1e XXXIX-B8, 2012
than on the host's CPU.
CUDA software library
eas to accelerate the
can be orthorectified
well-suited for GPU
the image-processing
it is possible to provide
o the thematic processors
s fully automatically road
s during the course of a
of a vehicle detector and
red for traffic processing
of brief image sequences
vith a high repetition rate
st is triggered, depending
that there is nearly no
bursts. This reduces the
mparison to a continuous
significantly. With this
automatic traffic data
image of the burst, road
overlaid, and vehicles are
> detection is done by
ost and support vector
sively on the detection of
010). Vehicle tracking is
e pairs within an image
in the first image. In the
luced for each detected
ed for in the consecutive
ium, 2010).
RMANCE
formance of the onboard
first the quality of the
he real-time performance
fic parameters should be
accuracy; 3 m absolute
sumed as sufficient in
road databases. Table 1
oeoreferencing accuracy
| real time case. For the
ised only on GPS/Inertial
tion SRTM DEM.
Áirect georeferencing:
RMSE'
«3m
<3m
n.a.
n error «0.1m, angle error of
ight 1000m AGL
of 3K+ camera system
(left). It is only based on
rements (right).
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 2 shows results of the vehicle detection and tracking
algorithm. Detected and tracked vehicles are marked by arrows
showing the direction of travel with its color representing the
vehicle velocity. The correctness of the traffic data obtained
from that scene was 95 %, the completeness was 85 %. This
results in a total quality of 81 % which is defined as
Onalte= truepositives 95 qq
truepositives + falsepositives + falsenegatives
Figure 2. 3K- scene obtained on 17? September 2011 close to
Cologne/Germany at a flight height of 1500 m AGL.
3.2 Real-time performance
During the first hours after arriving in an affected area rescue
forces often only need distances or spatial dimensions of
buildings or bridges to start working, so up-to-date
orthorectified images might be all they need in the beginning.
Therefore, it is interesting to know how fast the onboard system
is able to provide rescue forces with these images.
The image acquisition and the synchronization with the IGI
system hardly take any time compared to the succeeding
onboard processing modules and are neglected in the following.
As stated earlier, the 3K+ system can be installed across and
along flight track, respectively. This test uses the across track
setup (because of the larger coverage) in order to show
orthorectified land area as a function of processing time. It
mainly depends on the flight height and the changing GSD.
Table 2 lists typical flight heights and the resulting swath width.
The 3K+ system can cover a swath of 1280 m at 500m AGL
and can orthorectify 20 km? in 3.5 minutes (Fig. 3) with a GSD
of 6.5 cm.
3K+ camera system
Viewing directions 1x nadir, 2x £32° / variable
FOV +52° across
Coverage / GSD @ 500m 1280m x 240m / 6.5cm nadir
Coverage / GSD @ 1000m | 2560m x 480m / 13cm nadir
Coverage / GSD @ 3000m | 7680m x 1440m / 39cm nadir
35
Table 2. Coverage and GSD of the 3K+ camera system
140 T
xxx 500m AGL A
e*e 3000m AGL oo
1200 eee 1000m AGL os
S
o
$
$
S
1
60 md
Coverage in km?
i
40 eot aa
9 ey Ly
a ao echo?
x xX XXXXX
d o2o29? ka Kane x jx A
aco
KICK
Nm xc ic x (0008
298 23x x4 XXXAXÁ
96 50 100 150 200
Processing time in seconds
Figure 3. Coverage of onboard processed orthophotos as a
function of processing time during a flight.
If the operators at the ground station are more interested in a
larger overview of the scene the system can cover a swath of
almost 8 km if it climbs to 3000m AGL. 39cm-resolution
images of almost 140 km? can be sent to the ground after the
same processing time of 3.5 minutes (210sec in Fig.3). An
important result is the almost linear progress of the coverage at
all considered flight levels. If it were a logarithmic progress it
would mean that the image processing time cannot keep up with
the cruising speed of the airplane, which is typically at 136
knots due to the shutter speed of the cameras. In addition there
are almost always longer pauses between single flight strips
because of heading for other areas or limitations by flight
control. An example of in-flight generated images is shown in
Figure 4. In this case, the images cover an area of 35km”.
A good trade-off between coverage and resolution is flight
heights between 1000m and 1500m AGL because in this case
the GSD is small enough to get good results in further object
detection algorithms like the traffic processing module. The
already mentioned scene in Figure 2 was processed as part of a
larger image. When flying in traffic detection mode the time
between the bursts is used by the traffic processor to process the
last burst. With the current version (described in Section 2.2) it
is possible to complete the vehicle detection and tracking before
the next image burst is taken. After compressing the results the
system sends them directly to the ground with an average data
rate of 7 Mbit/s which is high enough to send all processed data.
These results show that the whole onboard processing system is
able to operate in real time.
