Research On Capturing Structure Priors And Property Of Structure Preserving In Object Tracking

Object tracking plays a key role in Computer Vision, because of its important theory significance and the strong application value. In object tracking, a typical algorithm composes of three components:appearance model, motion model and search strategy. Misclassification and mismatch causing by the appearance changes of object appearance are the greatest challenge. Therefore, a robust appearance model is the goal of researchers.In this thesis, we exploit structure information to enhance the descriptive power of appearance model. The structured model integrates video cues and inner structure of object in spatio-temporal domain. We then extend our structured model to multi-object task by introducing a structure preserving between objects. This thesis includes the following contents:(1) Based on appearance and motion model, we summarize fundamental theories, basic methods and main challenge in object tracking.(2) The structure relationship between object parts or objects was analyzed. We show how to describe the structure.(3) We introduce a structured super-pixel tracker that exploit the structure information by incorporating geometric constraints between the super-pixels of object based the cues of temporal and spatial. We use an undirected graph to model the super-pixels and their spatial position. The spatial position confidence help node to vote object state.(4) A structure prior model is presented to capture the data space association by establishing on-line updating. A dense sampling method is presented to infer space confidence. Using the well-established theory of circulant matrices, a fast learning method is achieved to infer the confidence for all possible position of target with the Fast Fourier Transform.