Multiple Pedestrian Tracking Based on Coordinate Attention and Camera Motion Compensation

Abstract

Multiple pedestrian tracking (MPT) is essential for social robots to perform various tasks for the convenience of humans. In these crowd scenarios, occlusions are very common, especially from the perspective of mobile robot. However, most current approaches ignore the mobile platform, and result in poor performance of a high number of ID switches and missed detections when deployed on mobile robots. To address the problem, a novel MPT algorithm named C2-FairMOT is proposed in this paper, in which the coordinate attention mechanism and camera motion compensation are introduced to achieve more stable feature extraction and data association, respectively. Compared with the other three baseline algorithms, the proposed C2-FairMOT can reduce IDS and FN by at least 4.0% and 2.9%, respectively.