Stereo Vision Based on Hierarchical Segmentation

Abstract

This paper proposes a new stereo vision method that can reconstruct 3D structures robustly by matching left and right images based on regions given by color segmentation. The proposed method enables correct matching of the regions by segmenting simultaneously stereo images into color regions and matching feature points on the edges of the same color regions. Furthermore, this method can reduce fault correspondences because the difference between segmentation results of right and left images is absorbed by carring out region matching using hierarchical color segmentation whose threshold to merge classes of color is changed step by step in order to obtain appropriate region correspondences independently of a select threshold value. This method can also detect and modify occluded parts by assuming that artificial objects consist of flat color regions approximately.