Automating the Extraction of Revolution Objects from Single Laser Scans of Architectural Scenes

This paper focuses on the reconstruction of surfaces of revolution. We propose a method for estimating those surfaces from point clouds. We are more particularly dealing with data coming from one unique laser scanner acquisition; indeed, our approach extensively uses the range image topology. It makes originally use of geometric properties of surfaces of revolution. We firstly get an approximation of the axis of revolution, thanks to a Hough transform of symmetry axis computed on range image contours. Robust and accurate axis estimation refinement of position and orientation is then reached by minimizing the rms error of the profile estimate. Image topology is also used to estimate the object’s profile. 3D points are connected following the image topology and the nearest neighbouring planes rotating around the axis of revolution. After projection on these planes, lines connecting points allow easy profile estimation. This method solves the measuring noise issue which is usually sorted out through moving least squares solutions. Once the object of revolution has been reconstructed, we retrieve automatically all occurrences of this pattern in the scene, with a scale adaptative matching algorithm.