Recognition of Polyhedra from Range Data

T he ultimate goal of machine/computer vision system based on planar-faced objects using range data emresearch is to develop and engineer machines ploying a modified version of the Hough transform on a with the ability to sense, understand, and act VAX 11/780. The results are encouraging. However, our upon their environments in an autonomous vision system also points to the limitations of the current manner. Machines with limited and specialized vision capavision systems compared to the ideal general-purpose vision bilities are already in industrial use. However, building a system and indicates directions for future research. general-purpose machine/computer vision system is still far off in the future. Overview. The use of range data for object recognition is An intermediate and possibly more immediate objective a relatively new field of computer vision. To make the in developing a vision system would be the machine's abilpresent article self-contained, a brief review of the relevant ity to sense the three-dimensional structure of objects literature is in order. Several techniques have been demonplaced in the field of view of the sensor and to recognize strated for extracting primitives (e.g., planes, cylinders, the objects from a library of a "limited" number of obedges, vertices, circles) from the range image to construct jects. The characteristics of the possible objects will be descriptions of the objects in the scene.49 Underwood and stored in a database in the machine. The complexity and Coates'" developed a program that used a series of two-didifficulty of this problem will depend upon the nature of mensional views of a polyhedron to construct a 3-D model. the objects, the dissimilarity of the objects, the nature of The model could then be matched to a single 2-D view of sensing, the availability of the computing resources, and the polyhedron. Bhanu" introduced a method using stothe type of the recognition techniques. chastic labeling techniques to match faces of an unknown To obtain a better feel for the problem, the reader may view with faces of the model. Another method is to segconsider the following possibilities. The objects may be ment the scene into planar and curved surfaces and to asplanar faced or curved; several of the objects may be of sociate with each surface a set of 2-D properties (e.g., area, the same type, as in the case of a bucket of nails; the oblength of perimeter, mean radius, etc.) and relationships jects may be selected from a finite number of dissimilar between each surface (e.g., angle of intersection between types; the sensor may be one or more video cameras or a adjoining planes). This method was demonstrated by Oshrange sensor; the available computer may range in computima and Shirai.'2 These relationships between surfaces can ing power from an IBM-PC, to a VAX 11/780, or to a be represented as a graph type structure and the matching Cray XMP; and finally, the algorithms may be based upon process as the graph/subgraph isomorphism problem'3 statistical theorems, Hough transform procedures, or rulewhich is known to be NP-complete.'4 Another method based results. demonstrated by Magee et al.'5 uses an intensity image to The research of the past 20 years has addressed the issue guide the use of a range image, and it computes all of the of building a vision system with a variety of sensors and aldistances between vertices in the image and attempts to gorithmic tools. It is difficult, if not impossible, to review match those with the models. Similarly, circles were all the past research here. Reviews by Aggarwal et al.' and matched by the distance between centers of the circles and Besl and Jam2'3 document the state of the art rather well. In by their radii. A distinction can be made between these two fact, one of the papers by Besl and Jamn3 is fairly encymethods since the Oshima-Shirai method is view-dependent clopedic. The present article describes a particular vision (a different model is generated for each unique view of the