Automatic reconstruction of large 3D models of real environments from unregistered data-sets

Towards photo-realistic 3D scene reconstruction from range and color images, we present a statistical technique for multi-modal image registration. Statistical tools are employed to measure the dependence of two images, considered as random distributions of pixels, and to find the pose of one imaging system relative to the other. The similarity metrics used in our automatic registration algorithm are based on the chi-squared measure of dependence, which is presented as an alternative to the standard mutual information criterion. These two criteria belong to the class of reformation-theoretic similarity measures that quantify the dependence in terms of information provided by one image about the other. This approach requires the use of a robust optimization scheme for the maximization of the similarity measure. To achieve accurate results, we investigated the use of heuristics such as genetic algorithms. The retrieved pose parameters are used to generate a texture map from the color image, and the occluded areas in this image are determined and labeled. Finally the 3D scene is rendered as a triangular mesh with texture.