3D Shape Recovery from a Sequence of Stereo Images

摘要

A method is presented for recovering the 3D shape of objects from a sequence of stereo images.For each pair of images, the object region in each image is detected using an active contour and the object surface visible is estimated by computing a depth map for scene points in object regions. Scene point correspondence required for deriving the depth of the scene point is determined by computing zero-mean normalized cross-correlation using multi-resolution dynamic programming techniques. The object shape is described by a deformable surface model, which successively fuses surface patches computed from each pair of stereo images through deformation. Aligned with the input surface path according to the object or camera motion, the model is deformed by attracting its corresponding surface part towards the input surface patch. Object motion between successive pairs of stereo images is estimated by tracking a set of image features. Experimental results are also included to show the feasibility of our new method.