Cobb%MIT-OZ@MIT-MC.ARPA (04/18/84)
[Forwarded from the MIT bboard by SASW@MIT-MC.] W. ERIC L. GRIMSON Local Constraints in Model Based Recognition and Localization From Sparse Data April 23, 1984 4:00PM NE43-8th floor playroom A central characteristic of advanced applications in robotics is the presence of significant uncertainty about the identities and attitudes of objects in the workspace of a robot. The recognition and localization of an object, from among a set of models, using sparse, noisy sensory data can be cast as the search for a consistent matching of the data elements to model elements. To minimize the computation, local constraints are needed to limit the portions of the search space that must be explicitly explored. We derive a set of local geometric constraints for both the three degree of freedom problem of isolated objects in stable positions, and the general six degree of freedom problem of an object arbitrarily oriented in space. We establish that the constraints are complete for the case of three degrees of freedom, but not for six. We then show by combinatorial analysis that the constraints are generally very effective in restricting the search space and provide estimates for the number of sparse data points needed to uniquely identify and isolate the object. These results are supported by simulations of the recognition technique under a variety of conditions that also demonstrate its graceful degradation in the presence of noise. We also discuss examples of the technique applied to real data from several sensory modalities including laser ranging, sonar, and grey level imaging. Refreshments: 3:45PM Host: Professor Patrick H. Winston