-B3, 2012
al segmentation
voxelspace. The
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
FINDING CUBOID-BASED BUILDING MODELS IN POINT CLOUDS
William Nguatem, Martin Drauschke, Helmut Maye
Y
Werner-Heisenherg-We
Institute of Applied Computer Sci
£3
=
ence,
34, 85
Bundeswehr University Munich
STT Neubiberg. Germany
william.nguatem,martin.drauschke helmut .mayerfunibw.de
Commission TEV 2
KEY WORDS: Point Cloud Segmentation, Surface Detection, Building Recognition, Building Reconstruction
ABSTRACT:
In this paper, we present an automatic approach for the derivation of 3D building models of level-of-detail 1 (LOD 1) from point clouds
obtained from (dense) image matching or, for comparison only. from LIDAR. Our approach makes use of the predominance of vertical
structures and orthogonal intersections in architectural scenes, After robustly determining the scene's vertical direction based en the
3D points we use it as constraint for à RANSAC-based search for vertical planes in the point cloud. The planes are further analyzed to
segment reliable outlines for rectangular surfaces within these planes, which are connected to construct cuboid-hased building models.
We demonstrate that our approach is robust and effective over a range of real-world input data sets with varying point density. amount
of noise, and outliers.
1 INTRODUCTION
The derivation of building models from 3D point clouds is a very
active research topic in computer graphics. photogrammetry and
geoinformatics (Wahl et aL, 2008; Meixner and Leberl, 2011:
Huang et al, 2011». Research on the derivation of building mod-
els is focused on (u) reduction of the amount o£ data, (h) auto-
matic object recognition and reconstruction from unstructured 3D
point clouds, and (c? integration of semantics into building aad
city models. Most of the work focuses on the analysis of terres-
trial amd airborne LIDAR point clouds, which are highly accurate
and very dense. In contrast to LIDAR point clouds. point clouds
reconstructed from multiple images (Snavely et ai. 2006: Hiep
et aL, 2009; Agarwal et al. 2010; Bartelsen and Mayer, 2010;
Frahm et al., 2000) are less precise and not so dense, but reliable
enough to recognize objects and their parts hy visual inspection.
Since point clouds can be generated very cost effectively from
images. we see a need for an automated analysis of these point
clouds. LOD | models representing buildings as cuboids with
planar and rectangular surfaces can be used for rectifying images
by means of homographies on the facades which can be used for
further facade analysis. eg, the detection of windows (Lee and
Nevatia, 2004; Reznik and Mayer, 20081. or the description of
repetitive patterns (Wenzel et aL, 2007). Furthermore, LOD 1
models will drastically reduce the amount of data. Triangulated
meshes of millions of points as obtained, e.g. by Poisson surface
reconstruction {Kazhdan et al. 2006) can easily reach several Gi-
gabytes of memory leading to slow or no visualisation possibil-
ities. Contrarily. cuboid-based building models consisting of a
few planar, rectangular surfaces only need a few Kilobytes.
The recognition and reconstruction of buildings or their parts is
either done based en domain knowledge, e.g. concerning the
shape of the objects of interest (Huang et al. 2011), on the re-
lation to other objects of the scene (Schmittwilken et al. 20091,
or on the piecewise intersection of planar, cylindrical, toric and
conical surfaces (Schnabel et al. 2007). In our work, we want to
find cuboid-based buildings. i.e., which either have the shape ot a
cuboid, or can be decomposed into several, possibly overlapping
cuboids. This is the first step towards a hierarchical modeling
of detached buildings in rural and suburban areas where build-
145
ings typically stand alone or in small groups. 1n urban downtown
scenes, our approach will lead to building blocks rather than sin-
gle buildings.
Our approach is based on the detection of the principal cuboid
from 3D point clouds obtained during image orientation. As op-
posed to LIDAR data, such 3D points are often not very dense.
We assume that the four main walls of a building correspond to
four faces of a cuboid without top and bottom. By further repre-
senting every face by a plane. we obtain a unique pair of paralle!
orthogonal planes for our principal cuboid.
The cuboid estimation algorithm thus can be reduced to the deter-
mination of pairs of unique orthogonal planes in the given data.
This is conducted by means of RANSAC (Fischler and Bolles,
19811. Restricting the first stage of modeling 3D buildings to this
case of a rectangular groundplan, our approach robustly produces
eliable results from which further refinements can be made, pro-
gressively increasing the level of detail. This is illustrated by ex-
perimental results, using terrestrial data collected at a small vil-
lage in Southern Germany.
The paper is structured as follows, In Section 2 we give an
overview on existing approaches for detecting planes and recon-
structing buildings in 3D point clouds. We describe the concept
of our algorithm in Section 3 and present algorithmic details in
Section 4. After showing and discussing experiments in Sec-
tion 5, we conclude with future work in Section 6.
2 RELATED WORK
In this section, we summarize related work, We discuss general
approaches for the detection of geometric primitives in 3D point
clouds w.r.t. the detection of vertical planes. I£ the vertical direc-
tion is known, it can be used as a constraint for plane detection.
thus. we also comment on research in this direction Because we
are interested in extending eur work towards the detection and
modeling of building parts. we also introduce related work on
this topic.
There exist several approaches for building recognition and re-
construction from point clouds which depend on precise and very
