Study On Video Segmentation For Object-Based Coding

Object-based coding method is introduced by the second generation video coding standard MPEG-4, which enables content-based interactive functionality and high coding efficiency. Object-based representation translating original video sequence into a more abstract and compact form, makes higher level video analysis and comprehension possible, which is a key from video processing to video analysis. Video object segmentation is a precondition of object-based coding and representation. Therefore, it is of great significance in theory and applications.Based on the correlative research at home and abroad, research in this dissertation includes two aspects: spatio-temporal segmentation scheme and object tracking segmentation scheme. And four video segmentation algorithms are proposed according to the two schemes. An object-based video coding system is implemented, which combines the joint processing of VOP segmentation and video coding. The major contents in the dissertation are as follows:A spatio-temporal segmentation method based on high-order statistics (HOS) significant test and max-tree filtering is proposed. Firstly, a high-order statistics hypothesis testing in inter-frame difference is used to automatic separate moving areas from background. The covered/uncovered background is removed by symmetrical difference. Watershed segmentation is implemented after the image is simplified by max-tree filtering. Finally, a combination of the spatial and temporal segmentation masks by projection from spatial segmentation to temporal segmentation produces VOP.An accurate method of object contour extraction based on morphological watershed algorithm is proposed. Video object coarse shape and location can be obtained according to moving edge matching. Foreground and background markers are constructed based on the coarse shape. Video object closed contours are obtained using watershed algorithm based on color multi-scale gradient reconstruction. It can detect new object appearance and old object disappearance.Object tracking is handled as a classification problem and implemented through an adaptive hierarchical Support Vector Machines (SVM) classifier, which can overcome essential limitation of conventional motion-based tracking algorithms. Feature vectors are extracted not only from local feature by DCT but also from neighbor feature by entropy. An adaptive hierarchical SVM binary decision tree is used for classification of foreground and background.A new video segmentation method which consists of motion object detection, object tracking, model update and contour extraction by watershed is proposed. Firstly,initial moving object is extracted by block-based affine dominant motion classifier that combines the benefits of the dominant motion and block-based affine modeling approaches. Video object motion is divided into two components of slow change and fast change, corresponding to rigid and nonrigid motion to track respectively. Moving object edge detection captures emerging moving component for fast change. The robust Hausdorff tracker can accommodate rotation and change in shape of the tracked object for slow change.An object-based video compression system is developed, which realizes the joint processing of VOP segmentation and video compression. Experimental results demonstrate that object-based video coding by our segmentation can achieve high compression ratio and acceptable image quality. Above all, it implements the object-based access of video content.