on MOT17 30 FPS and is currently ranked number one in the MOTChallenge rankings. But for situations where the camera is moving, or for completing target recovery tasks after a brief loss of position and track information, conventional MOT modules often do not perform as well as they should. To make bytetrack perform better when the motion background pattern is more complex, we used the re-identification module and added appearance information to enhance ByteTrack’s MOT process. We verified our improvement on MOT17, which achieved higher results than the original on the metrics Precision and Recall, and was also more suitable for more complex scenarios.