Research On Target Tracking Algorithm Based On Correlation Filtering

The technology of targets tracking is an important research direction in the domain of computer vision.It is also widely used in realistic society,such as unmanned driving and video surveillance.In recent years,the tracking method based on correlation filter has been widely concerned since it was proposed.Although researchers have put forward many novel methods and ideas,there are many defects in specific scenarios due to the complexity of tracking problems and the uncertainty of tracking scenarios.Therefore,it is still challenging to achieve efficient and robust target tracking,especially in long-term tracking,where the target may experience severe occlusion and move out of the field of view.In this thesis,aiming at the model migration caused by object deformation,occlusion and long-term tracking,a robust object appearance model and a detector for repositioning are established to realize the tracking process.The main work of this paper is as follows:(1)A multi-feature context adaptive tracking method is proposed to solve the problem of short-term occlusion and deformation in tracking.Firstly,a multi-feature correlation tracker is used to estimate the initial position of the target,and then the target position is predicted accurately by using the context-aware filter.Then the scale estimation of the target is performed by training scale filter.Finally,the average peak energy and the maximum response value are used to update the model adaptively.The simulation results with 9 kinds of classical trackers show that the proposed method can effectively deal with the drift caused by short-term occlusion and deformation,and the success rate and accuracy can reach 82% and 89%,respectively.(2)To solve the problem of severe occlusion in long time tracking,an adaptive long time tracking algorithm based on complementary model and heavy detection is proposed.Firstly,relative confidence is used to fuse the trained complementary model(multi-feature correlation filter and global color model)to obtain the initial position of the target.Then,a scale pyramid is built at the initial position for scale estimation.Finally,the learning rate is adjusted with the historical reliable information to realize the adaptive updating of the model.Where,when the response value of the target position is greater than the training threshold,the target position is credible,and the tracker and detector are updated respectively.When the response value of the target is less than the redetection threshold,itindicates that the tracker has failed to track,the detector is activated and the detector is used to relocate the target.Simulation experiments were carried out on 100 groups of video sequences.Simulation experiments were carried out on 100 sets of video sequences.Compared with the classic long-term tracking algorithm TLD,the proposed method improved tracking accuracy and success rate by 40.5% and 50.7% respectively;compared with LCT,they have increased by 7.3% and 5.6% respectively.