Stereo Video Object Segmentation Based On Disparity Estimation

New video standards, such as MPEG-4 and MPEG-7, do not concentrate only on efficient compression methods, but also on select better ways to represent, integrate and exchange vision information. Video object segmentation has more and more application in video coding, pattern recognition, and video index. However, the 2D video signal by description of plane image has not satisfied the needs of exchange and depth layer. Compared with 2D video signal, the stereo video signal has the more efficiency expression of vision. In addition, Segmentation of video objects becomes more effective in stereoscopic video because depth information can be reliably acquired. In recently years, new algorithm has been proposed continuously, such as based on the fusion of color and depth segments and on active contour. In this paper a stereo video object segmentation method combined depth and edge was proposed, and in which improved watershed algorithm is used.In order to acquire precise depth information which is indispensable to segmentation of stereo video, so a stereo matching algorithm based on adaptive weight was applied in this paper. In the method, rather than giving support-window size for each pixel, the support-weight of each pixel was adjusted according to the similarity and proximity between the pixel and their neighbors in a given window. A pixel from the same homogeneous regions as the pixel under consideration would have larger weight. Moreover, the disparity smoothness constraint term is introduced in matching cost function, and the disparity map was acquired.Because traditional watershed algorithm is sensitive to intensity change, a modified watershed algorithm was used to achieve initialize segmentation of disparity map in this paper. Although depth segmentation can provide a coarse separation of the semantic objects, precise contour of the object can hardly acquired due to occlusion in stereo matching. So edge information on coarse object area is detected to get precise contour of the semantic object. At last, an efficiency motion tracking algorithm based on area was used to extract object of adjacent frames quickly. The experimental results show our method can segment video object precisely.