Istanbul 2004
the graphical
ie and image
in the image
inding part of
d and tested
vith a digital
| area. Using
r parameters
be masked
' of different
ow the only
ed estimators
this manner
| be reduced
igital map
»
7
ids all street
region have
image. This
from image
ze area from
cted vehicle
igation data
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B1. Istanbul 2004
values (height over ground, attitude of the aircraft). The
vehicle recognition of LUMOS and "Eye in the Sky" works on
single images. Approaches based on difference images or
estimated background images do not work reliably for test
flights with airplanes due to their fast speed over ground.
The vehicles have a variety of appearances in the captured
images depending on sénsor type, object properties and
environmental conditions (e.g. weather, temperature). But most
of the traffic objects can be recognized as coarse rectangular
shapes which contrast more or less with the background.
Therefore the algorithm searches for characteristic contours (of
suitable sizes) in edge images.
If a higher pixel resolution is available (visible camera),
further properties of vehicles such as the existence of special
cross edges can be included in the search process. Pixel values
themselves from the original images give additional
information for consolidation or rejection of vehicle hypotheses
or indications of the probable driving direction (Figure 7).
Evaluating the number of vehicles per scene gives a measure
for traffic density that can be provided to a central processing
computer.
i
Figure 7. Vehicle hypotheses of different size classes
High frame rates allow the determination of velocities. The
frame based information is now processed for successive
images in combination to determine vehicle velocities.
Virtual car positions are obtained from real car position data
from one image and navigation data from the following image.
Velocity vectors can be extracted by comparison of these
virtual car positions and real position data from the second
image. The repeated recognition of a car in the following image
emphasizes the correctness of the car hypothesis.
Assuming a time difference of 1/5 s between two images and a
pixel resolution of 0.5 m, velocities of 9 km/h can be detected.
On the other hand, a small car moving with 80 km/h does not
change its position from image to image by a value of its
length.
The vehicle recognition algorithm delivers a coarse size
estimation so that accepted car hypotheses can be divided into
a number of length classes. Using three classes has proven to
be very practical; a larger number reduces the exactness of the
classification. Thus essential shapes of cars, vans and long
vehicles can be estimated.
The traffic data extraction within the airborne traffic
monitoring projects is done per image and road segment first.
Densities and/or velocities are calculated for each vehicle class
from the obtained vehicle numbers and positions. The
extracted data for single images are combined for completely
observed road segments using size and position of the
overflown streets. The calculated average velocities and
densities per road segment of the digital map and per
timestamp can used now as input data for simulation and
prognosis tools.
5. VALIDATION OF THE SYSTEM
During the last years, several test flight campaigns within the
projects LUMOS and “Eye in the Sky" took place to validate
the quality and reliability of the system and especially of the
real time image processing part. The applied sensors and
auxiliary equipment can be integrated within two hours.
Different scenarios were flown (hovering vs. moving, visible
camera vs. infrared camera, different flight heights, different
illumination conditions).
The evaluation of georeferencing quality using
photogrammetric methods for calibration flights gives
accuracies in the range of one meter which is sufficient for the
requested applications.
The comparison of automatic vehicle identification within
LUMOS vs. manually counted in the images is shown in the
Figure 9. The image sequence with rate of 12.5 frame/sec was
captured on May 6, 2003 over Berlin-City Highway from the
flight attitude of 600 m (Figure 8). The images showing at
least 60 96 of segment of the road of 90 m length were taken
into account.
Figure 8. LUMOS-image of the Berlin City Highway
southbound (left part of the figure)
i
8 | ee eee ms A @—___—|—+_— manvally |
T T T TTT T oT m
6.7 8: 9 10 11 12 19.44.15 18 17 18 19 20 21
Image Number
Figure 9. Vehicle counting automatically with LUMOS vs.
manually per image
On average, the number of automatically counted vehicles is
115 % below the manually generated value. From the
averaged detection rates and the length of the observed road
section, vehicle number densities of 59.3 vehicles/km from
manual counts and of 52.7 vehicles/km from automatic counts
are obtained.
To verify a quality of algorithm for velocity determination a
special test car of DLR have been overflown for some test
flight while driving. The velocity measured on board of a test
