3. MODELLING
3D modelling for both objects was based on point clouds
combined from airborne and terrestrial laser scanning data.
Both data sets were combined based on their common
georeference — EPSG:2180 coordinate system. There were not
performed any procedures designed to improve the matching of
both sets. In case of need or necessity to improve the matching
of the two considered data sets, the algorithms of transformation
must be used, for example the method proposed in the work of
Gruen and Akca (2005), which is based upon homologous
surfaces matching with use of generalized least squares method.
3.1 Geometry reconstruction
Modelling was performed using the Cyclone Model 7.1
software. The process of reconstruction a 3D vector model of
building can be divided into the following stages (Borkowski et
al., 2011):
Scanning data approximation by planes. Due to ease the
subsequent texturing, all the elements of the building and
architectural details were approximated by planes or theirs
compounds. The algorithm of growing regions was used to
identify the various planes in the data set.
Modelling the edges of the building. The edges of the building
were modelled as a result of the intersection between
neighbouring planes and were extended to the edges.
Checking and correction the topology. This problem was,
illustrated in Figure 3. It occurs most often in modelling roofs
of buildings or other at least four neighbouring planes. Three
roof planes define only two edges. The next roof plane defines
additional two edges, which often do not intersect with the
previous ones at one point.
Building walls extending. Each wall which has the connection
with the ground has to be extended to DTM. In this work DTM
was created as a mesh from ground points of airborne laser
scanning data set. Walls were extended to theirs intersections
with DTM mesh.
Final 3D vector model creating. After geometry reconstruction
each building model was exported to DXF format.
ight: correct topology.
: incorrect topology.
: Figure
3.2 Texturing
Texturing process consisted of imposing individual digital
images of the same spatial resolution for each plane of 3D
vector model. It was found that the size of one pixel on the
object (wall) equal to Scm is sufficient for the correct
visualization of the model, while created textures will not be too
big, so the final model can be presented via the Internet in an
effective way. Transformation and trimming digital photos to
the individual planes of the 3D model with simultaneous
resampling of images to the chosen spatial resolution is made in
original software developed at the Institute of Geodesy and
Geoinformatics. This software converts images using the
projective transformation (there is need to select a minimum of
4 homologous points on photo and model), and the resampling
is being done using bilinear interpolation. For some textures, it
was necessary to remove foreign objects such as cars or trees.
This task was performed with use of the free software Gimp 2.
Assignment and orientation of each to the appropriate model
was performed in another free software - Google Sketchup 8.
Visualization of the final models is shown in Figure 1.
4. ACCURACY ASSESMENT
4.1 Error sources
Discussing the accuracy assessment of created 3D model first
the errors that affect the final accuracy of the model have to be
considered:
1. The accuracy of TLS data.
2. The accuracy of ALS data.
3. Errors of integration of both data sets.
4. Generalization of the model (level of detail
modelling) and unambiguous identification of
individual components of the model (modelled
surface roughness).
5. Errors resulting from the topology correction.
6. Texturing errors.
The characteristics of both sensors show that the accuracy of the
TLS data, both horizontal and vertical is at a few centimetres
level and is about an order of magnitude better than the
accuracy of the ALS data. Method of the two point clouds
integration must therefore have an impact on the final accuracy
of the model. In the presented work, integration was based on
a simple combination of data sets. This approach is pragmatic
and results from the assumptions made by the user of models. In
Figure 2 there is shown that in practically all areas of the
building there can be found points from both airborne and
terrestrial scanning, but obviously with different resolutions.
This fact was used to assess some kind of models internal
accuracy, or evaluate the accuracy of the two point clouds
matching.
4.2 Internal accuracy
Figure 4. ector model of the same building created from
airborne (green) and terrestrial (grey) scanning data only.