Research On Video Object Tracking Algorithm Based On Correlation Filter

Object tracking is an important research task in the field of computer vision,which is widely used in video surveillance,intelligent navigation,medical services and other fields.Although the object tracking algorithm has made great progress,it is still a huge challenge to establish a robust target tracking system due to the widespread existence of pose,scale changes,occlusion and cluttered background during the actual movement of the object.In recent years,correlation filter algorithms use a large number of cyclically shifted samples for filter learning,and at the same time convert correlation operations in the time domain to point multiplication operations in the frequency domain.This not only improves the tracking accuracy of the algorithm,but also speeds up the tracking.It has become a framework widely used in target tracking.This dissertation mainly researches and improves correlation filter tracking algorithms.The main contents are as follows:(1)Aiming at the problem that the object bounding box is fixed when the kernel correlation filter algorithm performs scale estimation,it cannot adapt to the scale change of the target,and the tracking drift and loss are prone to occur in complex scenes such as occlusion or deformation,a robust kernel correlation filter tracker base on scale adaptive and occlusion detection is proposed.First,the combination of two complementary features of color and gradient can enhance the representation ability of the appearance model and improve the tracking performance of the algorithm;secondly,by adding the scale pool technology to solve the problem of fixed object scale,make the filter adaptive to the target scale Change,and use Newton iterative algorithm to find the maximum response value to better locate the center position of the object and the scale of the object;Finally,the occlusion detection setting is added when the appearance model is updated to avoid the tracking drift phenomenon when the tracking object is occluded and improve the robustness of the tracker in complex scenes.(2)The traditional correlation filter method uses the samples obtained from the cyclic shift operation of the foreground target to be disturbed by the boundary effect problem and cannot represent the real negative training samples.At the same time,the lack of real negative training samples will reduce the robustness of the tracker to the clutter background and increase the risk of tracking drift.Therefore,we propose a spatial and temporal regularization correlation filter tracker based on background awareness.Firstly,the clipping operation on the cyclic shifted samples can be generated by using the background awareness correlation filter,which can alleviate the influence of boundary effect.Moreover,the background information can be fully used to make the tracker distinguish the target and background well in the complex background scene.Secondly,the added temporal regularization term can make the filter adapt to the appearance change of the target.At the same time,the single histograms of oriented gradients feature is extended to multiple features to enhance the robustness of the object appearance model.Finally,using ADMM(Alternate Direction Multiplier Method)to solve the model can reduce the complexity of the algorithm and improve the tracking speed of the algorithm.