PMTrack: Multi-object Tracking with Motion-Aware

Xu Guo, Yujin Zheng, Dingwen Wang; Proceedings of the Asian Conference on Computer Vision (ACCV), 2024, pp. 3091-3106

Tracking-by-detection typically involves associating detection boxes across frames in a video sequence. A common approach is to use Kalman filter for prediction and matching with detection boxes based on IoU. However, the Kalman filter is a linear prediction method, which, in scenarios involving camera motion or nonlinear object motion, will result in issues like ID switching or tracking loss. To address the problem, we propose a method that leverages phase correlation to calculates the translational relationship between adjacent frames, maping target positions into the current frame's coordinate system. This positional correction effectively compensates for the shifts caused by camera movement, significantly reducing ID switches. Furthermore, our method distinguishes between the motion and stationary states of trajectories, thereby enhancing tracking stability and accuracy. Our experimental results demonstrate that the proposed approach attains real-time efficiency and excels in scenes with camera motion. It achieves an MOTA of 80.17%, IDF1 of 78.93%, and HOTA of 64.04% on the MOT17 test sets, surpassing mainstream works in terms of multiple performance indicators.