The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bib. Beijing 2008
483
oriented scene is about 4 degree and can be reduced to about
1.5 degree. In addition to a global adjustment, every estimated
surface plane can be corrected locally. So, for this scenario with
several façades, only the façade at the front of the street can be
extracted. Other façades are at least partially occluded and can
not be estimated correctly.
5.3 Comparison to other scenarios
In scenario 2 and 3, the façade is tilted in moving direction
caused by the small upwards viewing angle of the camera. The
distance to the building is adequate with about 1 meter and the
shift is in moving direction is about 4 meters. Both values are
somehow higher because the GPS coordinates are afflicted with
a bigger error. This is caused by the record situation. Whereas
in scenario one, you can find an open park landscape on the
opposite side of the building, in scenarios two and three, there
are buildings on the opposite side occluding the satellite signal
and so leading to a bigger positioning error. For scenario two,
there is a special case (see Figure 6). The first part of the façade
has a very regular structure and thus the algorithm finds many
wrong pairs of points of interest (Fig. 6a). This causes long
computation steps and wrong planes estimations. This effect
depends on the points chosen by RANSAC and produces
changing hypotheses for every computation. A second problem
of this façade is the gateway at the beginning (Fig. 6b). This
gateway is standing back from the façade generating points of
interest behind the surface plane. Together with pillars in front
of the gateway, which are standing in the surface plane, a point
cloud is computed with points jumping between the surface
plane and the gateway plane. In this area, no surface plane can
be estimated.
Fig. 6: Scenario two, a) regular façade structure causing wrong
point pairs, b) gateway behind the façade plane with
pillars in front.
In scenario 4 only parts of the façade can be extracted due to
the occlusion caused by the bridges. The façade parts before
and after the bridges can be estimated to about 70%. The
algorithm aborts when the bridges cover more than 40% of the
image. The positioning errors for the extracted planes are 5
meters along the façade and 2 meters to the correct façade plane.
6. DISCUSSION
Over all scenarios, some facts can be integrated. For all
scenarios, the upward angle of the camera view • cuased a
constant tilt of 6 ° in the found surface planes. Façades parallel
to the moving direction that are not occluded in parts can be
extracted with about 85 to 90 % when the façade has an almost
plane structure. The positioning errors are in the range of the
accuracy of the GPS. The scenario with the bridges can not be
extracted with appropriate quality, but this is a very special case.
An overview over the errors in the different scenarios is given
in table 1.
Scene
I
2
3
4
Pose error model
X
3 m
4 m
4 m
5 m
Y
Z
!4 m
1 m
1 m
2 m
Tilt error
6°
6°
6°
6°
Pose error GPS
3 m
4 m
4 m
5 m
Angle x-z-plane error
4 ‘/2°
4°
4 1 / 2 °
6 0
Completeness façade
90%
85 %
90%
40%
Completeness model
55 %
80%
90%
40%
Table 1: Errors in the different scenarios: Pose error model:
Error of the estimated position to the given model in
X, Y and Z axis, Tilt error: Tilt of the planes caused
by upward viewing angle of the camera, Pose error
GPS: mean error of GPS position to corrected
estimated position, Angle X-Z plane error: Error in
the estimated camera path direction, Completeness
façade: Percentage of completeness of a detected
façade, Completeness model: Percentage of
reconstructed façade surface of the hole scene
including non-detected surfaces
7. CONCLUSION
The used algorithm was originally designed for high resolution
images that show an object from different positions and viewing
angles. The results show that the 5-point-algorithm can also be
used for image sequences with constant viewing direction, if
there are façades parallel to the moving direction. Façades
observed in nadir view do not show enough movement of points
of interest to extract surface planes. The images taken for the
extraction do not need to be given in a high resolution. A small
resolution as given in the scenarios of this paper often contains
enough information to extract sufficient points of interest.
Problems occur if the façade contains only few structure and if
the structure is regular. When a façade is standing along the
street with only little occlusion, the façade can be reconstructed
with up to 90 %. For special façades like mentioned in chapter
4 the algorithm is not robust enough and reconstructs only small
parts of the façades.
The algorithm does not only provide estimated surfaces but also
an estimated camera path. Results show, that it is possible to
reduce the position error caused by the GPS by combining the
GPS and the given 3d model with the estimated surface planes
and the estimated camera path. The estimated camera path is
afflicted by a much smaller error than the GPS measurement.
The GPS position are only necessary for the scaling, rotation
and translation of the estimated model to fit the given 3d model.
The position refinement is then done without the measured
camera path comparing the positions of the façades and the
surface planes.
8. OUTLOOK
Two further steps are now possible: The corrected camera path
can be used to extract the infrared textures from the image
sequence. Or, the 3d model surfaces are included in the plane
estimation process to precise the estimated planes and thus