Motion-aware anti-similarity interference for multi-object tracking

Multi-object tracking (MOT) is essential for various UAV remote sensing applications, including traffic monitoring and disaster rescue. However, when appearance-alike objects follow intersecting or overlapping trajectories (similarity interference), the trackers often fail to distinguish them due to similar motion patterns. Many existing methods rely on the Kalman Filter for motion modeling, which only utilizes motion information from two adjacent time steps and easily leads to ID confusion during such interference. To address this issue, we propose a novel motion modeling framework based on state space models (SSM) to enlarge the temporal receptive field and improve association accuracy for similar objects. Specifically, multiple historical trajectories and camera viewpoint transformations are encoded to extract the long-term motion features. Afterwards, the multi-step future trajectories are predicted by the extracted motion features. Finally, the object is identified by the similarity in distance and velocity direction of predicted trajectories. Extensive experiments on VisDrone-MOT and UAVDT datasets demonstrate the superior performance of our method.