CollabMOT Stereo Camera Collaborative Multi Object Tracking

Abstract

The recent advances in deep learning techniques enable 2D Multi-object tracking (MOT) to achieve remarkable performance over traditional methods. However, most 2D MOT algorithms primarily utilize only single-camera view. Therefore, they are prone to frequent tracking losses and track-ID switching under conditions due to limited viewpoints and occluded objects. To alleviate this problem, we propose a stereo-camera-based collaborated multi-object tracking (CollabMOT) method that performs online and dynamic association of multiple tracklets from baseline MOT algorithms in overlapping views of stereo cameras. CollabMOT utilizes appearance similarity to generate global tracking IDs that unify the same tracklets between viewpoints of stereo cameras. It then leverages the transitive information from these global tracking IDs to reconnect the disrupted tracklets in each camera view. CollabMOT improves the overall performance of baseline 2D MOT methods on a single camera view by mitigating the problem of ID switching. Evaluation of CollabMOT on Argoverse-HD and KITTI dataset shows improved performance over baseline MOT methods. As a result, the proposed method improves the performance of the recent state-of-the-art method on the 2D MOT task of the KITTI dataset from 79.5 to 80% on High Order Tracking Accuracy (HOTA) score for vehicles.

Keywords

• Multi-object tracking
• stereo vision
• deep learning
• data association