Abstract

There is mounting evidence for believing that a select number of characteristic views of solid objects may serve as an adequate model for recognition tasks. This idea is examined in a constrained application domain involving machine part recognition and assembly. Methods of structural and syntactic pattern recognition are used to derive compact and efficient codes for 2-D views of geometric objects. The model data base stores information about characteristic views of objects and the visual events that lead to transition between these views. Such events denote change in relative position and orientation between sensor and object. The recognition method can be broadly described as exploratory, where data base search is conducted based on extracted 2-D features, and if necessary sensor movement is initiated to generate features leading to maximum disambiguation of the match set. The areas of investigation include syntactic representation of 2-D views, organization of hierarchical model data base, fast indexing into the data base using a hardware engine, and sensor guidance to achieve recognition with the minimum number of views.

© 1987 Optical Society of America