ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision“, Graz, 2002
3. The rotation of the new image is improved by mini-
mizing the same cost function we used above.
4. The corresponding points of the new image are either
used to calculate new or to improve old 3D points.
5. As long as adjacent images are left, go to step 2.
3.4 Image to 3D Model Fitting
So far we have only obtained a relative oriented sequence,
where the position and orientation in geo-referenced co-
ordinates as well as the scale is not known. The upgrade
from the relative orientation to a geo-referenced orienta-
tion of all images needs at least three well distributed con-
trol points. This upgrade involves a non neglecting manual
effort, especially working with large image sequences. In
our approach only two images of a continuous sequence
have to be fitted in a semi-automatic way, so the manual
work is minimized. Due to the fact, that the vertical di-
rection of the images is known from vanishing points only
two control points are necessary to transform the image
into a geo-referenced coordinate system. This two control
points are taken from the aerotriangulation. With the so ob-
tained camera positions in geo-referenced coordinates for
at least two images and the corresponding relative posi-
tions we calculate a transformation matrix that solves the
orientation upgrade for the whole sequence.
3.5 Bundle Block Adjustment
For each building block a bundle adjustment is carried out
to optimally fit the images into the existing block model. The
block geometry ressembles the classical photogrammetric
case of aerotriangulation: most of the image sequence is
connected with corresponding points only, while known
control points in some images help to fix the transforma-
tion to the world coordinate frame and stabilise the block,
particularly in long sections of translational camera move-
ment. The control point accuracy (i.e. the point accuracy
of the GIS data) is known, so that correct weights can be
assigned to the control points in order to account for their
limited relative accuracy. Since the vertical direction is
known from the detected vertical vanishing points in the
terrestrial images, a constraint is added to keep the mean
vertical direction of the cameras unchanged.
4 RESULTS
4.1 Accuracy
We distinguish between two different kinds of accuracy;
the relative and the absolute accuracy. The absolute ac-
curacy of the reconstruction is limited by the accuracy of
the ground truth, in our case of the GIS data acquired from
aerial images. The control points are the intersections of
the roof lines from aerial photogrammetry, which were mea-
sured with an accuracy < 10cm.
Our obtained relative accuracy is about 0.25 pixel, far be-
low one pixel reprojection error. The corresponding ob-
ject point accuracy strongly depends on the recording con-
figuration. It can be obtained from the bundle block ad-
justment, if the accuracy of the image measurements is
known. Thus, further analysis about the accuracy of the
corresponding points delivered by our algorithm has to be
undertaken. A good starting point will be the work of
Forstner [4] where an optimal estimation for uncertain ge-
ometric entities is given.
4.2 Geo-Referenced Orientation
Figure 6 shows two sequences of images which were ori-
ented semi-automatically.
t s Q mte
s artt site fa
UTE.
Figure 6: Geo-referenced orientations of two image se-
quences of the Schlossbergplatz. Each camera is repre-
sented by a small plane and a direction arrow. At each lo-
cation we shot two images with a short vertical basis. The
reconstructed POIs are indicated by dots.
43 Image to 3D Model Fitting
The graphical user interface (GUI) used for fitting of some
images to the 3D block model can be seen in Figure 7.
After an aerial and an image sequence is loaded, it is pos-
sible to update the whole sequence from a relative to the
geo-referenced coordinate system with only a few mouse
clicks. To verify the right geo-referenced orientation the
terrestrial images are superimposed by roof lines.
5. CONCLUSIONS AND FUTURE WORK
We have presented on image based documentation system
called MetropoGIS in which 2 1/2D GIS data is augmented
with terrestrial photographs. We showed that its possible to
determine the geo-referenced orientation of a large image
sequence with only a few mouse clicks without any posi-
tion estimation. Because we use a hand held camera, our
image acquisition is straightforward and allows high flex-
ibility. By exploiting improved lines and vanishing points
we developed a robust and fast method to determinate rel-
ative orientation even large baseline is present. So far we
have a semi-automatic system. An operator is still involved
in our work flow. To avoid this manual intervention we are
A - 191
