Who is partner: A new perspective on data association of multi-object tracking

Occlusion problem refers to the challenge of accurately tracking blocked or occluded objects. It occurs when mul-tiple objects are moving nearby or overlapping with each other in the scene. Despite its frequent occurrence in multi-object tracking (MOT) tasks, occlusion is often overlooked by researchers. This paper proposes a simple and effective method to partially solve the occlusion problems of multi-object tracking by developing the Partner Mining Module (PMM) and the Partner Updating Module (PUM). The PMM module mines the space relationship between objects, and the PUM module uses the relationship obtained by the PMM module to update lost tracklets' positions. Importantly, these two modules can be integrated into existing data association-based multi-object tracking methods without any additional training expenses. Furthermore, this study proposes novel methods for computing measurement uncertainty to enhance trajectory accuracy. Experiments conducted on MOT16 and MOT17 datasets show the effectiveness of the proposed modules. Integration of PMM and PUM into original methods substantially enhances the IDF1 score by 1 point.& COPY; 2023 Elsevier B.V. All rights reserved.