A feature point based scheme for unsupervised video object segmentation in stereoscopic video sequences

The video coding standard MPEG-4 is enabling content-based functionalities by the introduction of video object planes (VOP's) which represent semantically meaningful objects. In this paper, a novel fast, unsupervised semantic segmentation scheme is presented for stereoscopic sequences, which utilizes the provided depth information. Each stereo pair is first analyzed and the disparity field and occluded areas are estimated. Then a multiresolution implementation of the RSST segmentation algorithm is applied to the depth map for extracting the depth segments. For each depth segment, except the last. feature points are generated on its contour and a motion geometric space (MGS) for every initial point is defined. Afterwards one point per MGS is selected, which satisfies predefined intensity and curvature constraints so that the object boundaries are accurately extracted. Experimental results are presented to indicate the good performance of the proposed scheme on real life stereoscopic video sequences.