International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B8, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
Figure 1. The 3K+ camera system
The data from the GPS/Inertial system are used for direct
georeferencing of the images. Upon receiving the pre-processed
data from the airplane, the mobile ground station processes the
data and provides them to the end users via web-based portals
(Kurz, 2011).
GPS/IMU 3K* camera system CHICAGO on DLR glider
CHICAGO wing pod
Microwave datalink Ground station
Figure 2. Airborne hardware components and data flow of the
3K camera system for the real time processing chain
2.2 Onboard processing
The software running on the onboard computers must be
capable to process the incoming images in a way that the
produced data received on the ground is still up to date and of
use for the rescue forces. Moreover large data pile-ups caused
by a slow onboard processing module can stall the processing
system and must be avoided. These problems are quite likely to
happen because the detection and tracking of vehicles or
persons need high-resolution images in rapid sequence leading
to large amounts of data inside the processing chain.
Therefore, each camera has one dedicated computer for
processing the images. Before the actual detection of humans or
vehicles starts each image is pre-processed in two major steps.
Firstly, after the image is downloaded from the camera the IGI
system sends an event date with the exact time stamp, location,
and orientation of when the image has been taken to the
computer. The synchronization is done with the help of the
camera's external flash connector. Secondly, georeferencing
and orthorectification take place. The interior and exterior
camera parameters, determined by in-flight calibration (Kurz,
2012), and an SRTM DEM are loaded before take-off. After
determining the image bounding box the processor calculates
the intersection of each image ray with the sensor plane on the
34
graphics processing unit (GPU) rather than on the host's CPU.
The program works with NVIDIA's CUDA software library
and uses its special memory areas to accelerate the
orthorectification. As each pixel can be orthorectified
independently this calculation is well-suited for GPU
architectures. Only by leveraging the image-processing
capabilities of the video card’s GPU it is possible to provide
high-resolution orthorectified images to the thematic processors
on time.
One of the thematic processors extracts fully automatically road
traffic data from orthorectified images during the course of a
flight. This processing module consists of a vehicle detector and
a tracking algorithm. Images are acquired for traffic processing
in a so called burst mode. It consists of brief image sequences
of few images (3-5 images per burst) with a high repetition rate
(up to 3 fps). Every 5-7 seconds a burst is triggered, depending
on flight height and flight speed, so that there is nearly no
overlap between images of different bursts. This reduces the
amount of image data produced in comparison to a continuous
recording mode at high frame rate significantly. With this
technique we are able to perform automatic traffic data
extraction in real-time. To each first image of the burst, road
axes from a Navteq road database are overlaid, and vehicles are
detected along these roads. Vehicle detection is done by
machine learning algorithms AdaBoost and support vector
machine, which had been trained intensively on the detection of
cars offline prior to flight (Leitloff, 2010). Vehicle tracking is
performed between consecutive image pairs within an image
burst, based on the vehicle detection in the first image. In the
first burst image a template is produced for each detected
vehicle and these templates are searched for in the consecutive
images by template matching (Rosenbaum, 2010).
3. SYSTEM PERFORMANCE
In the following the quality and performance of the onboard
processing chain is evaluated. At first the quality of the
produced data is discussed and then the real-time performance
of the system.
3.1 Quality of Service
Products like ortho mosaics and traffic parameters should be
generated with sufficient geometric accuracy; 3 m absolute
horizontal position accuracy is assumed as sufficient in
particular for the import into GIS or road databases. Table 1
lists the horizontal and vertical georeferencing accuracy
separated for the post processing and real time case. For the
latter, the images are orthorectified based only on GPS/Inertial
system data and the global 25m-resolution SRTM DEM.
Post processing / Direct georeferencing:
Bundle adjustment
Otheor RMSE empir RMSE'
X |0083m |0.138m <3m
Y |0078m | 0365m <3m
Z | 0400m | 0.452 m n.a.
"Without DEM error, assuming GPS position error «0.1m, angle error of
inertial system <0.05°, flight height 1000m AGL
Table 1. Georeferencing accuracy of 3K+ camera system
given a bundle adjustment (left). It is only based on
GPS/Inertial system measurements (right).
Inter
Figure 2 €
algorithm.
showing tl
vehicle ve
from that :
results in 2
Quality —
Figure 2.
3.2 Real
During tl
forces of
buildings
orthorecti
Therefore
is able to
The ima;
system |
onboard |
As statec
along fli;
setup (b
orthorect
mainly d
Table 21
The 3K4
and can «
of 6.5 cn
