Research On Target Tracking Algorithm Based On The Combination Of Correlation Filtering And Depth Features

Target tracking refers to automatically tracking the target specified by the boundary box of the first frame in the subsequent frames of the video.The first frame initialization of target tracking is provided by the target detection results,and the tracking results can be used for other high-level tasks such as semantic segmentation.Therefore,target tracking is an intermediate level module for video content analysis research,which has great practical value and broad development prospects.At present,it has been widely used in national defense construction,aerospace,medicine and health,national economy and other fields.The algorithm based on correlation filtering has been a hot topic in the field of target tracking because of its good effect and fast speed.This paper focuses on the kernelized correlation filter algorithm,and improves the tracking framework to solve the difficult problems of scale changes,occlusion,loss,etc.,and achieves better tracking results.The main research contents of this paper are as follows:(1)A scale adaptive target tracking algorithm combined with context awareness(CASA)is proposed.First,the fusion features of color and texture are used to improve the target features representation ability.Then,the context information sample is added to the kernelized correlation filter tracker,which improves the discriminating ability of the tracker.Finally,different scale samples are used to learn the discriminate correlation filter to counter the target scale change.This algorithm has a better improvement in tracking positioning accuracy and overlap success of boundary box.(2)A multi-detector target tracking algorithm based on deep network features(MDDN)is proposed.The fusion feature of CASA algorithm is replaced by the convolution layer feature of the deep network,which can be roughly located according to the semantic information of the deep convolution layer feature,and then finely determined according to the spatial details of the shallow convolution layer feature.Then,for the case of target loss,a multi-detector mechanism is added.The historical detector is strategically retained according to the motion information of the previous frames,and the final target location is judged by synthesizing the corresponding response values of each detector.The improved algorithm is more effective than CASA algorithm in this paper,and can be re-tracked in time when the target appears..(3)The related demo software is designed in this paper.The software allows users to more intuitively understand the tracking effect of the two algorithms in this paper,as well as comparison with other classic algorithms.The two target tracking algorithms proposed in this paper are verified on OTB100 video library.Tracking performance is significantly improved on video sequences with different properties.