A Novel Stereo Object Segmentation Algorithm Using Disparity and Temporal Information

A novel stereoscopic object segmentation method based on disparity and temporal-spatial cues is proposed in this paper. First, a foreground layer map is generated from a patched disparity image and a motion object map is produced using an adaptive reference frame selection method. Then the intersection of these two maps is used to extract the foreground motion objects. In the foreground layer map generation stage, two intensity image sequences of the input stereo videos are separately mapped to the rank space by a rank transform to eliminate the parameter deviation of binocular cameras, and reduce the interference of the environment fluctuation and noises. A fast multi-window matching strategy is proposed to speed up the stereo matching process. Then, a disparity patching process is performed to bridge the disparity discontinuity in the smooth region to improve the accuracy of the disparity map. In the motion object map generation stage, an adaptive reference frame selection method is introduced to obtain salient motion and produce the motion mask of each view. Experimental results have indicated the good performance and the practicability of the proposed stereo segmentation algorithm.