MAT: Motion-aware multi-object tracking

Modern multi-object tracking (MOT) systems usually build trajectories through associating per-frame detections. However, facing the challenges of camera motion, fast and occlusion, it is difficult to ensure quality long-range or even tracklet purity, especially for small objects. Most frameworks depend heavily on performance re-identification (ReID) data association. Unfortunately, ReID-based association not only unreliable time-consuming, but still cannot address false negatives occluded blurred objects, due noisy partial-detections, similar appearances, lack temporal-spatial constraints. In this paper, we propose an enhanced MOT paradigm, namely Motion-Aware Tracker (MAT). Our MAT a plug-and-play solution, mainly focuses high-performance motion-based prediction, reconnection, First, nonrigid pedestrian motion rigid are blended seamlessly develop Integrated Motion Localization (IML) module. Second, Dynamic Reconnection Context (DRC) module devised guarantee robustness reconnection. The core ideas in DRC dynamic-window cyclic pseudo-observation trajectory filling strategy, which can smoothly fill fragments caused by occlusion blur. At last, present 3D Integral Image (3DII) efficiently cut off useless track-detection connections using Extensive experiments conducted MOT16&17 challenging benchmarks. results demonstrate that our achieve superior surpass other state-of-the-art trackers large margin with high efficiency.