Multi-Object Tracking Using Kalman Filter and Historical Trajectory Correction for Surveillance Videos

Abstract

Abstract In view of the problem of accuracy degradation, target loss, and the inability to continue tracking after target reappearance caused by video blurring, occlusion leading to sudden disappearance of the target in the process of surveillance videos multi-object tracking, we propose a multi-object tracking using Kalman filter and historical trajectory correction for surveillance videos. Firstly, the dynamic decoupling head is constructed to replace the original detection head, the loss function is optimized to improve the YOLOv8 detection algorithm. Then, the improved Kalman filter is constructed and the historical trajectory correction module is constructed to track the target pedestrians. Finally, the multi-object tracking module is constructed by combining the improved Kalman filter and the historical trajectory correction module, and the multi-object tracking module is outputted from the surveillance videos. The experiments on MOT17 and MOT20 show that evaluation metrics such as MOTA, IDF1, HOTA, etc. obtained by our method are the superior performance.