CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation 
178 
relation to the rate of false positives (number of false detected 
objects/number of detected objects). To be honest the false pos 
itives rate is not very objective, as the number of false detected 
objects does not depend on the real number of cars, and could 
turn out very bad just in case there is only one car in the image. 
Therefore in (Lei et al„ 2008) they consider the FP-number in 
relation to the length of streets. A still better way would be to 
take the street area, for example ’false positives per hectare’. In 
this example there is an optimal point, where detection reaches 
80 percent while the FP-rate is only ten percent or one car per 
hectare. 
A rather bad sample (the worst in our evaluation) represents the 
Figure 7: Detection rates on a highway (red) and narrow 
streets (blue) depending on false positives per detected cars 
Figure 8: Detection rates on a highway (red) and narrow 
streets (blue) depending on false positives per area 
blue graph in fig.7 and 8 where more than 300 cars have been 
clicked by hand. If we consider streets of all sizes in the Munich 
suburban area, on the one hand the detection time takes longer 
(more than 60 seconds) and the results become worse as well. 
The detection rate stays around two thirds while only the number 
of false positives rises from 5 up to 25 per hectare. 
A reason for the bad detection rate in the second example is the 
accuracy of street coordinates. As many smaller street elements 
are drawn next to the real street (fig.9) the algorithm misses many 
cars while detecting some rectangular structures next to the street. 
An approach to avoid this might be to improve the street accuracy 
by alternative street databases or street detection which should 
not be considered in this paper. 
Figure 9: False positives and negatives due to incorrect co 
ordinates (blue - existing car, red - found car, yellow - found 
truck) 
3.2 Tracking 
Figure 10: Correctness rate of tracks depending on the param 
eters cr and 7 
We implemented the tracking algorithm as explained above by 
using the vehicles distances on UTM-projection and the normed 
correlation coefficient of all three color channels in a 20-by-20- 
pixels window around them. As the images cover an area of 700 
on 1000 meters with hundreds of cars each, it is not easy to show 
how the whole set of tracks looks. That’s why only one street 
was picked out for visualization. In fig. 10 the resulting track 
ing rates depending on the parameters a and 7 are shown. As 
one can see the best results we get if a is between 20 and 30. If 
the value is too small (<j = 5) the dependence of the positions 
among each other is not respected enough. This results not only 
in incorrect assigned pairs but also in crude mistakes by assign 
ing objects together which are located very far from each other. 
This can strongly falsify the measured velocities. Furthermore 7 
should neither be too small nor too high. The best results yield 
values between 0.4 and 1.0. Around these settings a correctness 
of more than 80 percent (best value 85.7%) is achieved. 
As for the average velocities it is rather important to accept cor 
rect tracks than getting all vehicles tracked, after the SVD the 
acceptance is bound to the correlation coefficient of a pairing. If 
the pairing next to its ranking in row and column does not pass a 
threshold for the CC, it is discarded although it might be correct. 
In fig. 11 the remaining tracks are shown. In the upper half of 
the image 49 objects have been detected. 39 of them have been 
detected in the lower half as well which means they are possible 
to track. 36 of the objects have been assigned to another one, 30 
of them were assigned correctly. After the thresholding with a 
CC of 0.9 still 26 of the 36 tracks remain. So from end to end