A robust stereo-matching algorithm using multiple-baseline cameras

Abstract

Calculating the distance of various points in a scene relative to the camera position is one of the important tasks in stereovision systems. A fundamental problem in stereovision is to find corresponding pixels, points, or other features in both the left and right images that are taken from stereo cameras. For precise measurement and correct matching of repetitive patterns, a multiple-baseline stereo (MBS) algorithm has been proposed (see Kanade, T., ARPA Image Understanding Workshop, p.549-58, 1994; Jeon, J., "A robust stereo matching technique using multiple cameras", Ph.D. Dissertation, Chosun University, Korea, 2001). We develop a robust multiple-baseline stereo (RMBS) algorithm to solve continuous, as well as repetitive, multiple local minima through an adaptive window. Experimental results with both synthetic and natural scenes demonstrate that RMBS is faster in processing speed than MBS by 24.1%. In addition, the proposed RMBS algorithm is robust to different types of multiple local minima.