Automatic Reconstruction of Parametric Building Models from Indoor Point Clouds Sebastian Ochmann Richard Vock Raoul Wessel Reinhard Klein University of Bonn, Germany CAD/Graphics 2015
Motivation Digital 3D building models increasingly used for e.g.: Construction planning, renovation, retrofitting (Automatic) taking of measurements Acoustic or thermal simulations Long-term facility management Image from http://www.digital210king.org/ 2
Motivation Requirements on the models, applications: Analysis and editing on a high level, e.g. moving walls Parameterization of elements, e.g. size of a wall or room Relations between elements, e.g. wall/room adjacency 3
Motivation Suitable models often not available Especially for legacy building stock Use 3D point clouds as starting point for modeling Modeling still largely manual & time-consuming Automated reconstruction methods highly desirable 4
Related Work Active field of research, many recent reconstruction approaches Focus on reconstruction of separate surfaces or room volumes Lack of decomposition into building elements ( What is a wall? ) Our approach 5
Related Work Some methods employ a spatial partitioning and region labeling for finding a plausible room layout, e.g.: Oesau et al. 2014: Indoor scene reconstruction using feature sensitive primitive extraction and graph-cut Turner et al. 2014: Fast, automated, scalable generation of textured 3D models of indoor environments Mura et al. 2014*: Automatic room detection and reconstruction in cluttered indoor environments with complex room layouts * Image from Mura et al. 2014 6
Related Work Shortcomings of previous approaches: Only separated rooms or paper thin walls possible No notion of volumetric walls representing the building s structure, especially shared walls between adjacent rooms Projected planes Partitioning Possible labelings 7
Our Approach Starting point: Registered 3D indoor point cloud scans Labels: Ideally one per room (plus an outside label) Rooms unknown instead: One label for each scan 8
Our Approach Refinement of initial segmentation given by separate scans Diffusion process based on mutual visibility between point pairs Also automatically filters out clutter outside of the building 9
Our Approach Detection of vertical planes (RANSAC) Transfer into horizontal plane Generation of piecewise-linear partitioning? 10
Our Approach Generate candidates for walls (incl. estimated thickness) from detected vertical planes Edges of partitioning are centerlines of (potential) walls Projected planes Potential wall structures 11
Our Approach Generate edges for each single surface and pairs of parallel surfaces Edges are wall centerlines instead of wall surfaces Each edge attributed with (scalar-valued) wall thickness 12
Our Approach After suitable labeling: Connected graph of volumetric wall elements representing the building s structure Potential wall structures Region labeling Resulting wall graph 13
Our Approach We now have point labels and the partitioning: How to determine a suitable labeling of the partitioning cells? 14
Our Approach Labeling optimization as cost minimization problem: Σ (Unary costs for assigning each room label to a region) + Σ (Binary costs for assigning a pair of labels to adjacent regions) min. 15
Our Approach Minimizing the sum of labeling costs yields the desired labeling Graph multi-labeling problem solved using algorithm by Boykov et al. Edges between differently labeled regions are walls Room heights estimated from horizontal planes in the data 16
Opening detection & classification Further enrichment the resulting model Detect openings, i.e. doors and windows Classification by means of supervised learning 17
Multiple scans within a room In case of multiple scans within a room, implausible walls are removed in a post-processing step 18
Results Experiments on a variety of real-world point cloud datasets 19
Results Tested with up to 67 scans, 22.7m points, about 8.5 minutes 20
Results Regularization and hole filling in cluttered regions Controllable via smoothness property of graph cut based opt. 21
Results Comparison with a professional, manually generated model Good wall and opening localization 22
Conclusion The first method for reconstructing parameterized building models based on volumetric, interrelated building elements Direct export to industry-standard format (IFC) Enables automatic or semi-automatic further processing (editing, measurements, simulations, ) in industry-standard software 23
Thank you 24