the laser DSM is applied, whereas structuring is based on the
automatic interpretation of the given ground plans.
Figure 1 DSM from laser scanning with aerial image overlaid
Usually our DSM data is collected by airborne laser scanning
with a mean point density of one point per square meter. This
type of data can already be used for the generation of realistic
visualizations if an aerial image is wrapped over the DSM
surface (Figure 1). Nevertheless, this type of data provides only
iconic representations, there is no object related information
available at this state. This object related information is for
example required in order to enable queries on selected
buildings or to integrate the data with a 3D GIS. Additionally, if
the original DSM is applied for that purpose, the computational
burden caused by the large amount of 3D points to be
represented prevents the generation of animated visualizations.
International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 3W14, La Jolla, CA, 9-11 Nov. 1999
3 AUTOMATIC AND SEMI-AUTOMATIC
BUILDING RECONSTRUCTION
Decomposition
into primitives
D 2D primitives D
Estimation of s
3D parameters TN
D Merging ..
(CAD kernel) /
| Reconstucted building |
1 Texture —
rm mapping
Figure 2: Workflow for the automatic and semi-automatic
reconstruction algorithm
2D ground plan
3D Primitive
description
Digital Surface Model
Primitive surface
description
3D CAD description
(.sat, .dxf)
Terrestrial images
(facade views) VR Model
Figure 2 sketches the workflow of the reconstruction algorithm.
Input data is on the left, output on the right and the flash icon
marks the places where automatically derived data can be
modified or amended. Processing starts by decomposing the
ground plan polygon into 2D primitives (rectangles). Each 2D
primitive is the footprint of a corresponding 3D primitive. The
location, orientation, and the size of the 2D primitive applies as
well for the 3D primitive. What remains to be determined are
the parameters of the roof, namely roof type (currently one of
Internatione
a
Figure 4: Reconstructe
N
Figure 5: Boundary re;
flat, gable and hip), height of the building and roof slope. A
In order to derive the required 3D CAD-model for each least squares estimation computes the best fit of the models to
building, an abstraction and interpretation of the laser data is the given DSM. When several models are suitable, the one with ;
necessary. This can be solved relatively easy based on human the smallest residual is selected.
knowledge and interaction, but is very difficult to realize by
automatic procedures. Hence, the key idea of our algorithm is to
integrate the required human knowledge into the automatic data
processing. This can be achieved by applying existing
groundplans, which have been collected in advance by a human
operator. Since the human knowledge is implicitly contained in
these groundplans, an automatic reconstruction is feasible even
for very complex buildings. Additionally, compared to
approaches for automatic building reconstruction solely based
on laser data also a much smaller density of laser footprints is
sufficient. In the approach described by Vosselman (1999)
seven points per square meter are required, whereas in our
approach one point per square meter is sufficient.
In addition to these algorithmic aspects the integrity of 2D and
3D data sets can be guaranteed if the generation of the 3D urban
model is based on the analysis of an existing 2D GIS. This also
simplifies the common revision of both data sets. Finally, in
case there are no existing ground plans available they can be
generated before the actual reconstruction by manual
digitization of ortho images or maps. This can be performed Figure 3: Decomposition of groundplans to rectangles
with relatively small effort since a standard soft- and hardware-
configuration is sufficient for this task.
Figure 6: Reconstructe
After this step, the i
overlapping 3D solids
a list of solid descripto
desirable to find a b
parts. As this is a stan
to perform the necess.
Finally, a non-overla
Which can be exported
