The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
9 
weighed higher in the adjustment, which distorts the result. A 
solution could be to rotate the coordinate axes for slopes higher 
then 45° (Schwermann, 1995). 
The results of determining the exterior orientation of a chosen 
camera from the VIDS are shown in Figure 9. The RMSE is 
given in meter in object space to emphasize the maximum 
accuracy when using the camera as a traffic detector. All 
approaches based on control points converge to the same local 
minimum. Their RMSE is less than 0.05m up to a ground 
distance of 80m and less than 0.15m up to a ground distance of 
140m. This theoretical ground sampling accuracy is more than 
sufficient for a traffic sensor. It potentially allows a correct lane 
assignment of detected vehicles over the whole observation area. 
The accuracy is higher then the expected localisation accuracy 
of traffic relevant objects in the scene yielded by image 
processing. Nevertheless, the Gauss Markov approach based on 
lines as input doesn’t properly converge to the same local 
minimum, for the same reasons mentioned above. The RMSE is 
less than 0.8m up to a ground distance of 80m and less than 
1.5m up to a ground distance of 140m. 
Altering five of the input ground control points erroneous still 
leads to similar results for the algorithms based on adjustment. 
The worst case boundaries don’t change (Figure 10). The 
Newton method performs out of scale because of the absence of 
a statistical model. 
Figure 11 shows the results of the Gauss Markov approach for 
points using trigonometric functions and varying the amount of 
control points used. Surprisingly, this has no noticeable effect 
on the overall accuracy when using at least 10 control points. 
Figure 5. RMSE when adding noise to control points in 
simulated setup 
Figure 6. RMSE when adding erroneous image points in 
simulated setup 
^ <> Space Resection 
O o DLT 
-0—o—-o-' 
-0- -0 
20 30 40 
erroneous imoge points [%] 
Figure 7. RMSE when adding erroneous image points in traffic 
intersection setup 
Figure 8. RMSE when adding noise to control points in 
simulated setup