FUSING AIRBORNE LASER SCANNER DATA AND AERIAL IMAGERY FOR THE
AUTOMATIC EXTRACTION OF BUILDINGS IN DENSELY BUILT-UP AREAS
F. Rottensteiner * *, J. Trinder?, S. Clode^, K. Kubik®
* School of Surveying and Spatial Information Systems, The University of New South Wales,
Sydney, NSW 2052, Australia - {f.rottensteiner, j.trinder} @unsw.edu.au
° The Intelligent Real-Time Imaging and Sensing Group, School of ITEE, University of Queensland,
Brisbane, QLD 4072, Australia - {sclode, kubik} @itee.ug.edu.au
Commission III, WG 111/6
KEY WORDS: Building, Extraction, LIDAR, Multisensor, Fusion
ABSTRACT:
Using airborne laser scanner data, buildings can be detected automatically, and their roof shapes can be reconstructed. The success
rate of building detection and the level of detail of the resulting building models depend on the resolution of the laser scanner data,
which is still lower than the resolution of aerial imagery. Building extraction from aerial images alone is difficult because of
problems related to shadows, occlusions, and poor contrast. That is why it makes sense to combine these data sources to improve the
results of building extraction. This article deals with the fusion of airborne laser scanner data and multi-spectral aerial images for
building extraction. There are three instances in the overall process when exploiting the complementary properties of these data
appears to be most beneficial: building detection, roof plane detection, and the determination of roof boundaries. Building detection
is based on the Dempster-Shafer theory for data fusion. In roof plane detection, the results of a segmentation of the laser scanner
data are improved using the digital images. The geometric quality*of the roof plane boundaries can be improved at step edges by
matching the object edges of the polyhedral models with image edges. Examples are given for a test site in Fairfield (NSW).
1. INTRODUCTION
1.1 Motivation and Goals
The high potential of LIDAR (Z/ght Detection 4nd Ranging)
data for automatic building extraction has been shown in the
past, e.g. (Vosselman and Dijkman, 2001). The success rate of
building detection and the level of detail of the resulting
building models depend on the resolution of the laser scanner
data, which is typically still lower than the resolution of aerial
imagery. On the other hand, building extraction from aerial
images alone is difficult because of shadows and occlusions,
and also because the transition from 2D image space to 3D
object space has to be carried out. That is why it makes eminent
sense to combine these data sources to improve the results of
building extraction. There are three instances when exploiting
the complementary properties of these data appears to be most
beneficial (Rottensteiner and Briese, 2003):
(1) Building detection: The main problem in this context is to
distinguish buildings from trees. LIDAR data give parameters
of surface roughness, but with decreasing resolution of the
LIDAR data, the classification becomes more critical in areas
where the appearance of trees and buildings is similar. The
height differences between the first and the last echoes of the
laser pulse and multi-spectral images can be used as additional
data sources to improve the classification results.
(2) Roof plane detection: In order to reconstruct the buildings
by polyhedral models, roof planes have to be detected first.
Large roof planes can be detected in the LIDAR data. The
results thus achieved can be improved by taking into account
aerial images.
* Corresponding author.
S12
(3) Determination of roof boundaries: The geometric quality
of the roof boundaries at step edges, which in general is poor if
only LIDAR data are used, can be improved by image edges.
This paper deals with the fusion of first and last pulse LIDAR
data and multi-spectral aerial images for building extraction. It
consists of two main parts. The first part, describing our method
for building detection and the results that could be achieved by
it, is presented in section 2. In the second part, which is
presented in section 3, we will describe the current state of our
work for the fusion of LIDAR and image data for roof plane
detection and the determination of roof boundaries. Section 4
gives conclusions and an outlook on future work.
1.2 Related work
1.2.1 Building detection: The building detection starts with the
generation of a coarse digital terrain model (DTM) from the
digital surface model (DSM) provided by the LIDAR data, e.g.
by morphologic filtering. A further classification must separate
points on buildings from points on trees and other objects by
evaluating the local surface roughness and other cues. With
multi-spectral images, the normalised difference vegetation
index (NDVI) is well-suited for classification in this context
(Lu and Trinder, 2003).
Various classification techniques have been applied for building
detection, e.g., unsupervised classification (Haala and Brenner,
1999), rule-based classification (Rottensteiner et al, 2003),
Bayesian networks (Brunn and Weidner, 1997), and fuzzy logic
(Voegtle and Steinle, 2003). The probabilistic approaches
among the cited ones face the difficulty of modelling the
