A hybrid stereo vision method using a correlation function and variable window based on voxel components

Abstract

Existing stereo methods tend not to deal with depth discontinuity due to area-based problems, and so give inaccurate and noisy three-dimensional (3D) information in the surface morphology with steep disparity variations. In this paper, a hybrid stereo approach has been presented to estimate accurate, detailed and smooth information for the reconstruction and inspection of the surface morphology with both depth discontinuity and curved regions. It reduces the smoothing problems resulting from depth discontinuity by evaluating a new correlation function based on voxel components in the 3D disparity space. In addition, the connectivity of match candidate voxels is investigated to improve robustness to noise and then integer disparities are determined in the updated disparity space. A variable window based on voxel components is newly devised to cope effectively with curved regions as well as depth discontinuity. Sub-pixel disparities are accurately estimated through formulating the optimal error function in the variable window. We show how our results improve upon those of closely related techniques in terms of accuracy, robustness, matching density and computing speed.