er
2)
re
U-
3y
718
TS.
sed
>n-
ind
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
- [ni]
r^ a
EXE
.] D E
Figure 4: A Sensor alignment on a Japanese railway station,
where #1 to #12 and #C1 to #C8 represent the laser and camera
positions respectively.
3 EXPERIMENT
An experiment is conducted at the concourse of a railway station,
which has a dimension of 30m by 20m. Sensors’s alignment is
shown in figure 4. In this experiment, we applied two types of
sensors; video camera and laser scanner. Twelve laser scanners
are set on the floor to cover most part of the concourse. Every
scanner is controlled by computers that are physically connected
by 10/100 Base LAN.
Then, we set eight video cameras on the ceiling to assess the pro-
posed system. Figure 4 shows a concourse plan and a sensor
alignment. In this plan, six cameras film most part of the crowded
area by the nadir images (#C3 to #C8) and the rest two cameras
film whole aspects by the diagonal images (#C1 and #C2).
4 RESULTS AND DISCUSSION
4.1 Pedestrian Detection
Although we used twelve laser scanners in this experiment, all of
the sensors could not be integrated because network trouble was
occurred to #8 and #10-#12. Therefore, pedestrian detection was
conducted by using eight sensors excluding troubled four sen-
sors. Figure 5 shows a result of pedestrian detection, where white
circles show recognized pedestrians and white lines show those
trajectories. White points near the wall show laser scanners. This
image is captured when concourse was not crowded, but the pro-
posed method could track 120 pedestrians simultaneously by the
post processing.
Calibration of eight sensors could be conducted easily by using
the implemented interface. Generally, the calibration of camera
images require complex transformation, thus laser measurement
system has an advantage on calibration. By implementing an au-
tomatic calibration, it may be to conduct calibration more easily.
In addition, proposed method has an advantage on wide measure-
ment arca.
However, there are several examples where some pedestrians could
not be detected or tracked due to the reasons of high pedestri-
ans’ density. First, grouping processes were failed at procedure
of pedestrian recognition. For example, if density was getting
higher about 0.8 [persons/m?], there are failure examples where
1263
Figure
5: A Result of Pedestrian Detection. a: laser scanner, b:
detected pedestrian, c: trajectory.
a leg was grouped neighbouring pedestrian because they are too
close.
Moreover, in case of a pedestrian wearing a long skirt and pedes-
trian who have any bag or cart, pedestrian recognition tended to
fail. In addition, sudden acceleration and sudden deceleration
made tracking failures. For instance, tracking was failed when
other pedestrian suddenly change his or her direction or stop to
avoid other pedestrian coming from the opposite direction. This
example is due to the reasons that a value used in Kalman filter
between estimated a position to measured position surpasses the
state estimation error.
4.2 Movement-pattern Analysis
We analyzed the trajectories obtained by the above procedures
to investigate availability for the station design. Figure 6 shows
the oriented flow-lines and collision distribution at 5:30 pm us-
ing data of 50 second. Where, bright lines show passengers go-
ing to right (gate) from left (gate) and dark lines show the op-
posite flow lines. Dark zone means available area for measure-
ment, gray zone shows an area covered by obstacle such as a wall
and poles. In this image, white points show collisions; we as-
sume a point that some pedestrians come close within 60cm as a
collision. Moreover, we added another assumption restricted by
directional vector angle that made a 180 degree + 45 degree.
We can see the specific stratification between flows of rightward
to leftward. In addition, almost all people going to the platform
from the gate go to the escalator located on the left side. On the
other hands, people going to the gate from central stair approach
toward wall, because few people usc a stair located on the cen-
ter of concourse. Collisions were occurred in an area interfusing
each opposite flows. We think that proposed method is very use-
ful system for movement analysis.
4.3 Accuracy Assessment
Figure 7 shows overlapped results of geometrically corrected video
images, laser points and detected trajectories. We conducted an
accuracy assessment by using these data. At first, we calculated
the measurement ratio at the nine patterns of density and number
of sensors. Each pattern contains 50 images that was aggregated
ob vc
