Search

To recognize objects in a scene, a single view is sufficient if the scene is relatively simple consisting of easily recognizable objects or the image was taken from a good viewpoint. In other situations, a multiple views approach should be employed. A multiple view approach can overcome these difficulties of single view approach by taking additional images when initial images were not sufficiently informative. The images are acquired from selected viewpoints which will reveal some new objects or new aspects of the same object to help recognition rather than from r and om viewpoints, since more images would be needed if the viewpoints are selected r and omly and analyzing more images would be costly. This will be called active vision since it actively moves to selected viewpoints and acquires more informative images to accomplish the vision task. Active vision has received little attention in vision research in spite of its importance. The multiple view approach needs a single view object recognition as a subprocess which recognizes clearly visible and recognizable objects in each image. The single view object recognition process analyzes each image frame to supply the information on recognizability of an object, and the possible identities and poses of the object if not recognizable. The multiple view object recognition process accumulates this information in an accumulator which functions as an environment model and is used in selecting viewpoints. At the end of each frame, if any object is still unrecognized, then a hypothesis is selected from the accumulated information and a new viewpoint is selected based on the hypothesis. The viewpoint is selected such that it shows features of the hypothesized object clearly. Another image frame is acquired from this new viewpoint. The process continues until all objects are recognized. Two viewpoint selection algorithms, direct-down and vector-sum algorithms, are developed. The power of the active multiple view vision is demonstrated using synthetic images of polyhedral objects. We obtained views which show features of objects clearly using the viewpoint selection algorithms.