Object segmentation by fusion of multi-cues in stereoscopic sequences

An algorithm for object segmentation from stereo sequences based on fusion of multi-cues of edge, disparity, motion and color is presented in this paper. Firstly, the accurate disparity field is obtained using a two-level disparity matching method based on image edge information. The morphological operators are then performed on the given disparity field to obtain coarse objects segments. "Split and merge" process is applied to extract the objects regions, and "erosion and dilation" process is used to fill some small inner holes in the target regions or smooth the discontinuous regions. On the other hand, spatial-temporal segments are obtained with image edge structure and motion change detection. Different object boundaries can be articulated according to disparity and spatial-temporal segments. At last, the multi-objects are extracted by further fusion of the color information. Experiments indicate this algorithm is an effective method for segmenting multi-objects overlapped each other from stereoscopic video that usually is difficult to be done in the case of monocular video.