A Salient Feature Fusion Visual Tracking Via Spatio-temporal Context Description Using Bayesian Framework

Object tracking is the process of real-time accurate positioning the target of a campaign in the continuous video image sequence, it is widely used in the range of applications in the vision navigation, intelligent transportation, video surveillance and other fields. This paper focus on how to learning and building the most different description of the target feature model in the space background, and carries on the scale transformation operation corresponding to different tracking algorithms to further improve the performance of target tracking under complex scene. The research process is divided into the following three parts:First of all, considering that the traditional Meanshift algorithm step by step iterative search only in the local region that cannot be used to accurately positioning the fast moving target, we propose the method that extract the significant color information of the target, and pre search its range of motion by using the significant color information to meet the prerequisites for the Meanshift iterative search. This method reduce the computation of determine the specific location of target, so as to achieve accurate positioning in the video.Secondly, since the traditional Meanshift tracking algorithm using only color features without considering the spatial structure between target and the surrounding background, this part propose a combined color and texture feature Meanshift tracking algorithm. In this algorithm, we design a novel combined feature, which fuses the color and improved block LBP feature as the representation of the tracking object, it can extract the main information like the edge and the corner of the object. Besides, with the help of moment information from the candidate object area, it can solve the problem of the changes of the scale and direction of the object during the tracking process.Finally, in view of the complex problem that target and background region surrounded it are similar, we choose the approach to formulate the spatio-temporal relationships model between the object of interest and its locally dense contexts in a Bayesian framework, which models the statistical correlation between the simple low-level feature. Combining spatio-temporal relationships with biological visual system features to assess the target location of incredible figure in the new frame. The Fast Fourier Transform(FFT) is adopted for fast learning and detecting in this work. In addition, a scale tracking method based on the log-polar coordinates transform was investigated to resolve the problem of scale change object tracking.