3D modeling of urban areas using plane hypotheses

This paper proposes an efficient plane hypothesis matching technique for 3D mapping of urban environments using images obtained from a moving monocular camera. The algorithm is based on the assumption that urban environments are generally composed of buildings that have planar facades, and these facades are placed in the direction of gravity. A sparse 3D point cloud of the imaged scene is obtained using the classical Structure from Motion technique, and then the plane hypotheses are obtained by running an iterative Hough Transform on the 2D point set that is obtained from the projection of these 3D points in the direction of gravity. Superpixels are preferred instead of pixels for matching the image to the plane hypotheses. The superpixels are assigned to the plane hypotheses using their 3D point associations. As a result, a dense depth map of the urban scene is constructed successfully by means of the planar patches.