2.2 Observations
Observations are information that we can obtain from various
sensors. In human tracking under complex situations, people
may be occluded by others or close to each other even if there
are no occlusions. We may be able to identify and track multiple
people that are close to each other by color information, using
color difference of clothes, for example. However, color
information is not suitable for identification of people under
occlusions. On the other hand, range information is robust to
occlusion thanks to difference of distance to some people closer
and farther. Nevertheless, range information cannot distinguish
two people in proximity because the difference in distance is
slight. In addition, it does not bring information about
identification of each person because the observed shape is not
so different from person to person. In consideration of such
conditions, we use both color and range information, which can
redeem their demerit each other. We use stereo video camera to
acquire them simultaneously.
2.3 Forecasting
Forecasting step is a step to predict a current state of a system
from the last state by numerical model. In human tracking, it is
corresponding to the pedestrian behavior model. It predicts
pedestrian's current position based on the last position and
conditions around them. Among many models such as social
force model and cellular automaton, we use discrete choice
model in this paper. This is because discrete choice model
decides the next step of each pedestrian stochastically and can
deal with interactions between pedestrians. Besides, the
alternatives of next step are individual for each pedestrian for
each time. Thus, we consider this forecasting process as non-
linear process, contrary to many literatures on human tracking,
simply assuming random walk or linear process (e.g. Ali and
Dailey (2009)).
24 Filtering
Filtering step is a step to balance the predicted current state by
forecasting step and current observations. In human tracking, it
is corresponding to the problem to evaluate the likelihood of the
predicted state as person. Because background is not stable in
complex situations, we evaluate the similarity of foreground
area, the position and shape of human. Both color and range
information is used for this filtering.
3. GENERAL STATE SPACE MODEL
3.1 General State Space Model
This data assimilation system can be described in a form of
general state space model (Higuchi, 2003). General state space
model is widely used in many fields recently, for it can deal
with non-linear time-series model. As shown in figure 1,
general state space model is composed of state vector x, and
Observation vector z, State vector is a vector of variables of
human position and shape, which cannot be observed directly.
Observation vector is a vector of variables of color and range,
which we can observe directly from sensors. Then we define
observation model p(z/x,), a probability distribution of Z, on the
condition of x, and system model p(xjx,;), a probability
distribution of x, on the condition of x,,. After we obtain zy, =
Ui 7,,..., z,}, series of observations from time 1 to f, X, is
obtained by maximum a posteriori probability (MAP) estimate.
According to Bayes' theorem, the posterior distribution of X, is
as follows:
p(x, | z,) oc plz, | X,)p(x, | Zi)
(D
= p, | X, j| p(x, | X,, )p(x, , | Zp, Yd, ,
In this equation, p(z/x,) is observation model, P(XiX,.1) is system
model and p(x, ;[z;.,.;) is the estimation result at time #-1.
According to the general state space model, human tracking is
processed like this: First, prior probability of. X, is calculated by
applying system model to the probability distribution of Mel:
Then obtained prior probability of x, is combined with
observation z,, and posteriori probability of x, is calculated. In
this framework, we need to define state vector X, and
observation vector z, and model system model P(Xix.1) and
observation model p(z/x;)). We explain about their definition in
section 4.
State Vector x
- cannot be observed directly
(Human position and shape) System Model
p(x, |x.)
Observation Model
rz |x)
Observation Vector z
- can be observed directly
(Color/range information)
Figure 1. General state space model
3.2 Particle Filter
To estimate a state vector, we need to calculate probability
distributions in the equation (1) successively. We use particle
filter for this calculation. Particle filter is a method to
approximate the conditional distribution discretely by number
of particles sampled from that distribution (Gordon et al., 1993).
Calculation of the particle filter is processing as follows and in
figure 2:
I. Approximating the conditional distribution p(X,.i|z;..;) by
number of particles independently sampled with weight
(observation model).
2. Resampling particles with equal weight according to the
weight of each particle sampled at step 1.
3. Moving each particle obtained at step 2 according to the
system model p(x,|x,.,).
4. Weighting particles according to the observation model
(zx). j
5. Estimating x, as the expected value of weighted particles
obtained at step 4.
In th
state
4.1
State
persc
coor
figur
wheı
4.2
sens
both
J) is
whe