A Region-Based Stereo
Hiroshi Katsulai and Hirotaka Niwa
Department of Computer Science, Chiba Institute of Technology, 2-17-1 Tsudanuma, Narashino, Chiba, 275 Japan
The stereo, which is a method of obtaining depth information of scene from images obtained from at least two different directions, plays a very important role in applications to robots and similar equipments. The most difficult task in stereo method is to match individual parts of a two-dimensional projected image to those of another image.1,2) With respect to the method of matching, many studies have been conducted, and various techniques have been proposed. The stereo method based on features has attracted attention in recent years. However, it often fails in matching parts when attempting matching using points as features3) because it is difficult to specify points in images. On the other hand, proposals have been made for using line segments, which are easier than points to extract, as opposed to points for matching individual parts.6) Furthermore, methods have been developed which use regional features as an extension of the method which uses line segments.7) The method that uses regional features is considered to have a higher probability of success in matching images than the methods that use points or straight line segments because regional features contain a relatively large amount of description. However, no sufficient studies have been made yet on the region-based stereo. This study situation makes it necessary to conduct basic studies on regionbased stereo. This paper employs regions as features for matching, describes the stereo algorithm that directly employs region segmentation, and investigates the appropriateness of the algorithm by means of computer simulations. It is assumed that the three-dimensional object is a polyhedron, and each face of the object is projected onto a two-dimensional projection plane with uniform brightness using central projection. Region segmentation is delicate and does not necessarily ensure stable results. However, it is considered that a pair of two-dimensional projected images does not contain very large differences if the same scene is to be observed from slightly different directions. This paper uses the centroid of region, which represents the position of region, region shape, and the gray level of region as features for matching. Some consideration is taken on the matching technique to increase the accuracy of matching by performing an operation that is almost equal to enumerating all regional elements, using the sum of the similarity values of regional features as the evaluation function. A three-dimensional plane can be calculated from two matching regions by matching the two boundary points at the same height in the two projected images.