Multiple Object Tracking Based On Kernelized Correlation Filter

In the field of computer vision,Multiple Object Tracking(MOT)has always been a hot and difficult point of research.In this area,the amount of compute data is large and the processing speed of the algorithm require high,so that the tracking speed and accuracy cannot be well balanced.For this reason,the Kemeliezd Correlation Filter(KCF)with fast advantage in single object tracking is applied to MOT.In order to ensure high tracking accuracy while considering the tracking speed.This article is divided into the following four sections:(1)Object detection.Due to the misdetection and missed detection of 2D MOT 2015 and MOT 16 dataset targets,a Faster Region-based Convolutional Neural Networks(Faster R-CNN)is proposed,and the results of network detection are replaced with data sets provided.(2)Object tracking.In order to add target appearance information in MOT,and to combine speed and accuracy factors,this paper uses KCF model.However,it has been found that the tracking results are not ideal when performing multi-target tracking data sets,because the KCF algorithm does not support multi-scale.In order to overcome this problem,this paper solves the method of calculating the Intersection-Over-Union(IOU)between the detection target box and the prediction target box,and useing the detection target box instead of the predict.(3)Data association.For the calculation of the IOU of part(2),it is actually a mutual calculation problem between a batch of detection boxs and a batch of prediction boxs,which is essentially the task assignment and optimization problem.To this end,the Hungarian algorithm is used to determine the optimal correlation between the targets.(4)Object redisplay judgment.In the process of target tracking,there will be a situation in which the target disappears briefly and reappears.In order to improve the ability of the model to deal with such problems,it is proposed to use the similarity judgment method to judge the similarity of the emerging "new target",if the value less than a certain range,it is determined to be the same target.Because both the KCF and the IOU model have fast response,the algorithm can meet the requirements of processing data online.In addition,the KCF belongs to the model of the position search decision target,and the IOU is an overlap region calculation model.The combination of the two can effectively solve the target non-inertial motion and camera shake.Experiments were conducted on the public 2D MOT 2015 and MOT16 data sets:for the object detection part,experiments showed that the detection results using the Faster R-CNN network can be 6%higher than the accuracy provided by the original data set;for the multi-target tracking part,compared with other popular methods,the tracking accuracy is still higher than 10%while ensuring a processing speed of 30 frames/s or more.In order to further promote the development of KCF in the field of MOT,the source code of this article is open source,and the link is:https://github.com/HappyUncle/KCF_IOU_Tracker.