ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision“, Graz, 2002
LOCALIZATION AND GENERATION OF BUILDING MODELS
Ildiko Suveg, George Vosselman
Photogrammetry and Remote Sensing
Delft University of Technology, The Netherlands
Thijsseweg 11, 2629 JA, Delft
(I.Suveg, G.Vosselman] Ogeo.tudelft.nl
KEY WORDS: aerial images, 2D GIS map, building models, fitting, mutual information
ABSTRACT
This paper presents a knowledge based approach for 3D reconstruction of buildings from aerial images. The aerial images are
combined with information from 2D GIS database and specific knowledge of the buildings.
This paper describes the generation of building hypotheses in case of large variations in the terrain height. First the possible
locations of a building in the images are determined by using the ground plans of the building defined in the map and lines
extracted from images. These possible locations are verified by the 3D building model generation process. The generated
building hypotheses are improved by fitting them to the image data. An evaluation function based on information theory
principles is used to select the best model.
Experiments are presented that demonstrate the approach by reconstructing 3D building models.
1 INTRODUCTION
A lot of approaches have been proposed for the 3D recon-
struction of buildings from aerial images. However, despite
of intensive research, the current state of automation in the
3D reconstruction of buildings from aerial images is quite low.
Most approaches have focused on the reconstruc-
tion of specific building models: rectilinear shapes
[Noronha and Nevatia, 1997], [Roux and McKeown, 1994],
flat roofs [Jaynes et al., 1997], [Lin et al., 1994] or para-
metric models [Fischer et al., 1998]. But buildings show
a much wider variety in their shape. Other approaches
employ a generic roof model that assumes planar roof
surfaces [Bignone et al., 1996], [Moons et al., 1998],
[Schmid and Zisserman, 1997]. These 3D roof planes are
generated by grouping the coplanar 3D lines or corners com-
puted by matching of image features extracted from stereo
images. Hence, these approaches rely on the extraction of
image features, which raises a lot of problems especially for
aerial images. The feature extractors can fragment or miss
boundary lines, due to low contrast, occlusions, and bad
perspective.
A more robust approach should combine different
data sources. The image data can be combined
with scanned [Maitre et al., 1995] or digital maps
[Haala and Anders, 1996]. These approaches represent
the newest trend in 3D building reconstruction.
Our strategy for 3D reconstruction of buildings combines
pairs of stereo images with large-scale Geographic Informa-
tion System (GIS) maps and domain knowledge as additional
information sources. The 2D GIS database contains the out-
line of footprints of the buildings. The knowledge about the
problem domain is represented by a building library containing
primitive building models. Although, buildings reveal a high
variability in shape, even complex buildings can be generated
by combining simple building models with flat, gable or hip
roof.
This paper describes the localization and generation of build-
ing hypotheses using the ground plans of the buildings defined
in the map.
The paper is organized as follows: Section 2 presents a brief
À - 356
overview of the steps involved in our approach for 3D recon-
struction of buildings. The next section describes the ap-
proximate localization of buildings into images while section
4 presents the generation of building hypotheses. Section 5
presents experiments which demonstrate the effectiveness of
the method. The conclusions and future work are discussed
in the final section.
2 METHOD OVERVIEW
The complexity of the reconstruction process can be reduced
by a large amount by focusing on one building structure.
Therefore, it is desirable to localize the buildings in the im-
ages first and afterwards to do the actually reconstruction.
To cope with the complexity of aerial images, specific knowl-
edge about buildings is integrated in the reconstruction pro-
cess. Since most buildings can be described as an aggregation
of simple building types, the knowledge about the problem
domain can be represented in a building library containing
simple building models (flat roof, gable roof, and hip roof
building).
The building reconstruction process is formulated as a multi-
level hypothesis generation and verification scheme and it is
implemented as a search tree. The tree is generated incre-
mentally by the search method.
The first step of the actual reconstruction process is the par-
titioning of a building in simple building parts, which might
correspond to the building models defined in the building li-
brary. First, the partitioning is done using only the ground
plan of the building defined in the GIS map. If the ground
plan of the building is not a rectangle, then it can be divided
in rectangles, called partitions. Then, a partitioning scheme
can be defined as a subdivision of a building into disjoint
partitions. Usually, a building can have multiple partitioning
schemes.
All the possible partitioning schemes of a building are repre-
sented on the first level of the search tree. To avoid a blind
search method of the tree, partitioning schemes are ranked
based on their support in the images. This ranking provides
a means of giving higher priority to the partitioning schemes
with a more support in the images. The second level of the
