CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
Figure 1: Model refinement process
3 BUILDING RECONSTRUCTION FROM
TERRESTRIAL LASER SCANNING
Pu and Vosselman (2009) presents an automatic approach to ex
tract building facade features from a terrestrial point cloud. The
method first defines several important building features based on
knowledge about building facades. Then the point cloud is seg
mented to planar segments. Finally, each segment is compared
with building feature constraints to determine which feature this
segment represents. The feature extraction method works fine for
all facade features except for windows, because there are insuffi
cient laser points reflected from window glass. Therefore a hole
based window extraction method is introduce. Then polygons
to extracted feature segments and the merging of polygons to a
complete facade model. An advantage of this approach is that
semantic feature types are extracted and linked to the resulting
models, so that i) it is possible to get faster visualization by shar
ing the same texture for same feature type; ii) polygons can be
associated with various attributes according to its feature type.
Figure 2 shows a building facade model which is reconstructed
with the above approach. The generated building outline seems to
coincide with laser points well. However, if we take a close look,
it is easy to identify several mistakes from the model. By analyz
ing more models, we figured two main reasons for the modeling
errors. They are: •
• Limitations of outline generation method. For example, side
wall’s eave can ’’attract” the side boundary edges of the fa
cade, and result in a slight wider polygon in horizontal di
rection. The almost vertical or horizontal edges are forced
to be vertical or horizontal; however, this is not always ben
eficial.
• Poor scanning quality. Due to the scanning strategy of static
laser scanner, complete scanning of a scene seems impossi
ble. There are always some parts which contain very sparse
laser points, because of their visibility to any of the scan po
sitions. Occluded zones without any laser points are also
usual in laser point clouds. The lack of reference laser in
formation leads to gaps in the final model. For example, the
lower part of roofs are hardly scanned because the eaves oc
clude the laser beams. The directly fitted roof polygons are
smaller than their actual sizes. Sometimes these gaps are
foreseen and filled using knowledge. For example, we know
a roof must attach to the upper side of an eave, so we can ex
tend the roof polygon so that it intersects the eave. However,
knowledge based estimation are not always correct.
Figure 2: A reconstructed building facade model, shown together
with segmented laser points
4 PREPROCESSING
In order to extract straight lines, an image need to be in central
perspective and undistorted. The exterior orientation parameters
and focal length should be determined so that 3D model edges can
be projected to the same image space for comparison. These are
the two objectives of the preprocessing step. An omni-directional
panoramic image called Cyclorama is used in our method devel
opment, therefore conversion of Cyclorama to central perspective
are explained first in 4.1. A semi-automatic approach for exterior
orientation calculation is given in 4.2.
4.1 Perspective conversion of Cyclorama
The Cycloramas are created from two fisheye images with a field
of view of 185 degree each (van den Heuvel et al., 2007). The
camera is turned 180 degree between the two shots. The Cy
cloramas we used contain image data for the full sphere stored
in a panorama image of 4800 by 2400 pixels, corresponding to
360 degree in horizontal direction and 180 degree in vertical di
rection. Thus, on both directions the angular resolution is 0.075
degree per pixel. With the integrated GPS and IMU devices, all
Cycloramas are provided with north direction aligned at x=2400
and horizontal plane aligned at y=1200.
The equiangular projection of the fisheye camera model is de
scribed in Schneider and Maas (2003). The projection of Cy
cloramas to central projective can be understood as projecting a
218
