Recognition of multiple objects using geometric hashing techniques

This paper discusses the development of a robust robot vision system for implementation in a flexible assembly cell. The vision system is capable of recognizing the identity and returning the three-dimensional position and orientation of each object in a physical scene. The scenes of interest may consist of one or more (possibly occluded) industrial objects. A typical unstructured factory lighting environment is assumed. The resulting vision system is model-based and learns an object either through CAD data or by physically viewing the object via a black and white CCD intensity image. The geometric representation of the object is generated off-line and stored in a hash table. During recognition, the hash table is accessed to identify the object and return a transformation matrix encoding its three-dimensional pose. The hash table makes use of the straight- and parallel-line invariance properties of the affine approximation to the perspective transform, and allows for the bulk of the computational load to be shifted off-line. A feature point classification approach has been implemented which results in significant reductions in both hash table size and online recognition times. The latest version of the authors´ software recognizes relatively flat industrial objects in any three-dimensional configuration and from any viewing angle. Experimental results for multiple real-world objects (such as screwdrivers and pliers) in occluded scenes have resulted in recognition times of less than one second per object. In addition, this geometric hashing technique may be easily extended to the recognition of general three-dimensional objects