Research On MPEG-4 Based Video Segmentation Algorithm

MPEG-4 was published in 1999 by ISO and IEC, it is the first image coding standard aimed in content-based usage and also paved a way for image coding in the future. In order to fulfill the content-based functions, the image sequences must segment into video objects before coding, Accurate and general video segmentation is an unsolved problem now. Video segmentation was described in annex F of the MPEG-4 standard, Even there are many video segmentation algorithms in the text of the annex F, it was not an algorithm specified the use of video segmentation, that is, more algorithms which have better ability are expected to provided by researchers. On June, 2002, a work team AVS for audio-visual coding standard which will issue by China was established, the experts in video team approve video compression was founded in MPEG-4 profile, but there must be video segmentation first. For all reasons mentioned above, I select this topic in the thesis.Video segmentation can be divided into two classes: Automatic video segmentation and semi-automatic video segmentation. They have some same processing steps, like camera motion compensation, image gradient estimation, image border extraction, intra-frame segmentation, inter-frame segmentation, combination of intra-frame segmentation and inter-frame segmentation results. In this thesis, I focus on semi-automatic video segmentation, had researched on the topics of gradient estimation, border extraction, intra-frame segmentation and motion analysis in inter-frame segmentation:(1)The subtraction of pixel gray levels is an important survey variable, it will be used in many steps of video segmentation, the gradient can properly estimate the variable. In this thesis, I seriously analyze the problems presented in the gradient estimation, and give an algorithm on gradient estimation. The accuracy of the gradient values and the detection of gray level changes in horizontal, vertical and diagonal directions can have better results than traditional gradient estimators.(2)Border extraction are based on edge detection, Precise, easy coding completion and robust are the objects in edge detection for video segmentation. The researches of edge detection were beginning with gradient-based methods, I studied a confidence-based edge detector which was combined of gradient-based edge detector and template matching method, in the critical step, hysteresis threshold, from the experiments and research, I prepare the shape of the threshold curves for delineating the object and the background of which the gradient values were not for away.(3)In the intra-frame segmentation step, I had researched the mean shift based segmentation method. Mean shift is a nonparametric and robust clustering method, it is attracting to use mean shift in image processing topics. In mean shift based image segmentation, spatial domain bandwidth h_s and color range domain bandwidth h_r are two important parameters, they decide the resolution of spatial domain and color range domain in the segmented image. In this thesis I had researched the choice of h_r for the heavily textured and wide color contrasted images, I concluded the range of h_r from the experiments.(4)Inter-frame segmentation step was accomplished by doing motion analysis on video objects. I described a point correspondence algorithm in the text, the algorithm combined two methods, optical flow and color density distribution, it also took advantages of them. Then, I used this algorithm to do experiments for video sequence, the experiments verified its abilities and disadvantages.Finally, based on the four steps, gradient estimation, edge detection, intra-segmentation and object tracking inter-segmentation, which are described in the thesis, I gave a semi-automatic segmentation method and stated the completion steps in principle.