2D Non-Linear State Optimization Using Evolutionary Techniques

Ashish Dhiman
Rajiv Kapoor
Akshay Uppal
Dept. ofECE
Dept. ofECE
Dept. ofECE
Delhi Technological
Delhi Technological
Delhi Technological
University(Formerly Delhi College University(Formerly Delhi College University(Formerly Delhi College
ofEngineering)
ofEngineering)
ofEngineering)
Email:[email protected]
Email:[email protected]
Email:[email protected]

Abstract-

In

modelling

linearity behaviour in

the

non-Gaussian

and

t he systems accurately for

non
t he

estimation of the density function, Particle filter(PF) is

considered more precise compared to other filters like
Kalman filter. Particle filters are also known as Sequential
Monte Carlo method which used the sampling method in
implementing the recursive Bayesian filters.
Particle

filter

has

limitations

like

the

However,

degradation

of

particles and sample impoverishment (SI) which afford an

immense challenge in the non-linear state estimation of
particles. In order to triumph over the limitations (SI), in

this paper, we present novel implementation of 2-D state

estimation
problem

of particles based
using

on

bearing

on

tracking

PF-BBO

(Biogeography
based
optimization) and PF-PSO (Particle swarm optimization.
The efficacy of particle filter is expressed in the form of

root-mean-square-error

values

(RMSE)

and

show

the

improved estimation accuracy of PF-BBO over the PF

PSO.
Keywords- Particle filter; BBO; PSO; RMSE.

I. INTRODUCTION
The algorithm of Particle filter [1] is fit for the state
estimation of nonlinear, non-Gaussian noise system, hence is
an active topic of research and has found an extensive amount
of applications in the fields of object tracking [2], diagnosis of
faults [3], economic forecasting [4] and diverse other areas on
account of its stability, fast and robust nature. Sequential
Monte Carlo methods used the sampling approach to estimate
the state distributions with the set of particles to represent the
posterior density. Each particle has associated weights to
represent the probability of particle sampled from the
distribution. The sample-based estimation of the posterior
density of the vector allows the filter to handle both the
nonlinearity of the system, as well as the non-Gaussian nature
of noise processes making it a general purpose filter. In the
conventional particle filter, sampling of important density
functions different from the posterior probability density
function [5]. If the disparity of particle's weights increases
with respect to time, then the weight tends to collapse, hence
they can be neglected, leading to the leftovers of many
computations on the particles with small weights, therefore the
posterior probability distribution cannot be obtained from the

978-1-4799-8792-4/15/$31.00 2015 IEEE

particle set, this is known as particle degradation. Re

Sampling was introduced in order to defeat the particle
degradation which works by eliminating the trivial weighted
particles hence yield to sample impoverishment [6]. To
overcome these two limitations many optimization techniques
such as Biogeography based optimization (BBO), Particle
Swarm Optimization (PSO), Ant Colony Optimization (ACO),
Genetic Algorithm (GA), etc. were introduced.
Fucheng et al. [7] proposed the parameter transformation
filter (PTF) in 2D and 3D space. This approach was able to
transform the non-linear filtering problem of bearing only to
linear filter problem and showed that PTF can be a better
passive localization filter. Fei et al. [8] used the particle filter
(PF) algorithm on bearing only target tracking problem in case
of underway. They compared the tracking performance
between EKF, UKF, and PF and show PF has better
performance. Bhaumik et al. [9] presented the novel adaptive
grid-based method for risk sensitive. They studied the bearing
on tracking problem with linear and non-linear measurement
and observed that an absolute value of the mean position error
converges as the number of grid points is increased. Walia et
al. [10] also showed bearing tracking problem for non-linear
state estimation of particle filter based on cuckoo search via
levy weight. They also compared the traditional particle with
the particle fiIter with cuckoo optimization. Zhang et al. [11]
proposed the multiple imputations particle filter (MIPF) to
non-linear state estimation problem. They confirmed that
bearing only tracking problem can effectively deal with the
missing data in non-linear problems.
In this paper, we presents the application of particle filter
for bearing only tracking in 2D state estimation of the sample
particles to obtain the accurate positioning of the system. PF
BBO and PF-PSO are compared and incorporating the
optimization techniques offers more diversity in the particles
and prominent in target tracking. The main contributions of
our work are as follows:
E stimate the Non-linear state of sample particles based on
bearing on tracking the problem.
Performance analysis of PF-BBO and PF-PSO in terms of
RMSE.
Our paper is organized as follows: Section II presents the
description about the particle filter, evolutionary techniques.
Section III describes our approach related to bearing only
tracking problem and our findings from our experimental

2192

study. Section IV explains the results and Finally, Section V

presents the conclusion and the scope for future work.

P()
ZI-I

II. PARTICLE FILTER

f P()p(

XI-I

XI-I
)dxl_I
ZI-I

Our system is composed of two primary units, A. Particle

The Block diagram of
proposed methodology is explained in the block diagram (Fig.
1).

A set of random samples is taken as

A. Particle Filter

given by

Filter; B. Evolutionary techniques.

Particle filter was first introduced as a bootstrap filter by N.J.

Gordon et al. [12] and later on improved as survival of the
fittest, condensation algorithm, and interacting particle
approximation. It estimates the posterior probability of state
based upon probability distribution using a number of
weighted particles [13]. To represent the posterior probability
density function particle filter exploit the technique of point
mass representation and recursively sample the Bayesian
equations [14].

XI

FI-l (XI-I'WI-1)

(1)

,- ------- I
I
.
I Observation I
I
I
odel
, ---- ----,

( nt:l Sae
:

I
,

of Particles

'I

I
I

,------_/

r -I
(

Sae

odel

,-----

"

P O

- -'

b li Y

Estimation

,--

----

,
I
I
I

'

-----

Output

:
:
I

State

,-----_/

Optimization

of

states

P(!!"')
Z/

(5)

B. Evolutionary Techniques

We used two evolutionary techniques to remove the

sample impoverishment problem that generates due to the
degeneracy effect.

the position of a particle be represented by

-+

E Rn

f-

---7

---7

--+

--+

--+

vWv+prp(p-x)+grg(g-x)

Fig. I Proposed Methodology of the 2D non-linear state estimation of Particle

filter based on bearing only tracking a problem.

It is assumed that at discrete times, measurements are

available whose observation model is represented by the
observation equation as
(2)

where WI and VI represent zero mean, white-noise

sequence. The prior density of samples represented as
(3)
, the target is to successfully estimate the posterior from the
measurement sequences

Prior probability density function (PDF) is expressed as;

set

,------_/

z/=zk:k=I,2,3 .. .l .

X/-I (j) : j = 1,2,3 ...N.

the

, and let its velocity be v . The position as well as the

velocity are initially randomized and are subsequently updated
according to the formulae mentioned in equation (6). The
particle's velocity is updated as mentioned by Shi and
Eberhart:

,- -------- -

ZI HI(XI, VI)

estimate

1,2,3 ...N. which are distributed as

to

P(_z_'_)
P(Y')P()
:::;
x
z
,

_ x_'....:(",,-J..:..")
'
,.:.
.!.
,
.:::
:.!.
_
_
=
=
_
P( )
,Wj
z,
P()
P(_z_' )
k'
x,(k)
Z'-I

-->

,- -.
ReI I
I
BBO/PSO
j+-II Sampling
, ---------.

XI-l (j) : j

is

f : Rn -t R. Let

--------"

-,
I
I

aim

1. Particle Swarm Optimization(PSO)

In Particle Swarm Optimization (PSO)[15] the position of the
particles is randomized in the search area, to begin with,
moving in randomly defined directions. The subsequent
direction of motion of the particle is then set by the optimum
previous positions of the particle and its neighbours; the
search is then continued in the nearby particles in order to find
better positions keeping some fitness measure as a comparison

,- ------------ ,
Prior Next State
I
I
Estimation
I
I
'---

Our

(4)

(6)

The recurrence in the particle's velocity is controlled by the

....
....
inertia weight Rand p and g are represented as the particle's
and best swarm locations respectively via which the
communication between the particles takes place implicitly.

rp, rg

are defined as the stochastic variable belong to the

U (0, 1) and the values p, r E R defined by the user are

used to weight these particles.
Algorithm:

Step 1: For all particles i: l...s, initialize the particle's position

with a random vector X in the predefined boundaries (I, u )
positions.
Step 2: Starting from the initial position let the best know
position of the swarm be updated when
f (particle's best position) < f (swarm's best position).
Step 3: Initialize the velocity by vector such that

V E ( - u -/ 1 , u

-II )

2015 International Conference on Advances in Computing, Communications and Informatics (ICACCl)

(7)

2193

Step 4: Until a termination condition is met

For each particle with position x and velocity v do:
Update the particle's velocity using Equation (5)
Apply the boundaries on velocity.
Update the particle's position using Equation(6)
Re-Apply the boundaries on the updated values in case
they have been extended.
If f (updated position) < f (particle's best position) then
updates the particle's best known position.
Now if f( particle's best position ) < f (swarm's best
position) if this holds true update the value of the best
position of the swarm
And similarly for the entire swarm's best-known position is
known.

Step 3: For all the elite candidates Sj, till the termination
condition is met, set the emigration probability /!s directly
proportional to the fitness of Sj with /!s E [0,1] and for each

candidate immigration probability

2.

Biogeography Based Optimization (BBO)

BBO[16] is motivated from the immigration and emigration of

species between habitats. In case of BBO, immigration and
emigration rates are determined by choosing the fitness
function for each generation which makes it d istinct from
other methods. The emigration rate is linear to the fitness of
candidate solution whereas the immigration rate is inversely
proportional to the fitness of candidate solution[17]. In BBO,
habitat suitability index (HSI) is used to define the fitness
function for candidate solution, whereas candidate
optimization problem solution is denoted by a biogeography
habitat and population of candidate solutions is denoted by a
set of biogeography habitats. Migration and mutation are the
two steps in BBO optimization where migration is defined as
an operator which enhanced the candidate solution and
mutation increase the variance among the population. The two
operations are described as follows

Migration: under migration sharing of information is

controlled by the emigration and immigration rates, given
by

"tv

<D

K(I--)

(8)

where n is defined as the number of total number of candidate

solution and denotes the species count in the candidate
solution. K is the number of species of K th individual.
Mutation: After migration, its solution undergoes the
mutation operation controlled by the mutation rate

ms

mmax (

1- p

__

Pmax

(9)

where mmax is an user-defined parameter, and Pmax = arg

max Pj. The population consists of NP=n and the parameter i
vary from 1 . . . NP.
Algorithm:

Step 1:

Initialize the population of candidate solutions {Sj}

(500 in our case). Calculate weights and sort them in order of
most fit to least fit.
Step 2: Keep the fit particles determined by the elitism
parameter. (2 in our case)

2194

A. 1-l1s

new location Zj
Zj Si
Step 4: For an independent incrementing index decide if the
migration is to be done to the new location using the
immigration probability if the immigration is to be done then
select the emigrating particle using /!s
Step 5: Pick a habitat to obtain the feature. Mutate the worst
half of the solution according to the weight and replace the
habitats with their mutated versions.
Step 6: Next particle: i=i+1
SiZj
III. EXPERIMENTAL METHODS
Bearing only tracking problem is an important field of
research in the area of submarine tracking (radar and sonar
environment), aircraft surveillance, autonomous robots and in
mobile system. The objective of this problem is to find the
position and target of the moving object. However, due to the
non-linearity it i s difficult to construct the Bayesian filter for
this relatively simple problem.
In our method, the state space model
within the x-y plane defined as [10,12]

Dynamic State Model:

(10)
where state at time t is

Xk

Wk (wx, WY)t

noise

[x ,VX' Y ,Vy ]T t

and system

is zero mean Gaussian white noise

process. Vx and Vy represents the velocity in the x - Y

direction respectively. The online measurement by an observer
at the origin is represented by observation eq.
-I Yk
zk=tan (-) +Vk
Xk

where

Wk

(11)

and Vk represent the zero mean, white Gaussian

noise. The initial state of object XI is known.

XI

=l . J
0.7

-0.055

The prior for initial state is also predefined and assumed to be

Gaussian distribution with mean Xl and covariance (j is
defined as

a{f

0.3

-0.005
0

1 l O4 j
;1

-0.05

2015 International Conference on Advances in Computing, Communications and Informatics (ICACCl)

The likelihood of each particle is computed by the

Bhattacharya distance which leaves with the measure of
similarity to the target for every particle in the site. The
distance is used to update the weights of the particle
K

-d2

(n)

.J2Hu e2u2

.J2Hu e

-(l-p[pg(n),q)

2u2

(12)

(n)

(n)

(13)

()
LK i

IV. RESULTS AND DISCUSSION

PF-PSO Optimization Results:

.35

-O ..!5
5

"00

.35

0'

rfi

.0 . 45

.56

06

0 6

O, 1 4
--;;
8
:;;::: 0.12;--;;'
;--.,-: 0. 0
:!co.; a.:---:o.(I;
2----;
o!;;o
-O sg.
0';-.1 --;

(a)

.'
.

...... +

...

0.5

.55

.1---="012
0:'7
::
.(lSS02-70=-=.0 -700::00--::0-:00:---,

(b)

Iterations at Time Interval 'T=20'

1.62

1.64

1.66

1.68

II O oDdJ
0

1.n

1 . 62

164

-UiS

.1.7

-1T

178

'

:t*

.'

+.

174

-176

US

., ?

., ,,,

Q:]

17

n"

"'

"'

(a)

n,,'\

n",

"'

78

. 1. ,

025

"

035

(b)

Iterations at Time Interval 'T=40'

2.S

285
2 . 9

2.95

2.9

"

.,

'

3.05

05

3.1

3.1

-3.15

.36

2 . 8
285

., ,,

O.B

(a)

0.9

409

.411

315

o
1.1

"0.7

O.S

(b)

0.9

0.'

1,4

1.6

(a)

I.S

4.(13

4 .07

4.(18

-4.09

-4.1

J . 1

4.12

-4.13
22

2 .'

14

1.6

'6

(b)

Iterations at Time Interval 'T=80'

Fig. 2 (a) Particles without optimization as estimated by the Particle
filter (b) After PSO Optimization at different iterations

The proposed particle filter algorithm representing the bearing

only tracking problem was implemented on M atlab (2012b)
and tested on a core i5, 2.67 GHz PC with 4 G B Ram
memory. In our proposed method, we choose the particles size
N=500 and t=80 for the optimization results. In case of BBO
optimization we choose the elitism factor is=4, bounding
limits for lower and upper migration rates are 0 and I
respectively and the mutating factor=O. l. Fig (4) represents
the RMSE curves of particle filter trajectory in the x-y
direction for t=80 iterations where, red circle curve represents
the original state or true state of the sample particle and the
green squares curve represents the estimated states of the
particle filter.

0
o

'08

1.2

i=1

.07

-4.1

-4.05

413

.'(1;

"12

The weights are normalized by

'05

0.45

A.'

0.55

PF-BBO Optimization Results:











 DE
,WHUDWLRQVDW7LPH,QWHUYDO7 









 DE
,WHUDWLRQVDW7LPH,QWHUYDO7 









 DE

,WHUDWLRQVDW7LPH,QWHUYDO7 









 DE
,WHUDWLRQVDW7LPH,QWHUYDO7 


Fig. 3 (a) Particles without optimization as estimated by the Particle

filter (b) After BBO Optimization at different iterations.

Iterations at Time Interval 'T=60'

2015 International Conference on Advances in Computing, Communications and Informatics (ICACCl)

2195

techniques, the concept of mutation and migration helps in

avoiding the impoverishment of the particles. Our structure of
the algorithm is robust and simple, gives best results with
BBO optimization with least RMSE error. The Particle Filter
is a trending topic and the proposed algorithm find application
in the state estimation, surface tracking, 3-D human motion
tracking, distance estimation etc. Our future work is
integrating our algorithm with the real-time systems.

(VWLPDWHG6WDWH
7UXH6WDWH

REFERENCES

D
(VWLPDWHG6WDWH
7UXH6WDWH

E

Fig. 4 (a) Representation of RMSE curve of particle filter

trajectory for PF-PSO and (b) PF-BBO Optimization for T=80
Iterations.

From the Fig (4), we infer that with the every iteration the
estimated states are becoming closer to the original states.
Particles are more localized in case of BBO optimization
compared to the PSO optimization and RMSE curve
accurately follows the pattern with fewer error values. On our
simulations, it was found that RMSE values of the particle
filters which underwent optimization using PF-PSO and PF
BBO are 0.0236 and 0.0110 respectively.

V. CONCLUSION AND FUTURE WORK

This paper revolves around the 2D non-linear state estimation
based on bearing on tracking the problem of Particle filter
using the PSO and BBO optimization. The algorithm of
particle filter clearly indicates that it successfully represents
the particles and efficient in tracking and estimating the states.
The PSO clusters the particles into the areas with high
likelihood hence minimizing the degradation in particles. The
BBO is a cl ever way to replace the existing re-sampling

2196

[I] J. H. Kotecha and P. M. Djuric, "Gaussian Particle Filtering", IEEE

Transaction on Signal Processing, voL 51, no. 10, Oct. 2003.
[2] Xinting Pan, Xiaobo Chen, Aidong Men, "Occlusion Handling Based on
Particle Filter in Surveillance System", Second International Conference
on Computer Modeling and Simulation, vol.J, pp. 179 - 183, 2010.
[3] Predrag Tadic, Zeljko Durovic, "Particle filtering for sensor fault
diagnosis and identification in nonlinear plants," Journal of Process
Control, voL 24, no. 4, pp. 401-409, 2014 .
[4] Hedibert F. Lopes and Ruey S. Tsay, "Particle filters and Bayesian
inference in financial econometrics", Journal of Forecasting, voL 30, pp.
168-209, Jan 2011.
[5] M. Li; B. Pang; Y. He; F. Nian, "Particle Filter Improved by Genetic
Algorithm and Particle Swarm Optimization Algorithm", Journal of
Software, pp-666-672, 2013.
[6] Katsuji Uosaki, Yuuya Kimura, Toshiharu Hatanaka, "Nonlinear State
Estimation by Evolution Strategies Based Particle Filters", Proceedings.
Congress. Evolutionary Computation, voL I, pp. 2102 - 2109, 2003.
[7] Guo Fucheng, "A new method of bearing-only target localization based
on parameter transformation" International conference on Infonnation &
Automation, pp-1406-1410, June,2008.
[8] Zhang Fei; Zhou Xing-peng; Chen Xiao-Hui, Liu Rui-Ian, "Particle filter
for underwater Bearing-Only passive Target Tracking", IEEE Pacific Asia
workshop on Computational Intelligence & Industrial Application
vol:2,pp 847-851, Dec,2008.
[9] Bhaumik, S; Srinivasan, M; Sadhu,S; Ghoshal, T.K; "Adaptive grid risk
sensitive filter for non-linear problem" Signal Processing, IET,
vol:5,issue:2,pp-235-241,April 2011.
[10] Walia,G.S;Kapoor,R; "Particle filter based on Cuckoo search for nonlinear
state
estimation",3,d
International
Advance
computing
conference(IACC),pp-918-924,Feb 2013.
[II] Zhang,X.P; Khwaja, A.S; Luo, J.A; Housfater, A.S.; Anpalagam, A; "
Convergence Analysis of Multiple Imperfections particle for dealing with
Missing data in Non-linear problem", IEEE, International Symposium on
Circuits & Systems(lSCAS), pp-2567-2570, June 2014.
[12]N. Gordon, D. Salmond, A.F.M. SMITH, "Novel approach to
nonlinear/non-Gaussian Bayesian state estimation", IEEE Proceedings
Radar and Signal Processing, voL 140, pp. 107-113, April 1993.
[13]M. S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, "A Tutorial on
Particle Filters for Online NonlinearlNon-Gaussian Bayesian Tracking",
IEEE Transactions on Signal Processing, voL 50, no. 2, Feb 2002.
[14]Y. Liu, B. Wang, W. He, J. Zhao and Z. Ding, "Fundamental principles
and applications of particle filters", Proceeding 6th World Congress.
Intelligent Control and Automation, voL 2 pp.5327 -5331, 2006.
[15]Yang, X. "Particle Swann Optimization Particle filtering for Dual
Estimation", Signal Processing, IET,voL6 pp. 114-121,May 2012.
[16] D. Simon, "Biogeography-based optimization", IEEE Transaction
Evolutionary Computation voL 12 no.6, pp.702-713, Dec. 2008.
[17]H. Ma; M. Fei; Z. Ding; J. Jin; "Biogeography-Based Optimization with
Ensemble of Migration Models for Global Numerical Optimization"
IEEE World Congress on Computational Intelligence (WCC!), Jun, 2012.

2015 International Conference on Advances in Computing, Communications and Informatics (ICACCl)