FUSION OF 2D GIS DATA AND AERIAL IMAGES FOR 3D BUILDING RECONSTRUCTION
Marko Pagko, Michael Gruber
Institute for Computer Graphics
Technical University Graz
Münzgrabenstr. 11
A-8010 Graz
Austria
email: {pasko,gruber}@icg.tu-graz.ac.at
www: http: //www.icg.tu-graz.ac.at/” {pasko,gruber}
Commision Ill, Working Group 11/2
KEY WORDS: Fusion, 3D GIS, 3D Urban Database, Aerial Photogrammetry, Affine Matching, Roof Skeleton.
ABSTRACT
We present a method to reconstruct roofs of a cities’ buildings. The algorithm is based on a fusion process and exploits aerial
image data and the existing 2D GIS and DTM data. Special attention we offer to the 2D roof skeleton, which is developed
from the fusion result. The footprint of the building and the roof outline, which is part of the roof skeleton, are merged and
create the building box.
KURZFASSUNG
Der Beitrag befaßt sich mit der Rekonstruktion von Dachformen im Stadtbereich. Luftbilder und Daten des bestehenden 2D
GIS sowie Geländehöhen (DTM) werden gemeinsam genutzt. Spezielle Aufmerksamkeit widmen wir dem 2D Dachskelett,
das aus dem Fusionsprozeß der Quellendaten entsteht. Gebäudegrundriß und Traufenlinie (ein Teil des Dachskeletts) bilden
gemeinsam die Gebäudegrobform (” building box”).
1 INTRODUCTION
Automated recognition and reconstruction of objects from
images is one of the most important tasks in the ongoing
development of computer vision and object reconstruction
disciplines. Not at least it is an issue if the new skills and
tools of these working areas are used to enhance and improve
the procedures of automated mapping and documentation
of the environment [Forstner, Pallaske, 1993]. In the case
of urban data acquisition and administration we are now
going towards fully 3D models of towns and need to improve
our methods to increase the output and the quality of a
modern and powerful mapping and modelling procedure
[Leberl et al. 1994], [Gruber et al. 1995a].
Beside others we have focused on the reconstruction of
buildings from aerial images and the exploitation of the
existing knowledge from other data sources like the 2D GIS
[Gülch, 1992], [Lang, Schickler 1993], [Haala et al. 1994].
This leads to a fusion process, which creates the basic rela-
tionship between aerial images and the footprint of buildings,
as it is available from the digital map [Gruber et al. 1995b].
We present the key functions of our procedure and show
the tools and algorithms we use for the detection and
reconstruction of roofs of buildings. The height of the
detected roof-outline and the known terrain elevation leads
to a simple description of the entire building, the so-called
building box. This initial 3D data set of the entire object
shall be used for further processes like the enhancement of
geometrical detail. Finally we have to add the phototexture
from images to the geometry of the CAD model. This task,
which has to be done to involve the sterile geometry towards
a bright and well-rounded database for various usage will be
supported by image data, which have already been involved
in our procedure.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
2 MODELLING CITYSCAPES
The establishment of fully 3D models of urban environment
seems to be the next and unavoidable step in improving
the data for the management of our growing cities. The
3D data will not only present the geometric relationship of
objects, but also permit to create photo-realistic renderings
for various decision making and training tasks. Today we are
able to exploit powerful computers and strong software to
create such databases by hand. We have systems to handle
the source data (e.g. photogrammetric workstations like
DPWSs to set-up image data and to do stereo restitution),
to merge and refine data (e.g. 3D CAD systems) and we
have software to brush up and enhance phototexture (like
Photoshop from ADOBE).
There is no doubt, that ongoing work has to concentrate
onto methods and strategies to increase the throughput
of the entire modelling process. We are sure that there is
not only one field of interest where improvement urgently
needs to be done. Table 1 shows a number of different
activities, which are key members of the modelling process
for Cybercities. The amount of labour based on experiences
during manual modelling and the promise of automation
(5=heigh, 1—low) is estimated. In addition we show which
type of equipment is used during each activity. From Table
1 we see that texture processing and texture mapping needs
about half the cost of labour within the total procedure.
This may feed arguments against textured models, which
we easily contradict by comparing costs and benefits of the
phototexture. We have also take into consideration that one
single part of the entire process is not stand-alone but will
gain from adjacent development (e.g. the automation of the
texture mapping will be influenced by object reconstruction
procedure, if digital image data are source data of both, the
geometry and the texture).
In the current contribution we pick out a specific task of
257