Multi-cue multi-hypothesis tracking with re-identification for multi-object tracking

Multi-object tracking is an important research topic in the field of computer vision. In multi-object tracking, overlapping targets and dramatic changes in object appearance are major challenging problems. In this paper, we propose a multi-cue and multi-hypothesis tracking algorithm for multi-object tracking, which is based on re-identification (Re-ID) to alleviate the Id-Switch problem in the case of occlusion. Our approach has two major advantages: (1) Using metric learning with person Re-ID technology, our approach is able to reduce the distance within similar classes and increase the distance between classes. The appearance model we learn is more distinguishable than the appearance model obtained by the classification network trained on ImageNet. Our method can reduce the addition of different targets to hypothetical trajectory tree operations, thereby reducing hypothetical branches. (2) We build a re-identification search library. When the target is lost, we can find the trajectory of the target from the Re-ID search library and add it to the hypothesis trees. Therefore, our approach can be used to alleviate the problem of occlusion or target loss in multi-object tracking. To improve tracking accuracy further, we use the online training discriminative model kernel correlation filtering (KCF) to verify whether the branch added can be assigned to the current trajectory. Experiments show that our method outperforms other state-of-the-art methods on MOT challenge15, MOT challenge16.