Correlation Filters Based Visual Object Tracking Technology For Deformable Targets

The realization of target tracking algorithms occupies an important position in the design of intelligent visual information processing systems,and the research of concise and efficient correlation filtering target tracking algorithms is even more important.The focus of this paper is on the robustness of the algorithm when the target is deformed under visible light video.Unlike target translation or scale scaling,target deformation is a highdimensional image transformation.Therefore,the correlation filtering algorithm based on template features cannot predict the specific situation of target deformation in the next frame.Highlights of the research in this paper are as follows:(1)Research on the visual representation method of the target: This paper studies the target representation method based on the characteristics of the craft and the characteristics based on the Convocation Neural Network(CNN),as well as the fusion strategy,to help improve the algorithm's feature expression ability for the target;Combining the Similarity and Objectness information of the target and the template,CNN technology is used to construct a network framework for the feature template update strategy,so as to effectively maintain a reliable feature template.(2)Research on the spatial-temporal regularity terms of correlation filtering: This paper will systematically analyze and study the spatial-temporal regularization terms of the CF based algorithm,and design a spatial-temporal regularization term suitable for the deformable target according to the changeable characteristics of the deformable target.Specifically,this paper introduces the relaxation L0 norm to model the spatial attention of the correlation filter,so that the CF based algorithm focuses more on tracking the reliable area of the target.(3)Research on the CF algorithm based on the accurate appearance estimation model:This paper proposes a general target tracking framework that combines coarse positioning of correlation filtering and precise appearance estimation of deep learning,The combination of image estimation and the segmentation task in deep learning realizes the target appearance estimation method with pixel-level accuracy.(4)Evaluate the tracking effect of the above algorithm improvement work on public Visual Object Tracking(VOT)and Video Object Segmentation(VOS)benchmarks: It has been found in the tests that compared to the previous ones,our algorithm has improved significantly...