CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
230
which captures the repetitive patterns and particularities of the
building. Finally, new buildings can be generated in the
architectural style defined by the derived grammar. Even though
this approach provides individually representative grammars
instead of predefined ones, the crucial part of the inference
process, the facade interpretation, has to be done manually. In
contrast we pursue an approach which runs fully automatically
during all processing steps.
The automatic generation of a facade grammar, which is derived
from 3D point cloud measurements of a mobile mapping
system, are discussed in section 2. As demonstrated in section 3
top-down predictions can be activated and used for the
improvement and completion of the reconstruction result that
has already been derived from the observed measurements
during the bottom-up modelling. Moreover, the facade grammar
can be applied to synthesize facades for which no sensor data is
available. The discussion of 3D reconstruction results
demonstrated in section 4 will conclude the paper.
2. GENERATION OF FACADE GRAMMAR
The automatic generation of a facade grammar based on
terrestrial LiDAR data is the core of our facade modelling
approach. The first step is a data driven reconstruction process
aiming at the detection of geometric facade structures in the
observed point clouds. In this regard, a facade defines a planar
polygon with holes. Such holes indicate either windows, which
will be modelled as indentations, or salient structures such as
balconies, oriels or windowsills, which will be attached in the
form of protrusions. The result of the data driven facade
reconstruction serves as knowledge base for the generation of
facade geometries where no sensor data is available. This
knowledge, which includes information on dominant or
repetitive structures as well as their interrelationships, can be
inferred fully automatically and stored as a facade grammar.
While data collection will be described as a pre-processing step
in section 2.1, the basic concepts of the data driven
reconstruction and the subsequent grammar inference will be
addressed in section 2.2 and 2.3, respectively.
2.1 Data Collection
The StreetMapper mobile laser scanning system which was
used for our experiments collects 3D point clouds at a full 360°
field of view by operating four 2D-laser scanners
simultaneously. The required direct georeferencing during 3D
point cloud collection is realized by the integration of
observations from GPS and Inertial Measurement Units (IMU).
Figure 1 shows a 3D visualisation of the measured trajectory
overlaid to the 3D city model which was also used for the
following tests. This 3D city model is maintained by the City
Surveying Office of Stuttgart. The roof geometry of the
respective buildings was modelled based on photogrammetric
stereo measurement while the walls trace back to given building
footprints. The trajectory was captured during our tests within
an area in the city centre of Stuttgart at a size of 1.5 km x 2km.
The respective point clouds were measured at a point spacing of
approximately 4cm. Figure 2 depicts a part of the StreetMapper
point cloud at the historic Schillerplatz in the pedestrian area of
Stuttgart. The observed points are overlaid to the corresponding
3D building models in order to show the quality and amount of
detail of the available data. Another measured point cloud
overlaid to an existing coarse building model is shown in
Figure 3. This example is used in the following to illustrate our
bottom-up process for facade reconstruction. Within this
process, the geometric information inherent in the available
point cloud is exemplarily extracted for the facade marked by
the white polygon.
Figure 1. 3D city model with overlaid trajectory from mobile
TLS
Figure 2. Point cloud from mobile TLS aligned with virtual city
model
Figure 3. Lindenmuseum, Stuttgart: point cloud from TLS
aligned with existing coarse building model
