A cybernetic multiresolution system for disparity estimation in stereo vision

Abstract

PurposeThe aim is to obtain a dense and reliable disparity map from a stereoscopic pair under any luminosity condition. From the disparity map it is possible to obtain some information about the objects appearing in the scene, such as their position and their distance to the camera.Design/methodology/approachThe disparity map is obtained using a correspondence method based on a multiresolution scheme. It also uses the simulated annealing algorithm to minimize the energy function, which integrates information coming from several sources: grey levels, non‐parametric transforms, edges and geometrical constraints. The multiresolution scheme allows us to interpolate the disparity map obtained in each level and to use it as an initial estimation for the simulating annealing process in the following level.FindingsThe multiresolution scheme speeds up the convergence. The fact of adding information about the neighbourhood improves the results. A simple and suitable energy function has been proposed. A simple efficient scaling scheme has been used. Their results are as good as other costly techniques such as block‐to‐point.Research limitations/implicationsThe use of grey level images has the drawback that some bright areas can be confused. As a solution, some other additional features are proposed.Originality/valueA new metric is proposed to evaluate the quality of a disparity map. A low‐cost energy function is proposed. It integrates several type information to add robustness to the method. The convergence of the method is dramatically speeded up reusing the results of the algorithm thanks to the multiresolution scheme.