In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3AA/4 — Paris, France, 3-4 September, 2009
2.2 Data Driven Reconstruction
Frequently, the representation of buildings is based on
constructive solid geometry (CSG) or boundary representation
(B-Rep). In contrast, we apply a representation of the buildings
by cell decomposition (Haala et al., 2006). By these means,
problems which can occur during the generation of
topologically correct boundary representations can be avoided.
Additionally, the implementation of geometric constraints such
as meeting surfaces, parallelism and rectangularity is simplified.
Due to the applied representation scheme, the idea of our
reconstruction algorithm is to segment an existing coarse 3D
building object with a flat front face into 3D cells. Each 3D cell
represents either a homogeneous part of the facade or a window
area. Therefore, they have to be differentiated depending on the
availability of measured LiDAR points. After this classification
step, window cells are eliminated while the remaining facade
cells are glued together to generate the refined 3D building
model. These steps are depicted exemplarily within Figure 4
and Figure 5, and will be explained in the following sections.
The processing is based on the facade and point cloud marked
by the white polygon in Figure 3.
~ „
y~T
ti-
i fTf
1
•
[;
• j
- -
!
|*
i ■
i
1
f:
; _
M
\ i :
i I
|
:
1
1
1 M
1 1»
1
K
1
"I
j
Figure 4. Lindenmuseum, Stuttgart: LiDAR point cloud (left),
and detected edge points and window lines (right)
3 3
m n
Figure 5. Lindenmuseum, Stuttgart: classified 3D cells (left),
3D facade model (middle), and refined 3D facade
model (right)
edge points. Figure 4(right) shows the extracted edge points at
the window borders as well as the derived horizontal and
vertical lines. Based on these window lines, planar delimiters
can be generated for a subsequent spatial partitioning. Each
boundary line defines a partition plane which is perpendicular
to the facade. For the determination of the window depth, an
additional partition plane can be estimated from the LiDAR
points measured at the window crossbars. These points are
detected by searching a plane parallel to the facade, which is
shifted in its normal direction. The set of partition planes
provides the structural information for the cell decomposition
process. It is used to intersect the existing building model
producing a set of small non-overlapping 3D cells.
2.2.2 Classification and Reconstruction
In order to classify the 3D cells into facade and window cells, a
point-availability-map is generated. It is a binary image with
low resolution where each pixel defines a grid element on the
facade. The optimal grid size is a value a little higher than the
point sampling distance on the facade. Grid elements on the
facade where LiDAR points are available produce black pixels
(facade pixels), while white pixels (non-facade pixels) refer to
no-data areas. The classification is implemented by computing
the ratio of facade to non-facade pixels for each 3D cell. Cells
including more than 70% facade pixels are defined as facade
solids, whereas 3D cells with less than 10% facade pixels are
assumed to be window solids. While most of the 3D cells can
be classified reliably, the result is uncertain especially at
window borders or in areas with little point coverage. However,
the integration of neighbourhood relationships and constraints
concerning the simplicity of the resulting window objects
allows for a final classification of such uncertain cells. Figure
5(left) shows the classified 3D cells: facade cells (grey) and
window cells (white).
Within a subsequent modelling process, the window cells are
cut out from the existing coarse building model. Thus, windows
and doors appear as indentations in the building facade which is
depicted in Figure 5(middle). Moreover, the reconstruction
approach is not limited to indentations. Details can also be
added as protrusions to the facade (Becker and Haala, 2007).
However, the achievable level of detail for 3D objects that are
derived from terrestrial laser scanning depends on the point
sampling distance. Small structures are either difficult to detect
or even not represented in the data. Nevertheless, by integrating
image data with a high resolution in the reconstruction process
the amount of detail can be increased (Becker and Haala, 2007).
This is exemplarily shown for the reconstruction of window
crossbars in Figure 5(right).
2.3 Automatic Inference of Facade Grammar
2.2.1 Point Cloud Segmentation
At glass LiDAR pulses are either reflected or the glass is
penetrated. Thus, as it can be seen in Figure 4(left), by laser
scanning usually no points are measured in the facade plane at
window areas. If only the points are considered that lie on or in
front of the facade, the windows will describe areas with no
point measurements. These no-data areas can be used for the
point cloud segmentation which aims at the detection of
window edges. For example, the edge points of a left window
border are detected if no neighbour measurements to their right
side can be found in a predefined search radius. In a next step,
horizontal and vertical lines are estimated from non-isolated
As it is already visible in Figure 3, the given scan configuration
resulted in considerable variations of the available point
coverage for the respective building. Thus, the bottom-up
facade reconstruction presented in the previous section was
realized for a facade, which is relatively well observed. This
overall result is now used to infer the facade grammar.
Frequently, such formal grammars are applied during object
reconstruction to ensure the plausibility and the topological
correctness of the reconstructed elements (Muller et al., 2006).
In our application, a formal grammar will be used for the
generation of facade structure where only partially or no sensor
data is available.
231