4.1 Hardware 
The platform we used in order to carry out our experiments is a 
prototype device for mobile photogrammetry (see Figure 4). It 
consists of a standard resolution color video camera with 
extreme wide-angle lens, a GPS receiver, an electronic compass 
and a tilt sensor. By combining image data and orientation data 
we obtain an image with an approximate exterior orientation. 
The camera is a consumer style Sony DFW-500 video camera 
connected to the system via IEEE 1394 also known as FireWire. 
The camera was calibrated on a test-field using a ten parameter 
camera model, bundle adjustment was performed using 
Australis (Fraser, C. S. 1997). The additional parameters 
obtained were used for the computation of the resection, while 
rendering was done with a reduced parameter set. The GPS 
receiver is a Garmin LP-25, which can be operated both in 
normal and differential mode. We used the ALF service 
(Accurate Positioning by Low Frequency) of Deutsche Telekom 
for differential mode, thus obtaining a correction signal every 
three seconds. Our experience shows that the system allows for 
a determination of the exterior orientation of the camera to a 
precision of 7-10 m in planar coordinates. The orientation 
accuracy provided by the digital compass and the tilt sensor 
resulted in an error of 1° — 2°. 
  
Figure 5: Image of the opera house with building silhouette overlaid. On the left the approximate orientation is used, on the right the 
All the devices are connected to a laptop. While the camera and 
compass/tilt sensor are hand held, the GPS is attached to a 
backpack. Both the model computation and the image 
processing was implemented on a standard PC / laptop. 
4.2 Recognition Results 
Two examples for the recognition of buildings using our 
approach are given with the opera house in Figure 5 and with a 
museum in Figure 6. In both examples perfect matches were 
obtained. This is the case even though one can clearly see the 
problems, which are encountered in realistic situations. For the 
opera house a tree is obstructing the view partly occluding the 
building. A shadow is cast across the facade generating 
additional gradients. The statues on the roof are not represented 
in the model. Still the algorithm is able to detect the overall 
shape of the building, demonstrating the robustness of our 
approach. 
While the initial position of the silhouette in the image in Figure 
5 seems correct, it appears to be too large indicating an 
approximated position in object space too close to the building. 
For Figure 6 we can see a clear misalignment of the silhouette 
caused by false orientation. The method is able to correct these 
errors in both cases. 
result of the detection is displayed. Detection was successful despite the occlusions, shadowing effects and model imperfections. 
  
Figure 6: Image of a museum with building silhouette overlaid. On the left approximate orientation, right after detection. 
—494— 
sig 
bu 
rec 
ex| 
ou 
dis 
ele 
ori 
we 
ou 
Ta 
op 
Vie 
res 
gr 
use 
and 
SEV 
R 
Tal 
Ov: 
fro 
In 
is € 
pos 
ind 
ind 
str 
Fig 
fou 
in €