Direct Solution of the Least Squares Matching Problem

摘要

This work presents new approach to matching two images in the case of uncalibrated stereo rig. It is based on the direct use of well known in photogrammetry Least Squares Matching (LSM) technique. The developed algorithm (DLSM - Direct Least Squares Matching) uses modification of the LSM where the disparity map inside each matched window of the reference image is described by the affine transformation. This model accounts local disparity gradients and ensures efficiency of the DLSM for wide baseline stereo. The linear model is used to describe intensity transformation. The LSM is a nonlinear procedure that requires initial approximation of the estimated parameters for each matched window. Here we don't follow the usual practice based on utilization of the intermediate matching algorithms for obtaining such initial approximations. Our strategy is based on prediction of affine and intensity parameters and utilization of the recursive filling procedure to propagate solution of the least squares matching problem on the connected textured areas of the reference image where a smooth disparity map exists. Experimental results on real images are presented to illustrate the robustness and reliability of the DLSM in the case of wide baseline stereo and for matching images containing periodical textures.