Flexible computer vision systems for robotics and automation must be capable of recognizing three-dimensional objects in arbitrary positions and orientations from partial image data. Previous research by the author has led to the development of a system that can solve this task for a small library of possible objects. This project will extend these successful methods to much more general problems which will allow the system to work efficiently with large numbers of objects containing internal degrees of freedom. Research will continue on developing new methods for perceptual organization, in which image features are grouped into meaningful structures for matching to object models. Object models will incorporate variable parameters and articulations, and these internal parameters will be solved for simultaneously with the determination of viewpoint. Objects will be described as a hierarchy of models, in which components can be recognized and used to infer the presence of higher-levels. Low-level segmentation and perceptual grouping will be extended to allow for the detection of arbitrary curves. A major goal will be to develop an overall theory of recognition that can be applied to each level of the visual hierarchy. Such a theory would describe each level of recognition in terms of fitting over-constrained data to partial models, in which the evaluation of each level of recognition would be based on the use of similar statistical operations.
