Abstract
This paper describes modeling of ambiguous matchings in stereo vision and a viewpoint selection for resolving the ambiguities. Many iterative or active methods have been proposed for determining matches between features. These methods, however, do not consider the cost of recognition. In general, visual recognition of an environment requires much computation and the recognition result includes uncertainty. A trade-off, therefore, must be considered between the cost of visual recognition and the effect of visual information to be obtained. In order to consider such a trade-off, a new segment-based stereo matching method is proposed. The initial matches are established by using a local disparity histogram. All matches including ambiguous ones are kept with their reliabilities in the model. The ambiguities will be resolved by subsequent observations if necessary. An efficient method is proposed for determining whether a next viewpoint can resolve an ambiguity. Experiments with real data show the usefulness of the method.
