> given type
used as ob-
ge analysis.
1tilizing the
ie following
of two ma-
ition of ob-
;ored repre-
sed images.
tions of ob-
r goal is to
ral map to
e following
Xf a three-
perspective
| amount of
ther hand,
ot relevant
knowledge
)y a object
| from irrel-
ompleted.
roups. Spe-
detail rep-
resenting a unique object. An example are objects repre-
sented in a CAD system by fixed geometry and fixed topol-
ogy used for tasks like object recognition. Generic models
on the other hand are used to represent classes or groups
of similar objects. An example for this kind of model
are parameterized models, permitting the representation of
objects by fixed topology but variable geometry (Braun,
Kolbe, Lang, Schickler, Steinhage, Kremers, Fórstner &
Plümmer 1995). These models contain all primitives of a
certain object class without defining the parameter values.
Length, width and height of a quader type object are free,
but the number of its points, lines and faces as well as their
relations like paralellity of lines or coplanarity of points are
fixed. For this reason, a quader is a generic model, capable
to represent a whole group of objects.
In principle models for 3D object recognition can be
adapted quite well from preexisting CAD-like descriptions
of the visible objects (Flynn & Jain 1991). Even though
a 3D building model can not be derived definitely from
the existing two-dimensional ground plan, the existing GIS
data can at least be used to provide a first hypothesis of
the 3D shape of the building. Assuming parameter values
for the (unknown) roof slope a building of a certain (also
unknown) height, which are e.g. dependent on the given
usage of a building, the model lines defining the eaves and
ridges of a roof can be constructed. For this type of model
the parameters referring to the ground plan are fix (e.g.
length and width of the building), while other parameters
referring to the third dimension (e.g. roof slope or height
of the building) are free and therefore will have to be de-
termined by image analysis. For this reason a model gen-
erated using an existing ground plan can be interpreted as
a mixture between a specific and a parameterized model,
since some parameter values are fixed and some parameter
values are free.
3.2 Analysis of GIS data
Main goal while analysing the existing digital cadastral
map is to select or even create a suitable 3D building model
which is required for the object reconstruction by image in-
terpretation. This task implies steps like the extraction of
relations which are only contained implicitly in the avail-
able data, the elimination of unnecessary information (e.g.
details not visible in aerial images) and the generation of
hypotheses on the missing 3. dimension of the represented
building. Therefore the analysis process split up into the
generalization, i.e. the simplification of the given contour
lines and the combination of adjoining ground plans, and
the construction of buildings.
3.2.1 Generalization
In order to eliminate details not visible in the aerial im-
ages and to simplify the verification process the shape of
the ground plans extracted from the digital cadastral map
has to be generalized. If e.g. a building is covered by a
saddle roof at details of the given ground plan e.g repre-
senting bays or ledges will be hidden. Therefore we use
the assumption that the 2D contour of a saddle roof can
be defined by a rectangle which approximates the shape of
the given ground plan.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
à
Umax =
r2
mY, oe mm vm Dm mins sem we mm am i) cs €
> >
x0 rl
Umin
rm
Amin À max
Figure 2: Generalization of a ground plan
The rectangular contour of a saddle roof can e.g. be cre-
ated by a structural analysis (symmetry, similarity, close-
ness, unity, continuation) of the given ground plan. For
that purpose length and parallelism of polygon edges are
the most important features. The longest polygon edge of
a given ground plan can e.g. be used to adjust the longest
side of a roof. Parallel lines in the ground plan can be used
to build a set of rectangles which can be grouped together
(elimination of overlapping areas of the rectangles) to find
the best representation with rectangles. We use a very
simple approach to create a rectangular representation of
a given ground plan:
1. Find a right-angled corner in the given polygon P.
Only if a right-angled corner exists, the polygon
is generalized to a rectangle. The selected corner
represents the base of a two-dimensional coordinate
system as shown in figure 2. Now each point of
the given polygon can be described by the following
equation :
Pi = Zo + Aii urs (1)
2. In order to determine Amin, Amaz, min and [maz
all points p; € P are inserted in equation (1) to cal-
culate the minimum and maximum values A; and ju.
The points
Pini sHmaz | DAmin mar DA rain min XxDAmac LH min )
then represent the required rectangle.
3. The overlap between the constructed rectangle and
the ground plan can be used as measure on the good-
ness of the approximation.
Because the ground plans of the digital cadastral map de-
scribe properties, one physical building can be represented
by two or more adjoining ground plans, if the building is
owned by more than one party. Especially if the grouped
ground plans can be represented by a rectangle, it is very
likely that they are covered by one common roof. There-
fore adjoining ground plans are detected by searching for
line segments in the digital cadastral map which are used
by different polygons.
3.2.2 Construction of buildings
The creation of a 3D building is a task that cannot be
solved in a definite way because besides the unknown
height values there are many possible types of roof shapes,
e.g. desk, flat, saddle or hipped roofs. Because the usage
of a building provides a good hint on the possible type of
287