Research On Visual Tracking Based On Multi-Correlation Filters

Driven by the development of computer hardware and information technology,the related technical algorithms in computer vision and image processing have made great progress.Video tracking has always been the core research task in computer vision and wide applications.Object tracking can be seen as a continuous process of object recognition and detection in a video sequence.The objects are determined by manual calibration or automatic recognition in first frame,and in the subsequent video sequence tracking algorithm detects the current state of object,including position,size or other information.However,many complex factors will affect the robustness and accuracy of the algorithm,which makes the target tracking technology more difficult,such as rotation and scale variation,or illumination change,occlusion,interference from similar background,etc.In recent years,the introduction of correlation filter has made great success in target tracking,and the high accuracy and frame rate make it a popular research framework.Circular shifts generates a large number of training samples,but produces many false negative samples,which to some extent reduces the recognition ability of the filter template.Therefore,more and more researchers focus on the design of complex regularization priors to improve the discriminative ability.In this paper,a fusion framework of Plug-and-Play correlation filter(PPCF)is proposed,which can integrate different correlation filter trackers iteratively,so that different regularization priors are fused to enhance filter templates.The experiments on video database shows that,compared with the existing trackers based on single correlation filter,the proposed method has higher accuracy and robustness,which proves the effectiveness of PPCF algorithm.In addition to filter learning,search window is also an important part of determining the accuracy of object location estimation.Usually,the search area center is directly set at the location estimated by the previous frame,and the existing filter learning models often rely on simple and fixed regularization expression.To overcome these limitations,a robust location-aware and regularized adaptive correlation filter(LRCF)tracking algorithm is proposed.LRCF establishes a new bi-level optimization model to deal with both position estimation and filter training.The coarse-to-fine positioning strategy combines with the fusion strategy of multiple correlation filters to make the algorithm more robust.A large number of comparative experiments prove that LRCF tracking framework effectively integrates filter training and target location estimation,which enhances the filter and makes target location more accurate.Compared with the state-of-the-art tracking algorithms,LRCF tracking framework also shows the best performance.