togrammetry. Wuhan
ZP. Lu, T.D.. 2008.
unit quaternion. Acta
), pp. 30-35.
K. & Nlle, M., 1994.
ree point perspective
Journal of Computer
warm optimization. In:
onference on Neural
-1948.
zoulas, G.D., Vrahatis,
Analysis and Global
hers, Dordrecht, The
| schools: a distributed
ics (Proceedings of
dified particle swarm
ational Conference of
laska, pp. 69-73.
tudy of particle swarm
EE Conference of
DC, USA, pp. 1945-
n, J.H., 2006. Method
ents of exterior
once and Technology,
'riment and analysis of
jn. Journal of Xihua
), pp. 76-80.
04. Particle Swarm
9-13
02.Camera calibration
. with planar scenes.
chan University, 27(6),
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
A REFINING METHOD FOR BUILDING OBJECT AGGREGATION AND FOOTPRINT
MODELLING USING MULTI-SOURCE DATA
Y. Li**, L. Zhu, H. Shimamura?, K. Tachibana®
? International Institute for Earth System Science, Nanjing University, Jiangsu, China, 210093,liyan@nju.edu.cn
? Pasco corporation, 1-1-2 Higashiyama, Meguro-ku, Tokyo 153-0043
luihnz7801@pasco.co.jp
Commission III, WG II[/1
KEY WORDS: Modelling, polygon, image, DSM, feature, fitting
ABSTRACT:
Automatically detection, extraction and re-construction of 3D building modelling are difficult yet potentially high-payoff challenges
for photogrammetric applications. Solution usually requires integrating various sources, including LIDAR, imagery, and digital
surface models (DSM). However, highly automated and robust geometric modelling remains unsolved. We will present a 2D
modelling technique which represents a building's outline in an as-is way. It gives visually accurate corners and lines for buildings.
Aerial remotely sensed imagery and a DSM are used to detect and segment building masks. A refining footprint modelling is
implemented through line modelling, edge refining, and segment merging and generating. A district grouping based main orientation
algorithm is proposed. This approach has the ability of successive improvement, moving from a prototype to a subtle end product.
Experiments with Japanese data show that the models generated automatically fit the manual models very well.
1. INTRODUCTION
1.1 Motivation
There is an increasing demand of building models in various
GIS applications. Remote sensed data provides a cheaper and
more effective source for this demand. Yet highly automated
and robust building modelling remains a problem unsolved.
Large-scale production of building models greatly rely on 2D
digital line graph data (DLG) that are generated interactively,
while the 3D models are automatically derived using reliable
tools provided by softwares like TerraScan (TerraSolid, 2011)
CyberCity Modeler (Gruen, 2003) PhotoModeller (Zlatanova
2011) on the base of the 2D DLG. Generating the DLG or the
2D model usually take most of the workload in any building
mapping projects. Especially for the large amount of remote
sensed data, manual processing of the 2D modelling by fewer
workers is unpractical. So the automatically modelling with
fewer parameters or interactions is mostly needed.
1.2 Related researches
1.2.1 Building extraction
Building extraction technology uses computer science including
image processing, pattern recognition on single or multiple
images to detect and extract the information of a building, such
as contour, shape, location or height. Segmentation, feature
detection can be used singly or integrated. In early years,
researchers generally use image singularly to make it. For
instance, in 1990s, Irvin etc. proposed an idea extracting
buildings by their shadows (Irvin 1989). Off-terrain objects
have shadows if no other height data shadow can directly but
partly indicate the height information. When there are more data
sources available, multi-source data based methods become
popular. In conclusion, the building extraction technologies can
41
be divided approximately into two big categories, no matter for
imagery only or for multiple data.
Segmentation based
Region growing (Lari 2007). texture segmentation (Kim 1999,
Levitt 1997), dynamic contour (Fazan 2010) including level set
(Hao 2010) belong to this type. Some methods only partition
the image into regions. Further recognition or classification
(Stassopoulou 2000, Baatz 1999) is needed to tell which ones
are buildings. For texture analysis it is difficult to get a clear
edge between man-made and non man-made terrain. So it is not
proper to use this method for building modelling in high
resolution imagery. For large scale imagery, the intensity or
color of building roofs can vary extremely and can ruined the
segmentation.
Feature detection based
For better interpretation of a building, feature based detections
are mostly studied. Among all the features, line is the most
important one to identify the buildings. Grouping process (Lin
1998, Lee 2003) is proposed using the edges and line segments
derived from the image to group according to their spatial
relations. Because no restraint is given to a building area, this
kind of methods will be affected sensitively by the thresholds.
Moreover, in resolutions higher than 0.5m there will surely are
more lines of ridges which are not parallel each other.
1.2.2 Modelling methods
There is numerous building modelling theorems. In 2D case,
some studies assume that the house is composed by several
parts, and each one is modeled separately as a few of given
primaries, called parameterized methods (Tseng 2003, Braun
1995, Chow 2003). Others take the roof as several planar which
can be extracted from the data without certain shape of models
(Laramee 2003. For our case, we would like to consider the
contour polygon firstly. So the planar detection based methods
won't be employed.
