Research On Target Tracking Method Of Correlation Filtering With Joint Spatiotemporal Regularization And Deep Features

With the advent of the era of big data,massive data information needs to be efficiently parsed.Vision is the main way for humans to obtain information from the outside world,and there are rich values hidden behind these visual information.To automatically and efficiently analyze visual information,computer vision plays an important role.As a popular problem in the field of computer vision,visual object tracking aims to track the specified target and obtain its position information in each frame of the video.It has been widely used in smart cities,traffic control,national defense,military and other important areas related to national economy and people's livelihood.At present,exploring high-speed and accurate tracking models has become a hot topic in this field.Nowadays,the spatio-temporal regularized correlation filtering method with deep features shows excellent performance in visual tracking.On the one hand,the high-efficiency calculation speed meets the real-time requirements of tracking.showed better adaptability.Unfortunately,cyclic sampling is often used in correlation filtering,which often introduces a large number of unreal negative samples,which leads to the problem of boundary effects and weakens the learning ability of the filter.The tracking template is also susceptible to occlusion,resulting in a degradation phenomenon that affects the filter performance.In addition,the correlation analysis between the current deep feature channel and the tracking target is still insufficient,which can easily cause data redundancy and limit the ability of deep feature representation.Focusing on the above problems,the main work carried out in this paper is as follows:(1)Aiming at the problems of boundary effects and template degradation,a dynamic spatio-temporal regularized correlation filter tracking method is proposed.Use saliency detection to adapt spatial constraints to object appearance changes during tracking.On this basis,the original filter template is introduced into the temporal regularization framework to improve the robustness of the tracker and accelerate the target tracking through the alternating direction multiplier algorithm.Experiments on mainstream UAV datasets show that the proposed method can achieve higher tracking accuracy than the popular correlation filtering methods.(2)Aiming at insufficient expressiveness of hand-crafted features and redundancy of deep feature data,an adaptive deep feature channel weighted target tracking method is proposed.First,in the dynamic spatio-temporal regularization correlation filter tracking framework,the deep features are fused to improve the accuracy of the algorithm.Second,adaptive channel weights are introduced into filter optimization to remove irrelevant data interference in deep feature data.Then,the target scale is predicted by manual features,and the deep feature determines the target position to improve the tracking speed.Finally,experiments on mainstream tracking datasets show that the proposed method has stronger robustness in complex tracking scenarios than popular deep correlation filtering methods.