An Improved Association Pipeline for Multi-Person Tracking

Daniel Stadler, Jürgen Beyerer; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2023, pp. 3170-3179

The association task of assigning detections to tracks in multi-person tracking has recently been improved by integration of a second matching stage for low-confident detections that are usually discarded in the tracking process. Despite its success, we find that this two stage matching has some weaknesses. For example, high-confident detections are preferred over low-confident detections in any case, even if the low-confident ones are more accurate. Therefore, a Combined Matching (CM) is proposed which considers all possible assignments simultaneously in a single matching stage and thus improves the association accuracy. Moreover, shortcomings of existing motion and appearance distance combinations are identified and a novel Combined Distance (CD) for motion and appearance information is introduced that significantly outperforms previous fusion approaches. Furthermore, we propose an Occlusion Aware Initialization (OAI) which prevents the start of ghost tracks from duplicate detections under occlusion. The effectiveness of our components is shown with extensive ablative experiments and the competitiveness of our tracker is demonstrated on the MOT17 and MOT20 benchmarks, where the current state-of-the-art is notably surpassed.