Use of the surface-attribute probe for 3-D object characterization

This paper deals with the classical problem of 3-D object recognition, in which we are asked to establish unique correspondence between an arbitrarily oriented object apparent in an image and one of a set of known model objects, or to demonstrate that no such correspondence exists. We have developed a “smart” approach to this problem, which permits us to extract only as much feature data as is needed to perform the recognition task rapidly and efficiently. The object is assumed illuminated with patterned light, and both edge and surface-curvature features are extracted in a mutually reinforcing manner from selected regions of interest in the image. An “active model” is hypothesized and progressively refined by extracting feature data from an increasing number of judiciously chosen image regions by means of a so-called surface-attribute probe (SAP). The novel features of the method lie in the use of selectively placed, variable-size windows to extract image features in a progressive refinement manner. The method alternates between using surface-discontinuity features (edges) to guide the extraction of surface-curvature attributes, and using surface-curvature attributes to refine the extraction of surface-discontinuity features. The scheme handles planar as well as curved-surface objects, minimizes unproductive activity, and lends itself to parallel-processor implementation