Video Object Segmentation Based On Spatio-Temporal Information

The emerging multimedia standards MPEG-4 and MPEG-7 adopt content-based coding and description method, the content-based functionality was proposed, which includes content-based video compression, scalability, interoperability and so on. In MPEG-4, a video sequence is segmented into several meaningful video objects(VOs). Coding is performed on shape, motion and texture. MPEG-7 aims at describing all kinds of multimedia objects and standardizing multimedia interface universally to enable efficient content-based search and query. Thereinto, video object segmentation plays a critical role in realizing content-based coding and description, which has an immediate influence on coding efficiency and description validity. However, MPEG-4 and MPEG-7 just defined coding method and syntax principals, did not specify video object segmentation method. As a key supporting technique for video processing, studies on video object segmentation have far-going pragmatism significance and application importance. Towards this goal, this dissertation focuses on video object segmentation techniques.This dissertation introduces the applications and developments of video object segmentation techniques in aspects of multimedia standard, video surveillance and pattern recognition. We discuss its significance in current high-speed developing information epoch, and summarize research activities of video segmentation all the world, analysis interactive , automatic segmentation methods. On the basis of this, further and sufficient discussions are concentrated on several key techniques about object segmentation including spatial intra-frame segmentation, global motion estimation and compensation, temporal inter-frame segmentation and spatio-temporal projection. Meanwhile, a relative background sprite image reconstruction method is presented. Many series of simulated experimental results verified the proposed method of segmentation and background sprite reconstruction.Firstly, in aspect of spatial segmentation, an improved watershed strategy is presented in image pre-processing and region labeling to restrain over-segmentation in conventional watershed caused by noise and close textures , the main improvements are as follows: â‘  double opening and closing reconstruction is proposed to obtain morphological gradient, with the first one for restraining whole image noise and the second one for reducing light and dark details caused by textures. (2) a non-linear transform is proposed by integrating certain thresholding and scale grade classification, the former is to reduce the number of region minima and reserve main contours, and the latter is to overcome the influence of close textures inside objects. â‘¢ an improved watershed region labeling algorithm is proposed on the basis of pixels' connectivity, with the advantage of discarding distance transform. Experimental results illustrated that the number of regions by the proposed method is about one tenth of that by conventional method without post-process employment such as region merging, and exact edges are located.Secondly, as for motion detection for stationary background sequences, an idea of representing region motion with region's boundary information is created to resolve the problem existing in region-based method, this problem is that region-based method usually fails to identify motion due to the similar texture and gray inside the region. Only through Gaussianity test on boundary pixels and motion pixels ratio to whole boundary can motion regions be identified, which reduces computational complexity greatly and enhances the robustness of motion regions with similar texture and gray. Experimental results showed that the consumed time by proposed method was about one fifth of that by region-based method.Thirdly, Markov Random Field model on double scale neighborhood is established for motion detecting of dynamic background sequences, it breaks the restriction of the conventional single scale neighborhood to spatial relativity among pixels. Integrating gauss mixture distribution with MRF on difference images, a motion detecting model and convenient system energy function are proposed. Moreover, certain logic operation is performed on current and last binary motion mask to achieve static foreground regions.Finally, a video object segmentation strategy is implemented on the basis of spatio-temporal results. In addition, with regard to the sprite coding technique in MPEG-4, a background sprite image reconstruction method is realized accordingly based on the proposed video object segmentation method, a conditional mean composition principal is suggested to merging image, meaningful reconstruction result is attained in experiment.