In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
i 
473 
Time [seconds] 
Figure 5: The diagram shows the amount of data which is pro 
cessed by the Ortho Module but not by the Traffic Processor in 
case of full traffic data extraction including all road categories 
and urban core scenarios. 
Campaign 
BAUMA 1 
BAUMA 2 
Total 
Bursts evaluated 
29 
19 
48 
True Positives 
1002 
724 
1726 
False Positives 
75 
47 
122 
False Negatives 
36 
23 
59 
Correctness 
93% 
94% 
93% 
Completeness 
97% 
97% 
97% 
Quality 
90% 
91% 
91% 
Table 3: Evaluation of vehicle tracking quality. 
4 RESULTS AND ACCURACY 
Quality of traffic data obtained from the processing chain was 
evaluated on image sequences obtained on Cologne and ”BAUMA” 
Munich campaigns. The Cologne image data were used for val 
idating detection accuracy, whereas Munich data were used for 
evaluating tracking. The data represent a mix of scenarios rang 
ing from urban core and suburban roads to motorways. Due to 
the higher complexity of the scenes, the quality of traffic data 
extracted in metropolitan cores is somewhat lower than on mo 
torways and suburban areas. This deposits in vehicle detection 
notably. Therefore, the quality of vehicle detection is examined 
while distinguishing between these different scenes. However, 
each campaign contains a mix of both scenes, since German city 
cores are typically limited to a diameter of a few kilometers, even 
in German major cities. Therefore, a total quality containing both 
scene types is listed as well, giving the typical quality of the de 
tection to be expected in operational use. The evaluation of track 
ing distinguishes between the two Munich ”BAUMA” campaigns, 
since the weather and illumination conditions were better at the 
flight on 22.04. The definitions for completeness, correctness and 
quality are: 
true positives 
true positives-^ false positives 
true positives 
true positives+false negatives 
true pos 
true pos+false pos+false neg 
with true positives being the number of vehicles detected, false 
positives the number of non-vehicle detections, and false nega 
tives the number of vehicles missed. In tracking evaluation true 
positives is the number of vehicles tracked correctly, false posi 
tives the number of mistracked vehicles, and false negatives the 
number of vehicles detected, but not tracked (vehicles that are ob 
Correctness = 
Completeness = 
Quality — 
scured during image bursts and cannot be tracked are not counted 
as false positives). 
Table 2 shows the results of detection algorithm. With a value 
of around 90 % correctness of vehicle detection is at high level in 
total and in suburban and motorway situations. In urban cores the 
correctness drops slightly to 87 %. The completeness with 93 % 
in suburban and motorway scenes and 92% in total is at high 
level again providing precise estimates of local traffic density for 
traffic simulations or road level of service visualization. In urban 
core regions it drops a little to a value of 90 %. This all results 
in a total quality of 83 %. Table 3 presents the results of evalu 
ating vehicle tracking algorithm. The underlying image database 
represents a mix of urban core and suburban/motorway scenes. 
Tracking performs well with a correctness of about 93 % and a 
completeness of 97%. There is nearly no difference between 
the results obtained from two flights at different dates indicating 
that tracking algorithm might be robust against slight changes in 
weather and illumination conditions. 
The system accuracy is the product of detection and tracking 
quality. In the present case it is about 75 %. This makes the 
system accuracy competitive to that of ground based sensor net 
works. The quality of a typical induction loop is better than 90 % 
(e.g. Leonhardt, 2004) at time of production. During years of 
continuous operation it decreases slowly. In a complete sensor 
network of a metropolitan area there is a mix of new and older 
sensors, and some of them have even failed completely. This 
drops down the average quality or system accuracy of the sensor 
network. In Munich, the system accuracy of the complete ground 
based sensor network for traffic data acquisition is at a value of 
80%, as stated by Munich integrated traffic management cen 
ter. The quality of traffic data obtained by the presented remote 
sensing system being competitive to that of road sensor networks 
in connection with the good spacial resolution (including minor 
roads) and the reliability in case of ground damage makes the sys 
tem well suited for its operation during mass events and disasters. 
5 CONCLUSIONS AND FUTURE WORK 
The quality of traffic data obtained by the presented remote sens 
ing system being competitive to that of road sensor networks 
in connection with the good spacial resolution (including minor 
roads) and the reliability in case of ground damage makes the 
system well suited for its operation during mass events and disas 
ters. In the evaluated case with a mixture of urban core, suburban 
roads, and (local) motorways, the overall system accuracy is ap 
proximately 75 %. This accuracy value is comparable to that of 
typical metropolis ground based sensor networks with the advan 
tage of the remote sensing system, that traffic data on all road cat 
egories (including smallest minor roads) can be recorded. E.g. in 
case of traffic congestion on main roads and motorways not only 
the traffic density of the main road but also densities on alterna 
tive routes (that are not covered by induction loops or stationary 
traffic cameras) can be determined. However, the operating costs 
of this prototype system are quite high which limits its applica 
tion to temporally limited scenarios like mass events or disasters. 
With the ongoing development of UAV or HALE unmanned air 
crafts, further applicability of the system may occur in future. 
The system is sufficient for real-time traffic data extraction from 
aerial image time series on main roads, even in metropolitan core 
scenarios. In case of full traffic data extraction including all 
roads in urban core scenes with several thousands of vehicles 
per scene, a slight lack of performance in the traffic data algo 
rithms is present. In order to close this gap, two measures are 
taken. Firstly, Adaboost classifier will be cascaded. Secondly,