Planning Multiple Observations for Object Recognition

Abstract

Most cornputer vision systems perform object recognition on the basis of the features extracted from a single image of the object. The problem with this approach is that it implicitly assumes that the availible features are sufficient to determine the identity and pose of the object uniquely. If this assumption is not met, then the feature set is insufficient, and ambiguity results. Consequently. much research in cornputer vision has gone towards finding sets of features that are sufficient for specific tasks, with the result that each system has its own associated set of features. A single, general feature set would be desirable.However, research in automatic generation of object recognition programs has demonstrated that pre-determined, fixed feature sets are often incapable of providing enough information to unambiguously determine object identity and pose. One approach to overcoming the inadequacy of any feature set is to utilize multiple sensor observations obtained from different viewpoints, and combine them with knowledge of the 3D structure of the object to perform unambiguous object recognition. This paper presents initial results towards performing object recognition using multiple observations to resolve arnbiguities. Starting from the premise that sensor motions should be planned out in advance, the difiiculties involved in planning with ambiguous information are discussed. A representation lor planning that combines geometric information with viewpoint uncertainty is presented. A sensor planner utilizing the representation was implemented, and the results of object recognition experiments performed with the planner are discussed.