Research On Visual Tracking Method Based On Discriminative Scale Space Correlation Filter

Visual tracking is one of the research hotspots of computer vision.It has a wide range of applications in many fields,such as human-computer interaction,intelligent monitoring,automatic driving,smart city and so on.However,there is still no tracker that can be applied to complex scenarios where multiple challenges such as occlusion,rotation,and illumination changes.Therefore,it is of great practical value to continue the research in this field.In recent years,the idea of correlation filter has been applied in object tracking,and its advantages of high efficiency and excellent effect have attracted attention in object tracking.Based on the correlation filter tracking algorithm,this paper improves Discriminative Scale Space Tracker(DSST)in the aspects of feature representation,model building and update strategy.The main research contents of this paper are as follows:(1)In terms of feature representation,aiming at the limitations of DSST which uses a single feature,this paper is supplemented by a variety of manual features to better represent the target,while using the self-learning PCA algorithm to reduce the dimension of features,so as to remove redundancy and retain the effective information of the features,and further improve the efficiency of the tracker.Then,in order to make better use of the representation ability of each channel in the multi-channel feature,the adaptive weighting is applied to the multiple response channels.Finally,high-dimensional convolution features are incorporated into the algorithm,and its performance is improved by 4.4% through experiments.(2)In the aspect of model building,aiming at the defect of insufficient samples and boundary effect in the correlation filter trackers,this paper exploits context information to increase the number of real samples and restrain the response of the filter to the background area,so as to reduce the influence of boundary effect on the algorithm and improve the robustness of the tracker.At the same time,in order to enhance the discriminant ability of filters,subspace constraints are imposed on the original model.In essence,the constraint acts on the current filter.By using the distance between the historical filters and the current filter,the adaptive weights are applied to each filter.By summing the weights of these filters,the current filter is constrained which ultimately increases the success rate by 3.3%.(3)For the correlation filter trackers,the reliability of the template is crucial.The unreliable model participates in the process of update,which will reduce the tracking efficiency and may lead to tracking failure.Therefore,aiming at the model updating module,this paper studies and improves the idea of average peak correlation energy and minimum weighted interval,and proposes two model updating strategies,which can adaptively select high-confidence frames for updating,so that the tracker can abandon the contaminated model and guarantee the quality of filters.At the same time,the success rate of the methods is increased by 2% and1.1% respectively,and the tracking efficiency is greatly improved.In summary,this paper starts from three research points and improves DSST algorithm and its next version fast DSST(fDSST).Two improved methods are proposed,which are the real-time correlation filter tracker using weighted fusion feature and adaptive update and the real-time correlation filter tracker based on context fusion and subspace constraints.After a lot of analysis,comparison and verification,the effectiveness and superiority of the improved algorithms are fully proved.