Abstract Multi-object tracking plays a vital role in computer vision, widely applied in various applications, such as surveillance, autonomous driving, and video analytics. In real-world scenarios, fast object motion and frequent occlusions often lead to blurred or missing appearance information, resulting in reduced detection accuracy and failed trajectory association. Such challenges severely compromise the robustness and accuracy of tracking.To address the above problems, this study proposes MTrack, a novel multi-object tracking framework that integrates both appearance and motion cues. Inspired by state-space modeling, specifically, we first design a motion feature module for object movements in both horizontal and vertical directions across frames using the Mamba state-space model, capturing global motion features. Next, we design a motion-appearance coupling module that effectively integrates the appearance features from the previous frame with the modeled motion features, thereby enhancing the model's ability to localize objects when appearance cues are unreliable.MTrack achieves strong performance on multiple benchmarks, including MOT17, MOT20, DanceTrack, and SportsMOT. It significantly outperforms existing methods in key metrics such as HOTA and IDF1, demonstrating both its effectiveness and generalization capability.
Zhou et al. (Tue,) studied this question.