Abstract: Three-dimensional modeling of a scene from theautomatic analysis of 2D image sequences is a bigchallenge for future interactive audiovisual servicesbased on 3D content manipulation such as virtual vests,3D teleconferencing and interactive television. Wepropose a scheme that computes 3D objects models fromstereo analysis of image triplets shot by calibratedcameras. After matching the different views with acorrelation based algorithm, a depth map referring to agiven view is built by using a fusion criterion takinginto account depth coherency, visibility constraintsand correlation scores. Because luminance segmentationhelps to compute accurate object borders and to detectand improve the unreliable depth values, a two stepssegmentation algorithm using both depth map andgraylevel image is applied to extract the objectsmasks. First an edge detection segments the luminanceimage in regions and a multimodal thresholding methodselects depth classes from the depth map. Then theregions are merged and labelled with the differentdepth classes numbers by using a coherence test ondepth values according to the rate of reliable anddominant depth values and the size of the regions. Thestructures of the segmented objects are obtained with aconstrained Delaunay triangulation followed by arefining stage. Finally, texture mapping is performedusing open inventor or VRML1.0 tools. !11
展开▼
