A binocular vision system for 3-D object recognition is proposed. The objects to be recognized are assumed to have line features on their surfaces. Line segments are used as the units for object surface representation. A line correspondence algorithm is designed to find corresponding line segments in binocular images. Such line pairs are used to find corresponding corner points on the object surface, by which 3-D data can be computed. They are then used for shape matching which consists of 3-D registration and similarity comparison. The 3-D registration is accomplished by an extended line-based generalized Hough transform algorithm. And the similarity measure is defined in terms of the overlapping line segment length of two compared shapes. Experimental results show the feasibility of the proposed approach.
Information Science and Engineering (now Journal of Information Science and Engineering) 2 (1): 99-123