Determining 3-D structure of scene from image sequences obtained by horizontal and vertical moving camera

Abstract

The stereo vision is one of the simplest methods for acquiring three-dimensional (3-D) information. A problem then is that sometimes it is difficult to establish a correspondence between the left and the right images, in such cases as (1) multiple correspondence, (2) occlusion, (3) position reversals, and (4) horizontal edge. This paper proposes a method to determine the three-dimensional structure of the scene from the sequences of images obtained by the moving camera. As the first step, a two-dimensional image is constructed from the sequence of images, in which the apparent locus of motion is represented as segments. The correspondence problem in this step is simplified as the detection of segments on the synthetic image. By examining the relations among the detected segments, the occlusions can be detected, and the correspondence can be established where the positional reversals happen. By moving the camera not only in one direction, but also in orthogonal directions, the unique correspondence can be established, independently from the direction of the edge. An input system was constructed, which can accept a large number of motion stereo image sequence with a high speed, and some complex three-dimensional structures of the scenes were actually determined.