Building Roof Reconstruction Based on Residue
Anomaly Analysis and Shape Descriptors from
Lidar and Optical Data
A. Salehi and A. Mohammadzadeh
Abstract
Recently, combined use of lidar and optical data has been
extensively employed to achieve 3D modeling of buildings.
In the first step of this research, lidar point cloud is clustered
into roof’s major and minor planes. A FAST algorithm is used
to extract roof corners from a 2D optical data. Afterwards,
residue of the anomalies is obtained through subtraction of
main planes’ DSM from DSM of the lidar point cloud. Using
a Canny operator, edges of the residue’s anomalies image
are extracted and then superimposed on the optical image to
find appropriate corners. In the second step, shape descrip-
tors related to all possible polygons formed by the corners
are computed and compared against the ones, which are
obtained from the extracted edges. Then, the polygon with
the most similar shape descriptor would be chosen as the
optimal polygon representing the roof structure of interest. In
third step, planes are intersected and initial 3D roof model is
reconstructed. It is added to 3D optimized roof model formed
by 2D optimal polygon and minor planes which results in
secondary 3D reconstructed roof model. Finally, a wire frame
model is generated by vertexes. The results indicate efficiency
of the residue of the anomalies and shape descriptors for
buildings’ roof modeling.
Introduction
Nowadays, 3D geospatial data production has become one of
the major interests for many countries in E-Government (Xu
et al.
, 2012). It has variety of applications in map updating,
change detection, energy management, real estate, simulating
the weather pollution, tourism, and urban planning (Chen
et
al.,
2014). Photogrammetry and the remote sensing society
can play a major rule for providing such a 3D data. Especially
in the past two decades, buildings as one the most important
objects in the mentioned 3D geospatial data, has captured
the interest of many researchers. In 3D building modeling
research field, the accurate roof modeling of buildings is one
the most important challenges which is not completely solved
(Zhang
et al.,
2014) .The accurate roof modeling is essential
prerequisite for the pollution modeling of a city, analysis of
noise propagation and fluid expansion, and also determining
an optimized position for solar plates (Jochem
et al.
, 2009;
Vosselman and Dijkman, 2001; Suveg and Vosselman, 2002).
Moreover, current new advanced research fields like Micro-
Climate (city pollution modeling, heat islands) and renewable
energy need 3D city roof modeling in the LOD3 model (Asawa
et al.,
2012, Quan
et al.,
2015, Vuckovic
et al.,
2016). These
are new demands in remote sensing and photogrammetry
field, which conduct the major motivations of this research
work (Quan
et al
., 2015, Vuckovic
et al
., 2016). The lidar data
has great advantages such as direct acquisition of 3D data
from building roof and fast dense point cloud acquisition
with acceptable accuracy. The lack of accurate corner posi-
tions in lidar data leads for simultaneous use of optical and
lidar data for accurate detection of corners in roof reconstruc-
tion (Baltsavias, 1999; Satari
et al.
, 2012).
The previous building reconstruction methods from lidar
and optical data fall in two categories of (a) data-driven, and
(b) model-driven. In the data- driven methods, some elements
such as line and plane are used in building roof reconstruc-
tion, which gives more flexibility than model-driven ones.
Model-driven methods are limited to a library of predeter-
mined exclusive models (Satari
et al
., 2012) which should be
fitted to the building point cloud using optimal parameters.
Recently, the detail level of the building reconstruction is
one of the most important subjects in this field (Gröger and
Plümer, 2012). There are five level of details (LODs) for build-
ing reconstruction (Gröger and Plümer 2012) including: (a)
LOD0 which the buildings are detected and extracted with
2.5D horizontal polygon with determined roof height, (b)
LOD1 that is shown as a simple volume cube,( c) LOD2 which
the total roof surfaces such as gable roof surfaces are added to
LOD1, (d) LOD3 in which roof details are added to previous
level, and (e) LOD4 that interior structures of buildings are
considered. In the following, previous data-driven and model-
driven methods are discussed according to their LODs.
LOD0 which is the simplest representation of a building
was comprehensively studied by many researchers in the
past decades. Some of the most recent ones would be Ma,
(2005); Awrangjeb
et al.
(2012); Ahmadi
et al.
(2010); Aw-
rangjeb
et al.
(2013); Fazan and Dal Poz (2013); Yang
et al.
(2013); Awrangjeb and Fraser (2014); Mongus
et al.
(2014);
Rottensteiner
et al.
(2014); and Turker and Koc-San (2015). By
major advancements in building reconstruction algorithms,
considerable number of research activities is carried out in
LOD1 (Wang and Schenk, 2000; Suveg and Vosselman, 2002;
Lafarge
et al
., 2006; Dornaika and Brédif, 2008; Kwak
et al
.,
2012; Sirmacek and Lindenbergh, 2015). Recently, the neces-
sity of having more details about buildings’ roof in the virtual
city models has grabbed the attention of the researchers to
LOD2 (Gröger and Plümer, 2012). Sohn
et al.
(2008) detected
and clustered buildings from a lidar point cloud using height
and planar feature. They extracted lines and separated planar
surface and merged them to generate buildings. In
addition,
Zhang
et al.
(2009) separated ground and non-ground point
cloud and segmented the non-ground points then determined
topology of boundaries and roof direction. By intersecting the
planes, the building roof model was reconstructed. Moreover,
A. Salehi and A. Mohammadzadeh are with the
Photogrammetry & Remote Sensing Department, K.N.Toosi
University of Technology, Iran (
.
Photogrammetric Engineering & Remote Sensing
Vol. 83, No. 4, April 2017, pp. 281–291.
0099-1112/17/281–291
© 2017 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.83.4.281
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
April 2017
281
