Robust 2D-3D alignment based on geometrical consistency

This paper presents a new registration algorithm of a 2D image and a 3D geometrical model, which is robust for initial registration errors, for reconstructing a realistic 3D model of indoor scene settings. One of the typical techniques of pose estimation of a 3D model in a 2D image is the method based on the correspondences between 2D photometrical edges and 3D geometrical edges projected on the 2D image. However, for indoor settings, features extracted on the 2D image and jump edges of the geometrical model, which can be extracted robustly, are limited. Therefore, it is difficult to find corresponding edges between the 2D image and the 3D model correctly. For this reason, in most cases, the relative position has to be manually set close to correct position beforehand. To overcome this problem, in the proposed method, firstly the relative pose is roughly estimated by utilizing geometrical consistencies of back-projected 2D photometrical edges on a 3D model. Next, the edge-based method is applied for the precise pose estimation after the above estimation procedure is converged. The performance of the proposed method is successfully demonstrated with some experiments using simulated models of indoor scene settings and actual environments measured by range and image sensors.