3D Scene Reconstruction By Stereovision: Modelling-Solving-Experimenting

A computer vision system is presented. It provides an accurate and reliable dense depth map of a scene viewed from an "anthropomorphic " binocular sensor. At first we control the vergence movements of the two cameras by selecting a scene-directed fixation point. A rough matching is gotten from a set of Gabor filters. This matching is improved by solving a constrained optimization problem; the cost function and the constraints code several properties of the triplet (physical world, sensor, actual scene): uniqueness-occlusions, compatibility-similarity, smoothness. In order to achieve this optimization, an hybrid method combining simulated annealing and tabu search is used. We have obtained very good results on complex scenes in different domains ( random-dot stereograms, polyhedral objects, Peter-Pan mask, ...). Matching errors less than 0.7% (random-dot stereogram) and depth relative errors of about 0.05% are reached. KEYWORDS stereovision , vergence movements , stereomatching , Gabor..