Multiple Object Tracking Based On Tracklet-plane Matching

Video multiple object tracking(MOT)technology is an essential topic in computer vision.The purpose of MOT is to simultaneously track all the objects of interest and obtain the complete trajectories.This technology plays an important role in intelligent video surveillance and other applications.Current MOT methods aim to model the temporal relationship among detected objects and associate them into trajectories.Thus,two major challenges of MOT are the occlusion between the objects and the noisy object detection results.In order to solve these problems,we propose an MOT approach based on tracklet-plane matching(TPM)in this paper.TPM first constructs temporally high-related object detections into short tracklets.Then,a tracklet-plane matching process is introduced to organize related tracklets into tracklet-planes.Every end of a tracklet can be connected to a tracklet-plane.Finally,the tracklets in the different side of tracklet-planes can be associated into long trajectories with some post-processing oprations.In this way,the missing of the tracking object and the repeated tracking trajectory could be handled effectively.In addition,this paper proposes a tracklet-importance calculation method to evaluate the reliability of the tracklet.We will reduce the probability of the association between the low-importance tracklets both with the tracklet-planes and with other tracklets.The trajectories with low-importance will be removed.We also propose a similarity calculation strategy between tracklets based on the representative object.The most reliable object in the tracklet can be selected by the proposed representative-selection network.Then we can use the appearance similarity between the representative objects to measure the similarity between the tracklets.Therefore,the noisy information in the tracklets could be elimated effectively and the tracking performance of our approach could be futher improved.Finally,experimental results on benchmark datasets demonstrate the effectiveness of our proposed approach,which obviously outperforms the state-of-the-art MOT methods in tracking performance.