Shape Recognition In 3d Point-Clouds

While the recent improvements in geometry acquisition techniques allow for the easy generation of large and detailed point cloud representations of real-world objects, tasks as basic as for example the selection of all windows in 3D laser range data of a house still require a disproportional amount of user interaction. In this paper we address this issue and present a flexible framework for the rapid detection of such features in large point clouds. Features are represented as constrained graphs that describe configurations of basic shapes, e.g. planes, cylinders, etc. Experimental results in various scenarios related to the architectural domain demonstrate the feasibility of our approach.