부산대학교 도서관

Multi-object tracking (MOT) aims at associating bounding boxes and assigning identities of objects in videos. Data association is the core problem of MOT, which is essentially the solution to a bipartite graph problem, and real-time processing places higher demands on the performance and complexity of the algorithm. Generally, a spatio-temporal gating mechanism is employed for the accurate and fast association. However, the traditional distance gating mechanism could suffer from long- range ID switch in some situations. In this paper, we propose a cascade matching algorithm with Generalized IoU Gating Mechanism (GIGM) to prevent confused association, and leverage CIoU matching as a second association strategy after the cascade matching to obtain more reasonable tracking results. Experiments on the MOT16 dataset verify the capability of the proposed tracking method, the key metrics including IDF1, MOTA, and HOTA, have been boosted compared to the popular online tracking methods.