Research On Video Object Segmentation And Stereo Disparity Estimation

With advances in communication and information processing technologies, a lot of digital applications and services are emerging such as: digital TV, teleconference, videophone, interactive multimedia and stereo video. These diversified applications and services with a large amount of data demand more advanced digital signal processing technologies for efficient transmission and store. Video object segmentation and stereo image processing are hot topics in this field.The thesis focuses on the research of video object segmentation and disparity estimation. Firstly, a new spatio-temporal algorithm for video object extraction is proposed. It combines edge information and motion detection in twice-frame difference. Experimental results show that this method can effectively extract the video object from video sequence. Secondly, Delaunay triangulation mesh model is introduced to the image segmentation field. Improved DT model is used to extract"feature points"of image adaptively. Then the image object is segmented by local nodes density analysis and some basic image processing methods. It is creative because DT model was never used in image segmentation before. Finally, after making a deep research on stereo vision and the present methods of disparity estimation, the author designs a novel disparity estimation algorithm based on gradient and Markov Random Field (MRF) model. Block matching algorithm combining gray and gradient information is used to get initial disparity field. Then erroneously matched blocks are detected based on the order matching constraint and corrected iteratively by MRF-based causality prediction to achieve more accurate disparity field.