In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
471 
Figure 1: Topology of the onboard network. Each camera is di 
rectly connected to PC 1,2,3 which perform georeferencing and 
orthorectification. Vehicle detection and tracking takes place on 
PC4. PC5 sends the results via a proprietary S-band transmitter 
to the ground. 
suits presented in this publication are based on data processed of 
fline after flight at the ground for evaluating the processing chain. 
2.2 Test Sites and Imagery 
In the previous and current project the metropolitan areas Mu 
nich and Cologne have been selected for system flight tests and 
to demonstrate the airborne traffic monitoring system to the stake 
holders. Therefore the NAVTEQ road database installed on the 
system is limited to these regions. Hence, all data and results pre 
sented here are obtained during four campaigns in Munich and 
Cologne. 
First campaign was performed in Munich on 30.4.2007 with two 
flight strips across the city core and the main city ring road at 
sunny weather conditions and a flight height of 1000 m. Second 
campaign was flown in Cologne on 2.06.2009 at a flight height of 
1500 m. Third and fourth campaign were performed in Munich 
on 22.04.2010 and 24.04.2010 during BAUM A exhibition, both 
during fair to nice weather conditions at a flight height of 1000 m 
above ground. 
All images were resampled to a resolution of 20 cm GSD inde 
pendent from flight height, since vehicle detection is trained to 
that resolution. For evaluation in section 4 traffic data of the cam 
paign three and four were analysed. Due to the huge amount of 
images, not every flight strip of the campaigns was probed for 
manual reference. However, the analyzed flight strips give a good 
mix of scenarios ranging from downtown over suburban to rural 
areas including road categories from lowest category back road 
over inner city main road to interstate motorway. Results from 
the first and fourth campaign are shown in Figures 3, 4, and 6. 
During all campaigns images were taken in burst mode with 3 
images per burst and a repetion rate of 2 Hz. During flight strips 
the bursts were activated every 7 seconds. Image recording in 
burst mode makes vehicle tracking and measuring vehicle speed 
possible within burst image pairs but reduces amount of data pro 
duced significantly in comparison to a continuous mode with a 
high frame rate. Images were resampled to a resolution of 20 cm 
GSD 
3 PROCESSING CHAIN 
Automatic traffic data extraction works on the georeferenced im 
ages produced by the orthorectification process in the step before. 
Figure 2: Upper: Horizontal aligned original image. 
Lower: Confidence image with zero crossing. 
Preprocessing consists of a first module reading images from the 
camera ring buffer and receiving the corresponding exterior ori 
entation from the GPS/IMU navigation system in real time. Sec 
ond preprocessing module performs direct georeferencing / or 
thorectification (section 2.1). Each image burst is processed sep 
arately. Processing for traffic data extraction starts with vehicle 
detection on the first image of each image sequence (burst) fol 
lowed by a vehicle tracking based on template matching between 
the image pair consisting of the first and second image of the se 
quence. 
With this strategy the traffic parameters flux and density can be 
derived from image sequences. Vehicle detection performed on 
the first georeferenced image of the sequence gives the vehicle 
density whereas vehicle tracking applied on all consecutive (geo 
referenced) image pairs gives the velocity of the vehicles so that 
the flux of the traffic can be calculated. 
3.1 Vehicle Detection 
Prior to vehicle detection, image buffers along road axes obtained 
from NAVTEQ road database are cut out of the georeferenced 
images. This reduces the search space for vehicles in the images. 
Then all roads are straightened along road axes, so that vehicles 
travelling along each road in the image are aligned in the same 
direction. This reduces calculation time, since the classifiers of 
vehicle detection do not have to be rotated. 
In this work the vehicle detection is performed in three stages. 
The first step is based on the Gentle AdaBoost algorithm as in 
troduced in Friedman et al. (2000). Compared to other boosting 
variants, e.g. Real AdaBoost (Schapire and Singer, 1999) or Dis 
crete AdaBoost (Freund and Schapire, 1997), Gentle AdaBoost 
is less sensitive to errors in the training set as reported by Fried 
man et al. (2000) and Lienhart et al. (2003). Furthermore, Ad 
aBoost is the only classification algorithm with can handle the 
huge amount of features used in this work. Our training database 
contains 4662 sample of vehicles under different lightning condi 
tion and with various appearance. Furthermore, 15,000 negative 
samples are generated from designated areas. The complete set of 
standard and rotated Haar-like features as presented in Lienhart 
et al. (2003) are calculated for each sample in a 30 by 30 search 
window. The total number of 657,510 features for the horizon 
tal aligned gray value image of each training sample is reduced 
to the 200 most significant features during the training. Only 
these few features are used for later classification, which can be 
fast calculated by use of integral images as shown in Viola and 
Jones (2004). The application of the learned classifier results in 
confidence rated image reflecting the probability of each pixel be 
longing to a vehicle or the background respectively. Afterwards 
the confidence image is clustered by means of zero crossings as