Multi-target Multi-camera Tracking Algorithm Based On Re-identification

Multi-Object Tracking technology has huge application requirements in many practical fields: with the popularization of more and more surveillance equipments deployed in crowded places such as airports,schools,shopping malls,etc.,the demand for automatic analysis of surveillance video is increasing day by day.Recently,in the field of autonomous driving,there has been a need for surround view tracking,that is,using a couple of in-vehicle cameras to simultaneously observe traffic participants around the auto-driving vehicle.However,due to the existence of occlusion phenomenon,the current Multi-Object trackers cannot track targets accurately in dense scenes,and the phenomenon of target loss and ID exchange often occurs.And there is no open source Multi-Camera Multi-Object tracking algorithm that can help autonomous vehicles achieve surround view tracking.To this end,this paper proposed the following object tracking methods:A Multi-Object Tracking method for pedestrians in dense scenes combined with head tracking.In order to reduce the negative impact of severe occlusion in dense scenes on multiple object trackers,considering that the head is the highest and least occluded part of the pedestrian’s entire body,a new multiple pedestrian tracking method combined with head tracking was proposed,namely Trape Hat.First,a head tracker was used to generate the pedestrians’ head movement trajectories,and the pedestrians’ whole body bounding boxes were detected at the same time;Secondly,the degrees of association between the head bounding boxes and the whole body bounding boxes were calculated,and the Hungarian algorithm was used to match the calculation results above;Finally,according to the matching result,the head bounding boxes in the head trajectories were replaced with the whole body bounding boxes,and the motion trajectories of the whole body of the pedestrians in the dense scene were output.Experiments suggested that this method based on head tracking effectively reduced the negative effects of false negatives and false positives caused by severe occlusion in dense scenes.(2)A Multi-Camera Multi-Tracking method based on Re-identification.In order to achieve the effect of surround view tracking on auto-driving vehicles,and the higher-level auto-driving technology could be realized as soon as possible.This chapter proposed a Multi-Camera Multi-Tracking method based on Re-identification,which splited the Multi-Camera Multi-Tracking task into two sub-tasks: multiple object tracking under one single camera and target re-identification between different cameras.First,the detection results of the target vehicles were completed using the target detector,and the detected targets were tracked using the multiple object tracker.Then,the target re-identification algorithm was used to find the same targets under multiple cameras.After integrating the results of tracking and re-identification,the target motion trajectories were given in each camera and a globally unique id was given to each target.Such Re-Id based surround view tracking algorithm won the first prize in the Calm Car MTMC Tracking Challenge competition.