Detecting object surfaces by using occlusion information from active binocular stereo

We propose a method to detect polyhedral object surfaces by using “occlusion information” from active binocular stereo. Occlusion information announces whether each feature point in the scene is visible or not from each viewpoint from which an image has been taken; this is a useful geometric cue to suppose the existence of object surfaces. It is very difficult for conventional stereo methods to obtain this information. Two complementary algorithms based on occlusion information decide whether or not there is a surface between each hypothetical 3D segment called “arc” which corresponds two successive image feature points along an epipolar line. The final surface description of the proposed method is a set of these arcs classified into three states: 1) there is a surface, 2) there is no surface, 3) it is unknown whether or not there is a surface. Our method can reliably detect polyhedral object surfaces that are visible and occlude other objects in the scene. Experimental results of the proposed method applied to the actual scene are also presented