(XXIX-B3, 2012
1a, P., 2011. Pedestrian
' the art. IEEE Trans-
1e Intelligence, 34 (4).
6. Extremely random-
3-42.
ine boosting and vi-
e on Computer Vision
—267.
Auhle, D., Sester, M.,
ajectories: Improving
"organizing camera net-
'togrammetry and Re-
gnition using random-
Analysis and Machine
. Tracking colour ob-
nage and vision com-
ring and boosting. In:
001.
d Fua, P, 2010. Fast
s. IEEE Transactions
zence. 32(3), pp. 448-
, M. and Bischof, H.,
EEE ICCV Workshop
Gool, L., 2010. Au-
trians from a moving
immetry and Remote
rees and forests on a
tive background mix-
oceedings IEEE Con-
Recognition, pp. 246-
MULTIPLE HUMAN TRACKING IN COMPLEX SITUATION
BY DATA ASSIMILATION WITH PEDESTRIAN BEHAVIOR MODEL
W. Nakanishi ^ *, T. Fuse?
? Dept. of Civil Engineering, University of Tokyo, 7-3-1 Hongo, Bunkyo, Tokyo 113-8656 Japan
- (nakanishi@trip, fuse@civil).t.u-tokyo.ac.jp
KEY WORDS: Observations, Simulation, Integration, Image, Sequences, Tracking, System, Modelling
ABSTRACT:
A new method of multiple human tracking is proposed. The key concept is that to assume a tracking process as a data assimilation
process. Despite the importance of understanding pedestrian behavior in public space with regard to achieving more sophisticated
space design and flow control, automatic human tracking in complex situation is still challenging when people move close to each
other or are occluded by others. For this difficulty, we stochastically combine existing tracking method by image processing with
simulation models of walking behavior. We describe a system in a form of general state space model and define the components of
the model according to the review on related works. Then we apply the proposed method to the data acquired at the ticket gate of the
railway station. We show the high performance of the method, as well as compare the result with other model to present the
advantage of integrating the behavior model to the tracking method. We also show the method’s ability to acquire passenger flow
information such as ticket gate choice and OD data automatically from the tracking result.
1. INTRODUCTION
Recently in-depth understanding of pedestrian behavior in
public space is becoming significant with regard to achieving
more sophisticated space design and flow control. The difficulty
in space design in big stations, for example, is that we should
consider the congested level inside a station entirely, which
changes every second, and passengers' microscopic route
choices at the same time. Therefore, understanding passenger
flow in detail is necessary to accomplish good facilities
planning. The same is true in shopping malls and pedestrian
crossings. In order to understand such human behavior, the
main problem is to comprehend individual's behavior in
complex situation that people move interdependently.
Observation data from diverse sensors, which are informative to
understand human behavior, are accumulated these days thanks
to the development of sensing technology. As such data
increase, a strong need arises to acquire behavior information
automatically. However, automatic human tracking is still
challenging under the situations that people move close to each
other or are occluded by others. Human tracking is usually run
by color information obtained from video camera, for we can
get information of the entire field observed. As color
information is not robust to occlusions, range information
obtained from laser scanner or stereo video camera is also used
for human tracking recently (e.g. Munoz-Salinas, 2008).
Meanwhile some simulation models of walking behavior have
made progress recently (Bierlaire and Robin, 2009). In such
models, pedestrian's choice of next step is explained by not only
each individual's current position and velocity but also the
interdependency as the response to the presence of other
pedestrians. In order to develop simulation models, real data of
pedestrian behavior is necessary for calibration of parameters
and evaluation of reproducibility. In addition, the possibility to
improve behavior models by feeding the tracking result back to
them becomes greater if automatic tracking is achieved.
In this paper, we propose a new method of multiple human
tracking under the complex situations. The key concept is that
to assume a tracking process as a data assimilation process,
widely used in many fields of geosciences (e.g. Daley (1991)
and Wunsch (1996). As human behavior is uncertain and
human is non-rigid object, stochastic and non-linear tracking
process is suitable. Also as huge volumes of data are processed
for tracking, sequential process is suitable. An on-line data
assimilation system matches this two needs. It consists of
observations, forecasting and filtering. In human tracking,
observations correspond to observation data from sensors,
forecasting to pedestrian behavior model and filtering to
existing tracking method by image processing.
The rest of the paper is organized as follows. In section 2, we
describe how we apply data assimilation to human tracking. In
section 3, we present the calculation method of this assimilation.
In section 4, we define some components of the model
according to the review on related works, both human tracking
method by image processing and simulation model of pedestrian
behavior. Finally, we apply the proposed method to real data in
section 5 and conclude the paper in section 6.
2. DATA ASSIMILATION
2.4 Human Tracking as Data Assimilation
We assume a human tracking process as an on-line data
assimilation process as mentioned above. lt consists of
observations, forecasting and filtering step. In data assimilation,
after the current state is predicted by forecasting step,
observations of the current and past state are combined with
them by filtering step. In human tracking, the process is
repeated like this: In each frame, positions and shapes of people
being tracked are estimated by pedestrian behavior model
(forecasting step). Then estimated shapes and positions are
optimized referring to the new observation data (filtering step).
Each step in this paper in detail is described below.
