of each selected segment. These segments of the refined seg-
mentation form the basis for the extraction of the buildings'
2D information, whereas the height information is derived by
analysing the height information within the segments and the
related bounding box without segments. Figure 1 shows an
overview of the building detection using the ISPRS test data
set FLAT as example.
The use of a geometric criterion to distinguish between build-
ings and other objects higher than the topographic surface is
not always sufficient. Therefore, other criteria using other
sources of information have to be used, e.g. texture in-
formation [Eckstein and Munkelt, 1995] or edge information
[Baltsavias et al., 1995] from aerial imagery. This informa-
tion can also and — depending on the application — should
already be applied during the DSM generation.
4 BUILDING RECONSTRUCTION
The building reconstruction of our approach is based on the
use of parametric and prismatic building models. Parametric
models are used for simple separated buildings, which can be
described by a few parameters, e.g. a building with a symmet-
rically sloped roof, whereas prismatic models (ground plan
and height information) including generic knowledge about
regularities (e.g. orthogonalities, parallelisms, collinearities)
are used for complex buildings or building blocks.
4.1 Parametric Building Models
In our approach we use two different parametric models: flat
buildings and buildings with a symmetric, sloped roof. The
form parameters of these building models are the length,
width and height for flat buildings, and length, width, height
of eave-base and height of ridge-eave for buildings with a sym-
metric, sloped roof, assuming that the ground plans of these
buildings are given by rectangles. Furthermore, four param-
eters are needed to describe the position and orientation of
the building within the reference coordinate system.
In order to determine the x,y coordinates of a building's ref-
erence point and the orientation, the point of gravity and the
orientation of each refined segment using the heights within
the segments as weights are computed. The z coordinate of
the reference point is computed taking the mean of heights
within the background area of the bounding box. The param-
eters length and width are the length and width of a rectangle
approximating the segment and are computed along the first
and second main axis of the segment. The height parame-
ters are computed taking the height information in the orig-
inal DSM into account. For this purpose, we use a ranking
scheme, and use the median of the p% minimal and maximal
values within the segments as robust estimation of the mini-
mum and maximum, where p = 10 is choosen. The height of
a flat building follows from the difference between the mean
height of the segment and the mean height within the back-
ground. For buildings with symmetric sloped roofs the height
parameters are computed as: heightl = difference between
the estimated minimum height of the segment and the mean
height of the background; height2 = difference between the
estimated maximum and minimum within the segment.
The parameters of the parametric models are computed for
each detected segment. In [Weidner and Forstner, 1995], we
selected the model which is used for the description using
a geometric criterion, namely the slope of the roof. This
approach seems to be feasible for the selection of the para-
926
DSM
Segmentation
Extraction of closed contours
Elimination of discretization noise
Local application of MDL
Se
Polygon after
local processing
Derivation of hypotheses about regularities
Determination of a set of linear
independent hypotheses
Global robust estimation
- E I DRE
p
F "^.206.mg' -e— +
Reconstructed
Building
Figure 3: Building Reconstruction: Prismatic Models
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Th
els
wit
