Research And Design Of Multi-target Tracking Algorithm Based On Deep Learning

Multi-target tracking is an important branch in the field of computer vision and an important means of processing video data.It is widely used in many fields such as autonomous driving,intelligent surveillance,national defense and security,and intelligent robots.With the rise of deep learning and its application in the field of multi-target tracking,the performance of multi-target tracking algorithms has made a qualitative leap.However,the existing multi-target tracking algorithms still have many problems to be solved,such as: difficulty in obtaining the discriminative characteristics of the target,target occlusion,false detection of the detector and missed detection,etc.Aiming at the above problems,this paper uses the Tracktor algorithm as the basic framework to carry out the research and implementation of the multi-target tracking algorithm.The Tracktor algorithm realizes tracking with only detectors.It uses the bounding box of the detector to predict the position of the target in the previous frame in the current frame to complete the tracking.Tracktor has created a new tracking paradigm,achieved excellent tracking results in a simple way,but still has a certain potential for improvement.In order to accurately correlate data and extract discriminative features,this paper uses a new type of pedestrian re-identification network BoT Baseline to expand the Tracktor algorithm.Based on the pedestrian re-recognition benchmark framework,this network adds training techniques such as random erasure and uses batch normalized neck(BNNeck)to reasonably combine the two losses,achieving excellent performance with almost no loss in speed.In addition,this paper processes the MOT17 data set to obtain a pedestrian re-identification data set MOT17-reID,which is used to train the network.When the target disappears gradually,the tracking result must be partially occluded,which will affect the trajectory characteristics of the target.Therefore,this paper proposes a confidence-based appearance matching association.When updating the trajectory,filter the tracking results with insufficient confidence for more accurate data association.In the stage of judging the disappearance of the target,in order to reduce the influence of the occlusion on the multi-target tracking algorithm,the target disappearance judgment based on confidence is proposed,and the target that has been judged as disappeared is judged twice.Use the BoT Baseline network to extract the features of the target,and retain the targets whose features are close to the corresponding trajectory features,so that the algorithm has more continuity in the tracking of the target,and at the same time improves the detection result to a certain extent.This paper verifies the tracking performance of the algorithm on the MOT17 data set.An ablation experiment was designed to show the effectiveness of each innovative improvement.The experimental results show that the improvement schemes proposed in this paper have improved the performance of the multi-target tracking algorithm to a certain extent,and are progressively advanced.The previous improvements provide a basis for subsequent improvements.And support.By comparing with the various algorithms of the MOT Challenge,the algorithm in this paper demonstrates its advantages in target tracking integrity and achieved good tracking performance.