On locating objects by their distinguishing features in multisensory images

This paper reports preliminary work on a knowledge-based perceptual system for a robot that must function in an actual office environment. This system is distinguished by the following pragmatic considerations: (1) It is designed to find specific objects needed by the robot in the performance of a task rather than attempting the usually unnecessary and very much harder job of completely describing an environment of potentially overwhelming complexity. (2) It is based on the premise that in real scenes there is a sufficient redundancy of perceptual clues, as well as contextual constraints among objects, so that an intelligent system can devise a relatively simple strategy for distinguishing the specific objects of interest from others likely to be present. (3) It relies heavily on multisensory (i.e., color and range) data to increase the likelihood of finding distinguishing surface attributes for a particular object. Similarly, detailed descriptive representations for complex attributes (e.g., shape and texture) are avoided in favor of the simplest representations sufficient to distinguish the object of interest.