Deformable Object Tracking Based On Max-pooling Graph Matching

With the development of computer technology and networks, computer vision technology has become an important topic in the field of information science in the age of big data. As the basis for many high-level computer vision applications, visual target tracking has received more and more attention by domestic and foreign researchers. Depending on the application, visual target tracking mainly divided into two major directions:single target tracking and multi-target tracking. Although researchers have done a lot of study on single target tracking, all kinds of information and the scene limiting has not been fully exploited during the moving process of the target. The target may have large deformations and severe occlusions during the process of tracking. If it happened, the appearance of the target will be changed dramatically. In this case, if we continue to use the traditional bounding box to describe the target, we surely filter out some parts of the target or pull in background noise. We develop a novel deformable object tracking algorithm based on max-pooling graph matching, which can be applied in the scenes with large deformations and severe occlusions. The main content is concluded as follows.(1)We exploit the inner geometric structure information of the target to design a dynamic graph representation. This structure information is generated by oversegmenting the target into several parts(superpixels) and then modeling the interactions between neighboring parts. Both the appearances of local parts and their relations are incorporated into the dynamic undirected graph.(2)We choose max-pooling strategy for graph matching. Max-pooling strategy evaluates each candidate match using its most promising neighbors, and gradually propagates the corresponding scores to update the neighbors. As final output, it assigns a reliable score to each match together with its supporting neighbors. Based on max-pooling graph matching method, the matching relations between target parts and candidate parts are found to calculate the confidence map of target location.(3)Considering both the support of the holistic target and local parts, the optimal target location can be determined.Compared to state-of-the-art methods, experimental results on several deformable sequences demonstrate the effectiveness and robustness of the proposed method.