A Robust Coarse-to-fine Least-squares Stereo Matching for Automatic Terrain 3-D Reconstruction

Least squares methods are widely used for digital stereo matching and 3-D reconstruction from stereo pairs. A geometric transformation is used to shift, rotate and stretch the right-image search area until fitting its left-image target area in the least square sense. However, preliminary knowledge of the disparity field between the two stereo images is critical for the performance of the algorithm. Robustness, as well as accuracy, may be enhanced when adopting a coarse-to-fine approach. In this work, a multi resolution representation (Gaussian pyramid) of the stereo images and application of the least squares matching at each pyramid layer is proposed. Convergence results to be quicker and less critical, due to the progressively decimated disparity field. Also computation is speeded up, as search areas are roughly decimated as well. Finally, the correlation coefficient is jointly employed at the full resolution level to detect mismatches caused by the presence of local minima of the residual, thus reducing the mismatch error probability.