| The core task of multi-object tracking(MOT)is to predict the motion trajectories of multiple targets in a video sequence through observation and analysis.In recent years,with the development of object tracking technology,MOT tasks can be mainly divided into two paradigms,one is detection tracking paradigm and the other is joint detection tracking paradigm.In the detection and tracking paradigm,the multi-object tracking task is divided into two steps: First,the target is detected by the detection network frame by frame,and the object detection box is obtained.The detected targets are then assigned ids and associated with existing tracks through the association step.The joint detection and tracking paradigm is an algorithm framework that integrates target detection and target apparent learning into a unified network for multi-object tracking,which can improve the speed of MOT.Although the multi-object tracking algorithm has made great progress in recent years,it is still a very challenging problem to design a robust multi-object tracking algorithm when faced with many challenges in the field of multi-object tracking,such as motion blur,occlusion,fast motion,mutual interference between similar targets,crowded scenes,etc.Therefore,in view of the challenges in the field of multi-object tracking and the shortcomings of existing multi-target tracking algorithms,this paper mainly studies how to design a more robust multi-object tracking algorithm,and makes the following research achievements:(1)Based on the existing multi-object tracking framework Tracktor++,an improved multi-object tracking method is proposed,namely,combining relation network in Tracktor++ for multi-object tracking.The relation network is mainly integrated into Tracktor++ for interframe target association.The relation network can unify the apparent learning,similarity calculation,and ID allocation of the target under the overall network framework,enabling each target to learn a more favorable representation of ID allocation,reducing the frequency of ID conversion during tracking and thus improving the performance of the tracker.(2)A cascaded matching data association strategy is designed.In this task,the FRCNN detector is firstly used to generate object detection results.In order to further enhance the detector performance,the feature pyramid networks(FPN)is added to the backbone network to make the detection results more real.Then,GIOU and person re-identification network are used to match adjacent frame targets.Experimental results on two public data sets show that the proposed algorithm improves on IDF1 and IDs compared with some of the most representative algorithms. |
