AIPT: Adaptive information perception for online multi-object tracking

Information perception is crucial in MOT tasks. Recent approaches use positional, motion, and appearance information to model object states. However, in scenes involving camera motion, tracking tasks suffer from image distortion, trajectory loss, and mismatching issues. In this paper, we propose Adaptive Information Perception for Online Multi-Object Tracking, abbreviated as AIPT. AIPT consists of an Adaptive Motion Perception Module (AMPM) and an Asymmetric Information Suppression Module (AISM). In AMPM, we design an Adaptive Image Distortion Recovery Module (AIDRM) to perceive distortions in unknown scenes, allowing the tracker to autonomously recover distorted images as the scene changes. By designing the Information-Guided Trajectory Restoration Module (IGTRM), the tracker learns object motion states from prior information and constructs accurate reconstruction information during trajectory loss. Furthermore, our AISM module utilizes masking information to suppress potential relationships between asymmetric objects, thereby enhancing the ability of tracker to handle mismatches. Both AMPM and AISM exhibit excellent scalability, seamlessly integrating with most advanced tracking methods. Ultimately, our AIPT achieves leading performance on multiple benchmark platforms, including MOT17, MOT20, and KITTI.