Cluster Analysis and Priority Sorting in Huge Point Clouds for Building Reconstruction Wolfgang von Hansen Eckart Michaelsen Ulrich Th ¨ onnessen FGAN-FOM, Gutleuthausstr. 1, Ettlingen, Germany E-mail: wvhansen@fom.fgan.de Abstract Terrestrial laser scanners produce point clouds with a huge number of points within a very limited surrounding. In built-up areas, many of the man-made objects are dom- inated by planar surfaces. We introduce a RANSAC based preprocessing technique that transforms the irregular point cloud into a set of locally delimited surface patches in order to reduce the amount of data and to achieve a higher level of abstraction. In a second step, the resulting patches are grouped to large planes while ignoring small and irrelevant structures. The approach is tested with a dataset of a built- up area which is described very well needing only a small number of geometric primitives. The grouping emphasizes man-made structures and could be used as a preclassifica- tion. 1. Introduction In addition to photogrammetric imaging, laser scanning has become a major data source for the acquisition of 3D city models for tourist information or the documentation of cultural heritage. Airborne systems are widely used but also terrestrial laser scanners are increasingly available. They provide a much higher geometrical resolution and accuracy (mm vs. dm) and are able to acquire building facade details which is a requirement for realistic virtual worlds. How- ever, the operating area is limited due to occlusion and by the maximum range of the laser beam. It is not possible to capture an extended area from one position alone, leading to typical raw datasets consisting of several overlapping, huge point clouds. The overall objectives are • generalizing the point cloud to higher level geometric primitives and grouping them to objects for data reduc- tion and as preparation for high level cognitive tasks. • fusing the datasets and coregistering them into a single geometric reference frame using these primitives. While automatic matching of multiple point clouds still is a topic of research [2, 6], different approaches for seg- mentation and representation through geometric primitives exist. Segmentation techniques include clustering based on local surface normal analysis [1, 6], region growing using scan geometry and point neighborhoods [2], generation of a discrete grid [9], or a split-and-merge scheme applying an octree structure [10]. The objects are often represented by planar elements recovered through RANSAC schemes [1], creation of a triangular irregular network [4], tensor vot- ing [9] or least squares adjustment [10]. The tensor vot- ing scheme is able to detect not only planes, but also linear structures like high-voltage lines. In this paper, a preprocessing technique is presented, that transforms the point cloud into a set of object surfaces via a two stage process. The first stage is a RANSAC based gen- eration of locally delimited surface patches from the point cloud. The second stage groups patches belonging to the same surface in object space. 2. Generation of surface patch elements The input is a cloud of 3D measurements gathered by a terrestrial laser scanner and locally delimited planes shall be extracted as surface patches. These may represent a part of larger, planar objects but may as well coincide with small object surfaces. The transformation is split up into two sub- processes, a partitioning of the point cloud into spatial bins and the robust estimation of the dominant plane in each bin. 2.1. Spatial data partitioning The set of 3D points is partitioned and assigned to 3D volume cells using a Cartesian raster. All points in one of the raster cells will be denoted by X . Two objectives are satisfied by this partitioning: • The raw point cloud is organized into small blocks that can be processed seperately. • Since each cell only covers a small part of the complete scene, it emphasizes local features in object space. 0-7695-2521-0/06/$20.00 (c) 2006 IEEE