Research On Visual Multi-target Tracking Algorithm Based On Random Finite Set And Deep Correlation Filter

Visual multi-target tracking mainly uses the video context information to model the appearance information and motion information of multiple targets,realize the prediction of the target motion state,and mark the target trajectory.Visual multi-target tracking involves deep learning,machine learning,optimization algorithms and many other theoretical knowledge.With the development of computer hardware and mathematical theory,various civil and application systems for visual multi-target tracking have been realized,especially widely used in behavior recognition,traffic monitoring,intelligent driving,human-computer interaction and drone monitoring,etc.There are varying degrees of application in the depth and breadth of the target tracking field.This article focuses on the application of random finite sets in visual multitarget tracking,combined with the knowledge of deep learning and correlation filter,and launched a deeper research.The main results obtained are as follows:1.Aiming at the problems of target occlusion,missed detection,and increase in the number of target identification tags in the visual multi-target tracking task combined with the detection information,a multi-target tracking algorithm with motion information optimization and correlation filter is proposed.After the algorithm obtains the detection information of the target,it uses Kernelized Correlation Filter(KCF)to track the target and models the target's motion information.It is used to deal with the problem of inaccurate detector results that the target is blocked for a long time and cannot be tracked.At the same time,the smooth constraint of the confidence map is introduced on the basis of KCF to evaluate the degree of occlusion of the target,realize the adaptive update of the target template in KCF,and deal with the template pollution.Finally,the experimental results on the MOT17 dataset of MOT Challenge show that compared with the traditional detection and tracking algorithm IOU17,this algorithm can handle simple position prediction of occluded targets,and the accuracy of multi-target tracking MOTA index has increased by 2.43%.This algorithm has better stability and accuracy.2.In the visual multi-target tracking method of target detection with motion information,the modeling of simple motion information cannot cope with the complex motion of multiple targets,and there are a large number of missed and mis-followed targets.Therefore,a visual multi-target tracking algorithm combining correlation filter and GM-PHD is proposed.This algorithm classifies the tracking target through the association matrix,and integrates the idea of correlation filter to track the missed target.The idea of feature pyramid net is added to extract the deep image information and shallow image information of the target,and the occluded target does not update the target template and parameters,thereby reducing the pollution degree of the target template and reducing mis-following.For targets that have been occluded,Gaussian Mixture Probability Hypothesis Density Filtering(GM-PHD)will predict and update their position.If the target reappears in the later stage,the GM-PHD will re-associate the target label to reduce the fragmented trajectory and make up for the shortcomings of the detector that missed detection due to the occlusion of the target.Finally,the results on the MOT17 data set proved that compared with the current best tracking algorithm GMPHDOGM17 for GM-PHD,the accuracy of the multi-target tracking MOTA index increased from 49.9 to 50.3.3.In view of the large number of false detections and missed detections in visual multitarget detection and tracking,a GM-PHD visual multi-target tracking with multiple detections is proposed.This algorithm introduces a new type of target detector Efficient Det,and first trains the detector,and merges the trained Efficient Det detection results with DPM,FRCNN,and SDP,respectively,to reduce the number of false detections and missed detections of the original detector,and filter out some false detections.In addition,a multi-label window smoothing filter is added to the original algorithm to reduce the number of target identity jumps caused by longterm occlusion of the target,thereby further improving the tracking accuracy.Experiments show that this algorithm can improve the tracking effect on some video sequences.