An Enhanced Multiple Object Tracking Method With Improved Feature Fusion and Cross-Camera Trajectory Matching for Closed Scenes

Traditional MOT methods often struggle with issues such as occlusions, identity switches, and limited cross-camera trajectory matching accuracy. To overcome these challenges, this paper proposes a modified advanced multiple object tracking (MOT) method based on the CSTrack algorithm, named MACSTrack, which aims to address the competition between detection and re-identification (ReID) tasks in MOT and enhance its applicability to cross-camera scenarios. MACSTrack introduces a reciprocal inhibition network (REIN) to deeply mine and fuse the feature specificity and commonality between detection and ReID tasks, suppressing irrelevant features and enhancing task synergy. An improved scale-aware attention network (ISAAN) is employed to minimize information loss and feature distortion during feature fusion. Furthermore, a hierarchical association method based on ambiguous assignment modeling (HAAM) is proposed to improve the stability and robustness of object tracking in time series. To address the challenge of cross-camera trajectory matching in closed scenes, a two-stage strategy utilizing prior knowledge of the number of individuals is developed, which significantly improves matching accuracy. Experimental results on benchmark datasets demonstrate that MACSTrack outperforms the original CSTrack algorithm and achieves comparable performance to state-of-the-art MOT methods. Specifically, MACSTrack exhibits improvements in key metrics such as MOTA, IDF1, and a substantial reduction in identity switches. The proposed two-stage cross-camera trajectory matching strategy also achieves over 10% higher accuracy compared to conventional methods, highlighting the practicality and effectiveness of the approach.