tions, which are only contained implicitly in the available
data, by the elimination of unnecessary information (e.g.
details not visible in aerial images) and by the generation
of hypotheses on the third dimension of the represented
building. The pre-processing of the 2D GIS data will be
discussed in chapter of the article. The hypotheses, which
are generated in the preceding step then have to be ver-
ified and the unknown parameter values of the building
have to be determined in the second step by the analysis
of an aerial image (chapter 3.2.2).
2 AVAILABLE 2D DATA BASE
For most of the highly developed countries like the Fed.
Rep. of Germany, a vast amount of spatial and the-
matic information about man-made objects is already pro-
vided either analogous by conventional topographic maps
and cadastral plans or even digital by a usually two-
dimensional GIS. Basically in Germany exist two kinds
of data sources which contains information about a sin-
gle building. These are the urban development plan and
the digital cadastral map. The urban development plan
provides regulations about roof shape, roof slope, height
and direction of the ridge of the roof, number of floors
and use of the building that are obligatory for clients. A
rough estimate of the location of the building is addition-
ally available by the information about the region where
the building has to build up and the maximum area that
can be covered by the building. Therefore it contains a
lot of relevant information about a building to create or
select a suitable 3D object model, which can be used by
the following automatic image interpretation process. An
important aspect of the urban development plan is that
the information about a building does not depend on the
existence of the building. The main disadvantage of the
urban development plan is that it is not available in digital
form and therefore can not be directly used.
In contrast to the urban development plan the digital
cadastral map contains digital information about existing
buildings. The digital cadastral map is provided by the
land survey offices in Germany, it is build up as an area
covering data base and at the moment is available for 4096
of Germany. An example for this kind of data is given in
figure 1. The digital cadastral map provides information
about the distribution of property, including the borders of
all agriculture areas and the ground plans of existing build-
ings. Additionally it includes information about names of
streets and the usage of buildings (e.g. garage, residential
building, office block, industrial building, church, tower) in
form of text symbols. Because the digital cadastral map is
the only at least partially area covering digital data source,
which is available at present and because a similar type of
information can also be obtained by digitizing maps or
plans, it was exemplary selected as a 2D data source to
support the 3D building reconstruction. Procedures for
the automatic digitization of maps and plans resulting in
information similar to the digital cadastral map are e.g.
given in (Illert 1990) or (Carosio 1995).
For a task like the 3D reconstruction of buildings the
ground plans, which represent the shape of the floors, the
exact location of the buildings and their usage provide
valuable information. Even though the geometric informa-
tion is only two-dimensional, 3D descriptions of buildings
can by constructed by using certain assumptions on the
286
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
of building. These 3D hypotheses then can be used as ob-
ject models for the verification process by image analysis.
The generation of the 3D building hypotheses utilizing the
given digital cadastral map is described in the following
section.
uh nn
we NU € =
A
"eth
Y ee 7 as RAN
= gw s
—M At
A
7
Ë 777 7
Wh V E
gerere Fh dg ass A e
Ae V Qs em 7 Dr gr
2
Figure 1: Digital cadastral map
3 GENERATION OF 3D
HYPOTHESES
A system for 3D object reconstruction consists of two ma-
jor components, the object modelling or generation of ob-
ject representations and the matching of these stored repre-
sentations to descriptions derived from the sensed images.
The model database should contain representations of ob-
jects the system is expected to recognize. Our goal is to
use the information given in the digital cadastral map to
define a 3D model, which can be used for the following
image analysis.
3.1 Object Models
Images are two-dimensional representations of a three-
dimensional world. Due to occlusion and the perspective
view or also caused by lack of contrast, a certain amount of
object information is always missing. On the other hand,
images contain a lot of information, which is not relevant
for a specific task. Only by the use of a priori knowledge
on the sensed objects, which is represented by a object
model, important information can be separated from irrel-
evant details and missing information can be completed.
Object models can be classified in two main groups. Spe-
cific models are descriptions completely and in detail rep-
roof shape or building height, depending on the given type
resenti
sented
ogy us
on the
of sim
are pa?
object:
Kolbe,
Plümn
certair
Lengtl
but th
relatio
fixed.
to rep:
In pri
adapte
of the
a3D
the ex:
data c
the 3I
for th
unkno
usage
ridges
the pe
length
referri
of the
termii
erated
a mix
since :
values
3.2
Main
map i:
which
terpre
relati
able c
detail
hypot
buildi
gener
lines
the c
3.2.1
In or
ages
the g
has t
saddl
senti
the a
be de
the g