Dynamic State Estimation

Using Particle Filter and Adaptive Vector Quantizer

Takeshi NISHIDA, Wataru KOGUSHI, Natsuki TAKAGI and Shuichi KUROGI

Abstract— Particle filter (PF) is a method for discrete approx- iteration to converge, recently, an adaptive VQ (AVQ) algo-
imation of dynamic and non-Gaussian probability distribution rithm called competitive reinitialization learning (CRL) [4],
by using numerous particles, and its procedure can execute at [5] that is able to converge rapidly was proposed. It escapes
high speed and is suitable for on-line applications. However, in
conventional methods, a weighted average value or a maximum from unstable local minima by re-initialization procedure
weighted value of particles is used as a filter output, and and achieves high speed AVQ of time-varying unknown
information on most particles is disregarded. On the other environments.
hand, an adaptive vector quantization (AVQ) algorithm called Therefore, in this research, a method for extracting in-
competitive reinitialization learning (CRL) that can achieve formation on shape of probability density distributions from
high-speed adaptation without depending on initial conditions
has been proposed. Then, in this research, a method for extract- numerous particles by combining PF with CRL is proposed.
ing information on shape of probability density distributions The flow of the processing of the proposal method is as
by combining PF with CRL is proposed. Moreover, a rapid follows. First of all, the probability density distribution of a
adaptation performance and the robustness of the proposed target signal approximates discrete by numerous particles of
method are shown by the simulations. PF. Even if the probability density function is a time-varying
system, PF can track it high-speed and robustly. Second,
I. INTRODUCTION CRL is trained by using the particles as a sampling data
set of the target signal. At this time, all particles are used to
Particle filter (PF) is a method for discrete approximation
train CRL once as an input data set. In this procedure, CRL
of dynamic and non-Gaussian probffability distribution by
achieves a high-speed adaptation by the steepest descent
using numerous particles, and its procedure can execute
method similar to conventional VQ and “jump of weight
at high speed and is suitable for online applications [1].
vectors to an appropriate position” called re-initialization.
Although extraction of deterministic information is necessary
Since the particles of PF also jump greatly by the resampling
to use the estimated result of PF for the following procedure
procedure, re-initialization of CRL is effective when the
(e.g. system control), in conventional methods, a weighted
particles of PF are used as input vectors of VQ.
average value or a maximum weighted value of particles is
used as a filter output, and information on most particles II. PARTICLE F ILTER
is disregarded. For example, the filter outputs an unsuitable
or vibrating value when the probability density distribution PF is a generic name of Bayesian filters based on the
of target signal is a multipeak or an uniform distribution. sequential Bayesian estimation methods [6], [7]. Moreover,
If rough shape of the probability density distribution of the many improved algorithms that can be applied to practical
target signal can be extracted, it is thought that steady use various problems by giving the modification according to
of the estimated result of PF becomes possible. However, in the character of target signal have been proposed [1]. Since
general, when quite a lot of particles are used or the state the role of PF in the proposed method is only discrete
space becomes a higher dimension, it is difficult to measure approximation of the probability density distribution of a
the variance and the density from a distribution of particles. target signal, any algorithm proposed in these researches
On the other hand, many studies on vector quantization can be employed. Therefore, to explain the outline of the
(VQ) has been carried out as a methods for the optimal proposed method concisely, the algorithm of general PF is
quantization of numerous vectors [2], [3]. A vector quantizer shown below.
maps input signal into the finite number of representative
A. Particle filter algorithm
vectors called weight of cells, and each representative vector
is called a code vector or reference vector. Moreover, the We assume that the target input signals are able to repre-
performance of such a quantizer is measured by the distor- sent as a discrete time system, the discrete time step is shown
tion caused by reproducing an input vector with the corre- k = 0, 1, 2, · · · . Moreover, we assume that a state variable
sponding representative vector. Thus, an optimal quantizer xk ∈ Rl of the system is able to be represented by using the
minimizes the average distortion over all the representative state variables in time step k − 1 and k as follows.
vectors. Although conventional VQ algorithms require many
xk ∼ f (xk |xk−1 ) (1)
T. NISHIDA et al. are with Faculty of Engineering, Mechanical and
Control Engineering, Kyushu Institute of Technology, 1-1 Sensui Tobata, This equation is called a system model. Next, the process to
Kitakyushu, Fukuoka, Japan. obtain the observed variable y k ∈ Ro in the system is shown
by the following equation. index of cells. The purpose of the quantizer is to minimize
a distortion given by
y k ∼ h(y k |xk ) (2)
N N
1 (n)
2
(n)
This equation is called an observation model. In PF, the state Dk = xk − wk p(x)dx = Dk ,
N n=1
(n)
Vk n=1
xk in each
 time k is represented by a set of weighted M (7)
M
(m) (m) (m)
particles xk , πk . Here, xk ∈ Rl is a particle
m=1 (n)
(m) where Dk represents the subdistortion for a Voronoi region
in the sate space, and πk ∈ R is a weight of the particle.  
Moreover, m is index of particle, and l and o represent 
Vk
(n)
= 
xk  xk − w k
(n)
≤ xk − w k
(o)
, o = n .
the number of dimension of state variable and observation
variable, respectively. In a general PF algorithm, the set of (8)
particles in each time step is generated as follows [1]:
The CRL is an AVQ algorithm designed to minimize Dk
1pf: Apply a proposal distribution q(·) to latest esti- rapidly based on the equidistortion principle [8].
mated state variables and current observed vector
to obtain next states, B. Competitive Reinitialization Learning
(m) (m) (m)
The outline of CRL algorithm is given below.
x̃k ∼ q(x̃k |xk−1 , y k ). (3)
Given :initial weight vectors are given at random, initial
(n)
Here, the proposal distribution is a probability subdistortion dk = 0.001  1, initial time k = 1,
density function designed to estimate the current and parameters are set by the forgetting rate τ =
states based on the observed variables. Especially, 100, the threshold of distortion dth = 1.4, and the
the case assumed to be f (·) = q(·) is called Monte threshold of distortion entropy Ith = 0.95.
Carlo filter. 1crl : Find the index of winning cell for input vector xk
(m) (m)
2pf: Assign a weight π̃k to the state x̃k according (n)
c = argmin xk − w k .
to system model f (·), likelihood function h(·), and n
proposal distribution q(·) as follows, 2crl : Calculate subdistortion:
⎧
(m) (m) (m) 2
(m) f (x̃k |xk−1 ) · h(y k |x̃k )
⎨ (n) (n)
(m)
∝ (n) ηdk + xk − w k if n = c
π̃k πk−1 (m) (m)
. (4) dk+1 :=
q(x̃k |xk−1 , y k ) ⎩ηd(n) otherwise
k
(m)
3pf: Current estimation variables xk are sampled with where η = exp(−1/(N · τ )), and N represents the
(m) number of cells.
probability π̃k (m = 1, · · · , M ) as follows, (n)
⎧ (1) 3crl : Calculate normalized entropy Ik of dk , and the
mean value d¯k of all dk :
(1) (n)
⎪
⎪ x̃ with prob. πk
⎨ k
(m) . ..  
xk ∼ .. . (5) N (n)
dk
(n)
dk
⎪
⎪ Ik = − ln N ln(N ),
⎩ (M) (M) N (o) (o)
x̃k with prob. πk
n=1 o=1 dk o=1 dk
(9)
This procedure is called a resampling. The weights
(m) N
are normalized as πk := 1/M . When the resam- 1
d¯k =
(n)
dk .
pling is not carried out, the particles are modified N
(m) (m) (m) (m) n=1
as xk := x̃k and πk := π̃k .
4crl : If Ik < Ith and dk > dth d¯k , go to step 4crl(a),
(c)
4pf: Current particles
 M else, go to step 4crl(b).
(m) (m) 4crl(a):Find the sth cell with minimum subdistortion, and
xk , πk ∼ p(xk |y 1:k ) (6)
m=1 update cth and sth cell:
are acquired, return to step 1pf. 
(n) xk if n = s,
wk+1 := (n)
III. C OMPETITIVE R EINITIALIZATION L EARNING wk otherwise,

(n) η d¯k if n = c or n = s,
A. Vector Quantization and Equidistortion Principle dk+1 := (n)
ηdk otherwise,
The problem formulating of VQ is shown as follows; let
xk ∈ Rl be a l-dimensional input vector at time step k with 4crl(b):Update the winning cell:
the underlying probability density function p(xk ), and the  (n) 
(n)
(n) w k + α(k) xk − w k if n = c
vector quantizer consist of a number of cells with weight w k+1 := (k)
(n)
vectors wk ∈ Rl (n = 1, · · · , N ), where n represents the wk otherwise,
 where α(k) = 1 − Ik . where,
5crl : Set k := k + 1 and go to step 1crl.  
(1)  (1)
xj,k xj,k
(1)
For further details of CRL or its parameters, see the literature Xk  ∼ N (μ1,k , Σ), j = 1, · · · , J/2 ,
[4], [5].  
(2) (2)  (2)
Xk  xj,k xj,k ∼ N (μ2,k , Σ), j = J/2 + 1, · · · , J ,
IV. S IMULATION
and J = 2000. The MSE of this data set and the particles
A. Approximation performance of multipeak probability den- was calculated as follows,
sity distribution
J
1 (m)
1) target signal, system model, and observation model: M SEkP F = min xk − xtest
j,k . (18)
J m
In order to generate of an observation signal with multipeak j=1
probability density distribution, a time-varying function is
The simulation result of time evolution is shown in Fig. 5.
consists as follows,
In this figure, since two Gaussian distributions overlapped
 
1 1 1 T −1 completely at each 180 step cycle, the MSE dropped in
yk = exp − μ1,k Σ μ1,k
2π|Σ|1/2 2 2 the same cycle. It was found from the results that the
 nonlinear time-varying target probability density distribution
1 1 T −1
+ exp − μ2,k Σ μ2,k , (10) had been approximated by PF appropriately. Therefore, in
2 2
the following, the effectiveness of the proposal method is
where verified by using this result.
3) vector quantization performance of CRL: Next, time
μi,k = x − xtarget
i,k , (11) evolution of VQ by CRL is shown in Fig. 3 and Fig. 4,
xtarget
i,k = Ai sin(kπ/180), (12) the former is Voronoi diagrams and the latter is Delaunay
diagrams drawn by using weight vectors of CRL. It was
i = 1, 2, xtarget
i,k ∈ R2 , Ai ∈ R2×2 , A1  diag(0.3 0.3), trained by using the particles of PF as input signals, and here,
A2  diag(0.14 0.14), and Σ  diag(0.016 0.016). Namely, the number of the cell of CRL was set as N = 30. Moreover,
the target probability density distribution consists of two to evaluate the approximation accuracy of VQ quantitatively,
Gaussian distribution that move on two dimensional plane. the MSE between the test data sets shown in eqn. (17) and
Fig. 1 shows the time change of this function, and such a the weight vectors of CRL was calculated by the following
cyclically change of the input distribution is rapid compared expressions,
with convergence of conventional VQs. J
1 (n)
Next, a proposal distribution q(·) in eqn. (3), a system M SEkCRL = min wk − xtest
j,k . (19)
J n
model f (·), and an observation model h(·) in eqn. (4) of PF j=1
were set as follows.
The simulation result of time evolution is shown in Fig. 5.
xk = xk−1 + dk + v k (13) Since the number of weight vectors of CRL is considerably
  less than the number of particles of PF though the test data
kπ
sin 180 − sin (k−1)π set for calculation of these values is equal, the value of
dk = 180 (14)
kπ
sin 180 − sin (k−1)π
180
M SEkCRL is larger than the value of M SEkP F . Moreover,
v k ∼ N (0, Σ) (15) it is understood from this results that an appropriate quan-
tization had been achieved by the proposal method because
Moreover, the relation between the following is assumed for a small MES value of 0.02 or less had been maintained.
easiness. Furthermore, it is understood that the feature of the multipeak
probability density distribution was not dismissed and was
f (xk |xk−1 )  q(xk |xk−1 , y k ). (16) maintained from the arrangements of the weight vectors of
VQ and the Voronoi diagrams drawn from them.
Namely, a prior knowledge that target distribution is hardly 4) conditional methods for particle filter: In conventional
changed in a time step was used. methods, although the weighted average value or the maxi-
2) approximation performance of PF: At first, the ap- mum weight value of the particles is used as a filter output,
proximation performance of PF to the target distribution was an undesirable value might be output. Fig. 6 shows results
tested. The number of particles was set as M = 2000. of applying the conventional methods to above-mentioned
In order to evaluate the approximation performance of PF exercise; the former generated the filter output in the vicinity
quantitatively, the test data set according to the distribution of the center of the two peaks, and the latter vibrated the filter
of the observation model was generated as follows, output between the two peaks in short course. It is understood
  from these results that the extraction of an appropriate and
test (1) (2) 
test 
Xk = Xk ∪ Xk = xj,k j = 1, · · · , J , (17) steady estimation value from multipeak probability density
distributions by the conventional methods is difficult.
 N(x,y) N(x,y) N(x,y) N(x,y)
8e4 8e4 8e4 8e4

5e4 5e4 5e4 5e4

0.4 0.4 0.4 0.4

0 0.2 0 0.2 0 0.2 0 0.2
0.4 0.4 0.4 0.4
0.2 0 0.2 0 0.2 0 0.2 0
x
0 -0.2 x 0 -0.2 x 0 -0.2 x 0 -0.2
y -0.2 y -0.2 y -0.2 y -0.2
-0.4 -0.4 -0.4 -0.4
-0.4 -0.4 -0.4 -0.4

(a) k = 0 (b) k = 30 (c) k = 90 (d) k = 270

Fig. 1. Target probability density distribution changing in time. The centers of two Gaussian distributions cyclically move with period of 360 steps.

(a) k = 0 (b) k = 30 (c) k = 90 (d) k = 270

Fig. 2. Results of approximation by PF. Black points represent the particles.

(a) k = 0 (b) k = 30 (c) k = 90 (d) k = 270

Fig. 3. Results of VQ by CRL. The number of cells is N = 30. Points and lines represent the weight vectors and Voronoi bounds, respectively.

(a) k = 0 (b) k = 30 (c) k = 90 (d) k = 270

Fig. 4. Delaunay diagrams drawn by the results of VQ. Only line segments less than the length of 0.07 are drawn.

5) complex of calculation : In a series of procedure of method by adjusting the number of particle M and weight
this proposed method, all particles are used only once to train vector N .
adjustment CRL. This reason depends on the distribution of
the particles of PF not rapidly changing at each time and on B. VQ of dynamic scene
the rapid convergence of CRL. Moreover, since the compu- Finally, the example of online detection of a blue color
tational complexity of the proposed method is O(M 2 N ), a object from a dynamic scene by the proposal method is
general computer can execute a real-time calculation of this shown. The size of the dynamic scene is 640×480 [pixels]
 MSE k
PF Fig.8(c). In these figures, green lines and red points represent
0.03 CRL
MSE k Voronoi boundaries and weight vectors, respectively. More-
over, the Voronoi regions with less than 2000 [pixels] were
MSE

0.02 colored with blue, and the narrower size region was colored
deeper. It was clarified from these results of experiments
0.01 that the proposed method executed appropriate VQ of the
particles which were result of discrete approximation of the
200 400 600 800 1000
uniform distribution and the rough shape of the distribution
0
sampling step was extracted online. Moreover, it was confirmed that this
proposal method adjusted appropriately even if the shape of
Fig. 5. Time evolution of the mean squared error.
the uniform distribution and the distribution of the particles
changed rapidly. In this experiment, the volume of informa-
tion of an input image was about 900 [KB] for each frame,
and its frame rate is 30 [fps]. Here, the experiments were and it was decreased to 800 [B] (i.e. 1/1152) by proposed
conducted under the condition that multiple blue objects method because information was consolidated only in the
were in the scene, and the initial positions of the particles position of the weight vectors.
of PF were set to the center of the image. The image V. C ONCLUSION
processing procedure in these experiments is as follows.
First, after conversion from a rgb input image (Fig.7(a), In this paper, we proposed a method to approximate
Fig.8(a)) to a hsv image, only areas with the hue value of online probability density distribution by particle filter and
120±10 were extracted. Next, the binarization, erosion, and an adaptive vector quantization algorithm CRL. This method
dilatation processing were executed for noise reduction and is effective for extracting information on ”rough shape” from
the preprocessed images are shown in Fig.7(b) and Fig.8(b). time-varying non-Gaussian probability density distribution. It
Afterwords, the pixel values of these preprocessed images becomes possible for a detailed verification of the estimated
were used as a likelihood function in the procedure of PF. In result of particle filter to become possible by using this
this experiment, the target objects moved at random by the information, and to obtain a steady presumption output.
blurring because the experimenter had been taking the scene For example, waiting for the execution of control until the
with the camera by the hand. Therefore, the systems model estimated result converges to a Gaussian distribution if it is
and the observation model were set as follows, understood that the estimated result of PF forms multipeak
or uniform distribution by the proposal method becomes
xk = xk−1 + v k , v k ∼ N (0, Σ), (20) possible.
y k = H(xk ) (21) VI. ACKNOWLEDGMENTS
T
where, xk = (xk [pixel], yk [pixel]) , Σ = diag{20, 20}, This work was partially supported by the Grant-in Aid for
and H(·) represents the binarized value at xk . Moreover, Scientific Research (C) 21500217 of the Japanese Ministry
the number of particles was 1000, and the number of weight of Education, Science, Sports and Culture.
vectors CRL was 100. The results of particle filtering are
shown in Fig. 7(c) and Fig. 8(c). In these figures, green lines R EFERENCES
and red points represent Delaunay line and weight vectors, [1] A. Doucet, N. Freitas, and N. Gordon (eds), sequential Monte Carlo
methods in practice, New York, Springer, 2001.
respectively. Furthermore, Voronoi diagrams corresponding [2] T. Kohonen, Self-organization and associative memory, Springer-
to these Delaunay diagrams are shown in Fig.7(c) and Verlag, 1984.
[3] N. Ueda and R. Nakano, ”a new competitive learning approach based
on an equidistortion principle for designing optimal vector quantizers”,
Neural Networks, vol.7, no.8, 1979, 1211–1227.
[4] T. Nishida , S. Kurogi, and T. Saeki, ”Adaptive vector quantization
using re-initialization method”, Electronics and communications in
Japan, Part 2: Electronics, Vol. 88, No.5, 2001, 60-69.
[5] T. Nishida and S. kurogi, ”An analysis of competitive and reinitializa-
tion learning for adaptive vector quantization”, Proc. of IJCNN2001,
2001, 978–983.
[6] G. Kitagawa, ”Monte Carlo Filter and Smoother for Nong-Gaussian
Nonlinear State Space Models”, Journal of Computational and Graph-
ical Statistics, Vol. 5, 1998, 1-25.
[7] A. Blake and M. Isard, Active Contours, Springer, 1998.
[8] A. Gersho, ”asymptotically optimal block quantization”, IEEE Trans.
on Information Theory, vol. 28, no. 2, 1979, pp. 157–166.
(a) k = 88 (b) k = 90 [9] K. I. Diamanaras and S. Y. Kung, ”Principal component neural
networks”, John Wiley & Sons, Inc, 1996.
Fig. 6. Red circle and blue cross represents the weighted average value
and maximum value of particles, respectively. The former generated in the
vicinity of the center of the two peaks, and the latter vibrated between the
two peaks in short course.
 (a) (b) (c)

(d) (e)

Fig. 7. Results of applying proposal method to object detection in dynamic scene. (a) Input image, (b) preprocessed image, and (c) particle filtering image.
The yellow dots represent the particles. (d) Voronoi diagram. The green lines and the red points represent the Voronoi bounds and the weight vectors. The
Voronoi regions with less than 2000 [pixels] were colored with blue, and the narrower regions were colored deeper. (e) Delaunay diagram. The green lines
and the red points represent the Delaunay lines and the weight vectors.

(a) (b) (c)

(d) (e)

Fig. 8. Results of applying proposal method to blue object detection in dynamic scene. A part of the target object was disappeared by the hand. (a) Input
image, (b) pre-processed image, and (c) particle filtering image. The yellow dots represent the particles. (d) Voronoi diagram. The green lines and the red
points represent the Voronoi bounds and the weight vectors. The Voronoi regions with less than 2000 [pixels] were colored with blue, and the narrower
regions were colored deeper. (e) Delaunay diagram. The green lines and the red points represent the Delaunay lines and the weight vectors. Only the line
segments with the length less than 50 pixels are drawn.