Chapter 1: Introduction 1
1.4 Three dimensional building reconstruction from remotely sensed data 24
As introduced in the previous section, a building does not have a constant height. Generally, building roofs can have many parts such as ridges, hips, and eaves, as shown in Figure 1-12. One building height measurement extrapolated to include an entire roof is not enough to support detailed 3D reconstruction. To accurately describe a 3D building model, the heights of each roof part need to be estimated individually; consequently, advanced sensors and very high resolution data are required for detailed 3D modelling. For example, LiDAR data with high point density and aerial photos with centimeter spatial resolution are popular for building reconstruction applications.
Typical roof types have been investigated in the current literature, such as the flat, gable, hip, and shed roofs identified in Figure 1-13. A successful reconstruction of typical roofs provides a solid basis for complicated roof reconstruction, as complicated roofs are based on typical roofs and often contain individual components derived from typical roof types. Pre-defined roof types can simplify the reconstruction task and provide a prior knowledge for the automatic reconstruction. A roof type assumption can largely improve the reconstruction accuracy, and is thus widely used in studies (Henn et al., 2013; Huang et al., 2013).
Figure 1-12. An example of complicated rooftop with definitions of roof elements (Photo courtesy of Wikipedia: http://en.wikipedia.org/wiki/File:Roof_diagram.jpg)
Figure 1-13. Examples of some popular roof types. Roof
types
Conceptual roof shapes Example roofs
Flat
Gable
Hip
Automatic reconstruction of digital 3D building models based on 2D building footprints remains a challenging task. Some 3D building methods simply assign a constant height to a 2D building footprint and extrude the building up (Lafarge et al., 2008); essentially, the resultant extruded building is not a 3D, but a 2.5D model. Most methods are developed on LiDAR data for the building roof reconstruction (Haala & Kada, 2010; Khattak et al., 2013; Kong et al., 2013). A systematic review of current 3D building reconstruction is studied in Wang (2013), where reconstruction methods are roughly categorized based on their data as image-based, LiDAR-based, or image-LiDAR fusion methods. This classification framework is introduced in Figure 1-14 and current popular software used for 3D building reconstruction is provided in Table 1-1.
Furthermore, algorithms required to automatically reconstruct 3D buildings can be divided into model-driven and data-driven methods. In model-driven methods, typical roof types are predefined with input point clouds or DSMs fitted to the predefined roof types for modelling (Henn et al., 2013; Huang et al., 2013). In data-driven methods, point clouds and digital surface models (DSMs) are segmented and grouped, features are recognized, and 3D models are built accordingly (Lafarge & Mallet, 2012; Zhang et al., 2012). As in the example provided in Figure 1-15, aerial image and LiDAR data are used to cluster heights, compose planes, detect edges, reconstruct facets, and conduct post- processing.
Figure 1-14. Classification of 3D building modelling methods. Adapted from Wang (2013).
Table 1-1. Existing commercial 3D building reconstruction systems and research prototypes. Adapted from Wang (2013)
System Developer/researcher Input data Description
CC-Modeler CyberCity AG & ETH, Zurich Calibrated stereo pan of aerial images Semi-automated photogrammetric 3D reconstruction system inject Inpho GmbH & Bonn
University, Germany Calibrated single, stereo, or multiple overlapping aerial images Semi-automated
Constructive Solid Geometry based approach
Ascender University of Massachusetts
Calibrated multiple aerial (nadir and oblique) images
Automated 3D building model reconstruction SiteCity Digital Mapping
Laboratory, CMU,
Calibrated multiple aerial (nadir and oblique) images
Semi-automated photogrammetric 3D reconstruction system ImageModeler RealViz & INRIA,
France
At least two photos taken from
different positions
Accurate 3D measurement and modelling from photos PhotoBuilder Oxford University, UK Uncalibrated two or
more photos
Vanishing points based method to 3D reconstruction Nverse Photo Precision Lightworks,
USA
Two or more aerial images
A series of plug-in components Shape Capture ShapeQuest Inc. &
NRC, Canada,
Single or more photos
Accurate 3D measurement and modelling from single or more photos
PhotoModeler Eos Systems, Canada Single or more photos
Accurate 3D measurement and modelling from single or more photos
PhotoGenesis Plenoptics Ltd, UK Uncalibrated single or more photos
Semi-automated model- based 3D reconstruction system
Photosynth Microsoft Internet photos Sparse 3D model generation for navigating images in 3D space
Pix4UAV Pix4D, Switzerland Aerial images Automatic 3D model generation from aerial images
Table 1-1 (Continue). Existing commercial 3D building reconstruction systems and research prototypes.
System Developer/researcher Input data Description
C3 Apple, USA Aerial images Automatic 3D model generation from aerial images
Edgewise ClearEdge3D, USA Range data 3D modelling using range data
One reoccurring issue in current 3D building reconstruction is that often, rooftops alone are reconstructed while walls are assumed to be featureless (Haala & Kada, 2010). The evaluation process for reconstructed 3D buildings has not been thoroughly discussed, with most studies interested in developing sophisticated reconstruction methods. Currently, evaluation methods for 3D building reconstructions are derived from 2D building footprint evaluation methods. For example, 3D buildings are evaluated based on building roof completeness, topological consistency between reference and sample 3D buildings, and geometric accuracy in XY and Z directions (Rottensteiner et al., 2013). Often, building wall correctness and shape similarity are ignored.
(a) (b)
(c) (d)
(e) (f)
(g) (h)
(i)
Figure 1-15. The procedure of 3D building reconstruction based on an aerial image and LiDAR data. Adapted from Sohn et al. (2012): (a) aerial image, (b) LiDAR data, (c)
extraction, (g) building reconstruction with distortion errors indicated as arrows, (h) shape regularization, and (i) 3D polyhedral building model.