Research On Active Matting-Based Object Contour Tracking

Object tracking is to continuously find the trajectory of a specific target from a video sequence recorded by the camera. As an important research topic in computer vision, it promotes the theory and practical applications in image processing, pattern recognition and artificial intelligence. The main difficulties in object tracking is the similarity between the target and background, the object appearance variances caused by the external factors, such as pose and illumination as well as shading. In view of drifting issue in object tracking, the classical methods are prone to model degradation due to the inaccurate appearance modeling. Benefit from the image matting that can obtain the precise object contour, we study and implement an active matting-based object contour tracking algorithm.This thesis includes the following three aspects:(1) We discuss the importance and difficulties of object tracking and introduce the image matting theory. Furthermore, we summarize and analyze the advantages and disadvantages of existing methods of the object appearance modeling, contour tracking and image matting, which provides theoretical foundation to the object contour tracking.(2) We establish the relationship between image matting and object tracking. The foreground and background labels are automatically obtained by matting in the video sequence. We improve the conventional LOG interest point detector based on the color surface coding theory. Then we transform the motion estimation of successive frames into robust interest point matching to complete the accurate appearance modeling of the color sequence. Therefore, we achieve the color surface active matting-based object contour tracking.(3) For the issues of the large deformation and fast moving of the target, SIFT flow algorithm is adopted to locate the interest points. To promote the robustness of the contour tracking in the complex videos, we introduce a new template update strategy to improve the performance of the object tracking algorithm based on the optical flow.