Abstract
The paper describes a method for recognition of an overall shape of a polyhedral object. The method utilizes multiple stereoscopic image pairs of the object which consist of conventional half-tone images representing different views of the object. For saving processing time, necessary images are selected automatically and dynamically by use of a geometrical assumption on objects and calculation on images. Through preprocessings, geometrical features, such as vertices, connected edges and surfaces, are extracted from the original images, and described by graphs including their connective relations. An approach utilizing the following three constraints is proposed for solving correspondence problems : (1) geometrical constraints on vertices, (2) topological relations of connected edges, and (3) those between corresponding surfaces. By applying stereoscopic calculations to each pair of images, three dimensional locations of vertices which are involved in the respective field of view can be obtained. Then an overall shape of the object can be reconstructed by identifying common features emerging in different pairs and merging them into a descriptive graph of the overall shape of the object. Experimental results are shown to insist effectiveness of the method.
