The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
2.3.3 Expression of Time Series Transition 
Since crowds flow pattern changes at every moment, it is 
necessary to adjust the time interval how previous frames 
should be dealt with (e.g. a minute, an hour or a day). In order 
to adjust time interval flexibly, a leaky bucket algorithm is 
applied. For example, if we need to assess the trajectories in the 
past 10 seconds, an additional value of one KDE process is 
defined as 10, and the added value is subtracted 1 at every 
seconds. Therefore, the probability density will be zero after 10 
seconds, so that we can adjust the time interval by adjusting the 
additional value and subtracting value. Accordingly, if some 
pedestrians pass at same position and direction, the probability 
density is higher than other places. 
3. EXPERIMENTAL RESULTS 
3.1 Overview of Experiment 
In order to test our proposed system in a real environment, we 
conducted an experiment at a concourse of railway station in 
Japan. The dimension of the measurement area is about 60m x 
30m (Figure 3). In this experiment, we exploited 8 laser 
scanners to cover entire target area. In rush hours, over 200 
pedestrians occupy the concourse as shown Figure 4. 
J Dimension : 60 x 30m j 
Platform 
Platform platform 
1 
1 
LJ II 1 
■ 
L 
■ 
J 
-|„A 
Rest Room 
Ticket Gate 
j Store 
Measurement Area 
Platform Platform 
■ Laser Scanner 
Figure 3. Layout of 8 laser range scanners 
Figure 4. An image of experimental site 
3.2 Results of Tracking Pedestrians 
Figure 5 shows a result of tracked 69 persons. The line symbols 
stand for the obtained trajectories. The validation period is 
selected from 7:00 AM to 7:05 AM including congested and not 
congested situation. We assess the tracking accuracy as the ratio 
of the number of trajectories tracked from entrance to exit 
completely. In relatively congested period from 7:00 to 7:02, 
the tracking accuracy is 81.2%, and in unoccupied period from 
7:02 to 7:05, the accuracy is 88.9%. Conventional researches 
using CCD camera could not achieve these accuracy and 
successive tracking in wide space and high congested situation. 
Figure 5. Tracking result (non-congested period) 
Figure 6. Tracking result (congested period) 
3.3 Results of Crowds Flow Detection 
Figure 7 shows a result of visualization of calculated crowds 
flow as a probability density distribution. The sampling data 
period is selected from 7:00 AM to 7:10 AM. In this figure, z 
axis corresponds to the value of probability density or the 
likelihood. In other words, the probability density is in 
proportion to the traffic density. The color of crowds flow 
stands for the direction that pedestrians walked most frequently. 
Figure 8 stands for a result that shows prominently congested 
areas. The highlighted areas show where the calculated 
congestion index mentioned in section 2.3.2 is over a certain 
threshold. These all processes are achieved on the real-time 
basis. 
955