An accurate and robust stereo matching algorithm with variable windows for 3D measurements

This paper describes the application of a stereo algorithm with novel variable windows to the measurement of three-dimensional (3D) shapes. Due to heavy variations of disparities within a rectangular window, previous area-based stereo algorithms have given rise to relatively large matching errors around depth discontinuities. In this paper, a new correlation function based on a robust estimator is formulated to improve matching accuracy in the area with depth discontinuities. Four-direction line mask including a rectangular window is devised not only to reduce noise effects but also to improve matching accuracy. The connectivity of match candidates selected by the correlation function is investigated to include the physical sense that disparities near an interested pixel are constant, and then integer disparities are estimated by finding the disparity plane with maximum connected match candidates. A novel variable window derived from a set of maximum connected candidates with positive intensity-gradient based similarity is used to minimize matching ambiguity of existing area-based methods. In addition, sub-pixel disparity estimation is performed to reduce local deformation effects of the variable windows resulting from projective distortion. Experiments confirm that the proposed method estimates accurate and robust results in the vicinity of depth discontinuities as well as in continuously smooth areas. (C) 2004 Elsevier Ltd. All rights reserved.