TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.6, Jun
e
201
4, pp. 4731 ~ 4
7
3
9
DOI: 10.115
9
1
/telkomni
ka.
v
12i6.549
3
4731
Re
cei
v
ed
De
cem
ber 2
9
, 2013; Re
vi
sed
March 7, 201
4; Acce
pted
March 22, 20
14
Severe-
Dynamic Tracking Problems Based on Lower
Particles Resampling
Xungao Zho
ng
1
, Xiafu Peng*
2
, Xun
y
u
Zhong
3
Dep
a
rtment of Automatio
n
, Xi
amen U
n
ivers
i
ty
Siming R
o
a
d
, Xi
ame
n
36
100
5, Chin
a, T
e
l.
:
+
86-13
7-20
88-
542
9, F
a
x: +
8
6
-
592-
258-
00
05
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: zhong
xu
ng
a
o
@1
63.com
1
, p
e
ngx
i
a
fu
@
126
.co
m
2
,
zhon
g
x
u
n
y
u@
xmu.e
du.cn
3
A
b
st
r
a
ct
F
o
r a target as it
w
i
th large-
dyna
mic-cha
n
g
e
w
h
ich is still
challe
ng
ing fo
r existing
met
hods to
perfor
m
rob
u
st tracking; the sampli
ng-
base
d
Bayesia
n
f
ilter
ing often suffer
from co
mputat
ion
a
l co
mp
lexit
y
associ
ated w
i
th large n
u
m
b
e
r
of particle de
ma
nd
ed an
d w
e
ighi
ng
multi
p
le hy
pothes
e
s
. Specifically,
this
w
o
rk propos
es
a ne
ura
l
a
u
xil
i
a
ry Bayes
i
a
n
fi
lterin
g sche
m
e
base
d
o
n
Mo
n
t
e Carl
o res
a
mplin
g tec
hni
qu
e
s
,
w
h
ich to ad
dre
sses the co
mp
utation
a
l
i
n
ten
s
ity that is intri
n
sic to al
l
p
a
rticle filter, i
n
clu
d
in
g those
hav
e
bee
n
mo
difie
d
to overco
me th
e de
ge
neracy
of particl
es. T
r
ackin
g
q
ual
ity for sever
e
-dy
n
a
m
ic
exp
e
ri
me
n
t
s
de
mo
nstrate that the neura
l
vi
a comp
e
n
sate
the Bayesi
an fi
lterin
g error, w
i
th high acc
u
ra
cy and inte
nsiv
e
tracking
perfor
m
a
n
ce
only
r
equ
ire l
o
w
e
r
particl
es co
mpare w
i
th s
e
q
uenti
a
l i
m
port
ance r
e
sa
mpli
ng
Bayesi
an filteri
ng, meanw
hi
le,
our method
al
so w
i
th strong robustn
ess for low
number of particl
es.
Ke
y
w
ords
:
Bayesi
an filt
eri
ng, ne
ura
l
n
e
tw
ork, Monte
Carlo
resa
mpli
ng, partic
l
es c
onstrai
ned, r
o
bus
t
tracking
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Rob
u
st tracki
ng is an
acti
ve re
sea
r
ch
topi
c in
com
puter vi
sion,
and it h
a
s
receive
d
extensive ap
plicatio
ns in
cl
uding in intell
igent su
rve
ill
ance, robot visual
servoi
ng
, etc. In terms of
tracking
algo
rithms, large n
u
mbe
r
of ap
p
r
oa
che
s
h
a
ve
been
pro
p
o
s
ed in the
pa
st decade
s, su
ch
as the
state-estimation
-b
a
s
ed Kalm
an f
iltering (K
F
)
[1], and sam
p
ling-b
a
sed B
a
yesia
n
filtering,
namely pa
rticle filtering (P
F) [2-4].
For the
sce
nario
s
with
severe-dyna
m
ic motio
n
, the KF wa
s wi
dely re
p
l
ace
d
by
sampli
ng
-ba
s
ed tracking
m
e
thod
su
ch
a
s
the
PF,
wh
i
c
h i
s
a m
u
ltiple-hypoth
e
si
s
solutio
n
a
b
le
to
estimate a
r
bi
trary di
stribu
tions throug
h eval
uation
of rand
om
sampl
e
s i
n
a state
sp
ace,
therefo
r
e, sa
mpling is a
vital step for PF
, while
the traditional se
quenti
a
l Monte Carl
o
sampli
ng
(SM
C
S) meth
od
s face the deg
eneracy of p
a
rticle
s p
r
obl
em, whi
c
h so
metimes i
s
very
severe leads to only a few particles are us
ed to represent
the co
rresponding probability
distrib
u
tion.
Therefore,
th
e exten
s
ion
s
to PF, rece
ntly, mainly focu
s
on th
e
sampli
ng
met
hod
s to
overcome th
e
degen
eracy
of particl
es
problem
s an
d to
improve the
diversity of p
a
rticle
sa
mpl
e
s,
including
aux
iliary variable PF (AVPF)
[5], fission
bootstrap PF
(FBPF)[6].
In order to further
improve the
sampli
ng efficien
cy, the famous M
a
rkov
chain Mo
nte
Carlo
(MCM
C) meth
od h
a
ve
obtaine
d the
con
s
ide
r
a
b
l
e
develop
me
nt [7,
8], and the ada
ptive MCM
C
[9,
10] have sh
own
more
supe
rio
r
ity in in
crea
sing the
mixin
g
an
d a
c
cept
ance
rate
s, in
[11] the
auth
o
rs p
r
op
osed
a
intensively a
daptive M
C
MC
(IA-MCM
C
)
sam
p
le
r to improve th
e sa
mplin
g
ef
fi
ci
en
cy
,
w
h
ich
combi
n
e
s
a
d
ensity-g
r
id
-ba
s
ed
p
r
edi
ctive mo
del
with
the sto
c
h
a
sti
c
a
p
p
r
oximati
on M
onte
Ca
rlo
(SAMC) algo
rithm
[12].
In this pa
per, a method
with ra
dial b
a
si
s fun
c
tion
neural net
work
(RBFNN) auxiliary
particl
e filteri
ng algo
rithm i
s
propo
se
s, whi
c
h wa
s m
o
tivated by re
duces
the co
mputational cost
and imp
r
ove
s
the robu
st
ness for vi
sual tra
cki
ng,
and we ap
ply this neu
ral-pa
rticl
e
filtering
schema
in l
a
rge a
b
ru
pt mo
tion tra
c
king
probl
em
s.
We first having
a de
scriptio
n
to the Baye
si
an
state e
s
timati
on fra
m
e
w
ork for visual tra
cki
ng
ta
sk, then
we int
r
od
u
c
e
sam
p
ling
-
based Baye
si
an
filtering, whi
c
h the sampl
e
r able to e
s
timate ar
bitrary distribution
s
thro
ugh ev
aluation of ra
ndom
sampl
e
s in
a
state sp
ace. But for the unco
n
stra
ine
d
abru
p
t motio
n
trackin
g
problem, a cert
ain
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4731 – 4
739
4732
numbe
r of
sa
mples are
still req
u
ire
d
to
captu
r
e th
e a
b
rupt
motion
due to th
e b
r
oadn
ess of t
h
e
whol
e state
space, whi
c
h
stack i
n
favour of comp
utational
co
st.
Therefo
r
e we
fu
rther pro
p
o
s
e
d
a
neural-Baye
s
i
an resamplin
g filtering
(NBRF)
ba
sed
on
lo
we
r sa
mpling hypot
hesi
s
,
utili
ze
s
the
Monte Carlo
sampli
ng al
g
o
rithm
whi
c
h
with con
s
trai
nt cou
n
t of particle
s
to
si
milar
comp
utation
of the compl
e
x integratio
n
conju
gate in
Bayesian
filtering, a
nd th
e neu
ral net
work (NN) vi
a by
comp
en
sate
the Baye
sian
filtering
e
r
ror
ca
n b
e
o
v
erco
me th
e
high
comp
u
t
ational b
u
rd
en
cau
s
e
d
by large numb
e
r of
particle
s
p
r
o
b
lem,
mean
while the ra
nd
om abrupt motion ca
use the
model u
n
fitab
l
e whi
c
h
will
dire
ctly decre
ase th
e
tra
cki
ng pe
rform
a
n
c
e al
so b
e
im
proved
by NN.
Many
comp
are expe
rime
nts d
e
mon
s
trated that th
e
NBRF
can
be
effective an
d
pre
c
ise tracki
ng
largely un
co
n
s
train
ed ab
ru
pt motion with robu
st
even
using le
ss nu
mber of samp
ling parti
cle
s
.
2. Resampli
ng-Base
d Ba
y
esian
Filtering for Tra
cki
ng Problem
For a ta
rget
as it with ab
rupt motion, it
is a challen
g
i
ng problem t
o
achi
eve th
e rob
u
st
tracking,
since the seve
re
dynam
ic target wa
s difficult repr
esent
ed by a linea
r app
roximat
ed-
model. Herei
n
, we intro
d
u
ce the
re
sa
mpling
-
ba
sed
Bayesian fil
t
ering meth
o
d
s which is
to
enlarge the sampling va
ria
n
ce to cover t
he po
ssi
ble
motion un
ce
rtainty.
Sampling te
chniqu
e su
ch
as Mo
nte Ca
rlo s
equ
ential
importa
nce resam
p
ling
(S
IR) [13]
that is recursi
v
ely estimate
s po
st
e
r
ior
probability de
nsity function (P
DF)
}
/
{
)
:
1
(
(t)
t
p
Z
X
by sel
e
cti
ng
and
simulatin
g
a stati
s
ticall
y relevant su
bset of
p
o
ssi
b
le sy
stem st
ates, form
ally, the goal of
SIR
is to obtai
n
a set of
N di
screte
sam
p
l
e
N
1
i
i
(t)
}
~
{
X
and th
eir
correspon
ding
weig
hts
N
1
i
i
(t)
}
~
{
W
to
approximate
poste
rio
r
PDF, then
the i
n
tegratio
n in
Bayesian f
ilte
r
ing can
be approximated
by
point mass
es, as
follows
[12]:
)
-
(
~
}
/
{
)
(
(t)
1
i
(t)
)
:
1
(
(t)
i
t
N
i
t
p
X
X
W
Z
X
(1)
The pa
rticle
s-wei
ghts
pair
can th
en be
use
d
to co
m
pute an e
s
ti
mate of the system state
)
(
ˆ
t
X
,
the estimatio
n
equatio
n gi
ven by:
i
(t)
1
)
(
(t)
~
~
ˆ
X
W
X
N
i
i
t
(2)
As sho
w
n Al
gorithm
1, we intro
d
u
c
e
a se
que
ntial
importa
nce
resampli
ng B
a
yesia
n
filtering (SIRBF), whi
c
h
an implem
en
tation of
the Markov chi
na Monte
Carlo resampli
ng
algorith
m
for trackin
g
probl
ems.
Algorithm 1. SIRBF for target tracking
for i=
1:
N do
/
/
Parti
c
les
s
a
mpling with SIR
)
z
,
~
/
(
~
~
)
:
(1
)
1
-
:
(1
)
(
)
(
t
i
t
i
t
i
t
x
x
q
x
, and set
i
t
i
t
t
x
x
x
)
(
1)
-
:
(1
)
:
1
(
~
,
~
end for
for i=
1:
N do
/
/
weights
updating
x
x
q
p
p
w
w
t
i
t
i
t
i
t
i
t
i
t
i
t
i
t
)
z
,
~
/
~
(
)
x
~
/
x
~
(
)
x
~
/
(z
~
~
)
(
)
1
-
(
)
(
1)
-
(
)
(
)
(
t
)
1
(
)
(
N
t
t
t
w
w
w
1
j
j
)
(
i
)
(
i
)
(
~
/
~
~
end for
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Seve
re-Dyna
m
ic Tracking
Problem
s Ba
sed o
n
Lo
wer Particle
s Re
sam
p
ling (Xu
ngao Zh
ong
)
4733
Let
N
j
t
eff
w
N
1
j
2
)
(
)
~
(
/
1
if Neff <
Nthres
hol
for i=
1:
N do
//
Resampling
with SIR
N
i
i
t
N
i
i
t
i
t
N
x
w
x
1
)
(
1
)
(
)
(
}
,1/
{
}
,
{
)
(
)
(
end for
else
N
t
t
N
t
t
w
x
w
x
1
i
i
)
(
i
)
(
1
i
i
)
(
i
)
(
}
~
,
~
{
}
,
{
)
(
)
(
The SIRBF
result
s a
r
e
sh
own
in
Figu
re 1, th
e
p
o
ssible
ta
rget po
sition
i
s
presented by
sampli
ng p
a
rt
icle
s with th
ei
r corre
s
pon
di
ng weight
s,
i.e., the su
cce
ss
of the SIRBF highly reli
es
on its ability to maintain a good approximation to
the posterior dist
ribu
tion, but there
exist
potential p
r
im
ary drawback of sampli
ng
approa
che
s
f
o
r a la
rge
nu
mber
of parti
cle
s
are re
qui
r
ed
to guara
n
tee
suf
fi
cie
n
t sam
p
ling in the broad state
spa
c
e,
and the e
s
timation a
c
cura
cy is linea
rly
with the num
ber of particl
e
s, as shown in Figur
e 2(a) the tracking accu
racy will improves
with
the co
unt of particl
es i
n
crease, Figu
re
2(b
)
sho
w
that the com
p
u
t
ational cost i
s
fit with cu
b
i
c
polynomial b
a
se
d on num
ber of sampli
ng, given by:
5
6
2
8
3
14
10
749
.
6
10
947
.
3
10
062
.
1
10
335
.
4
-
)
(
x
x
x
x
f
(3)
Therefore, th
e high comp
utational bu
rden cau
s
ed
by a large nu
mber of pa
rticle
s often
make
s the SI
RBF infe
asi
b
l
e
for practi
cal
appli
c
at
io
ns.
In view of a
b
o
ve proble
m
s, we
propo
se
d
a method
asso
ciated
with
neural net
works to ai
ds
t
he re
sa
mplin
g-ba
se
d Bay
e
sia
n
filtering
to
redu
ce th
e p
a
rticle
co
unt, while
maint
a
ining
trackin
g
quality an
d
the co
mputa
t
ional dem
an
ds
remai
n
lower
comp
are with
the tradition particl
e filteri
ng.
(a)
(b)
Figure 1. Illustration of the Possi
ble Ta
rg
et Po
s
i
tio
n
wa
s
Pr
es
en
te
d b
y
Particles with
Num
b
er
100, 150
0 Re
spe
c
tively, the Estimation
Accu
ra
cy of
Target Po
sition
is Linea
rly wi
th the Numbe
r
of Particle
s, (
a
) re
sult of SI
RBF, (b) the
den
sity distrib
u
tion
of samp
ling parti
cle
s
meet norm
a
l
distrib
u
tion
10
0
120
140
160
18
0
200
220
0
0.
02
0.
04
0.
06
0.
08
0.
1
0.
12
100
120
140
160
180
200
220
0
0.
01
0.
02
0.
03
0.
04
0.
05
P
a
rti
c
l
es
de
n
s
it
y
Possible position
Possible position
Particles densit
y
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4731 – 4
739
4734
(a)
(b)
Figure 2. Illustration of the Estimation Accuracy
of Possible Ta
rg
et Position Agai
nst to the
Comp
utation
a
l Co
st of the SIRBF Algori
t
hm, (a
) the e
s
timation result of SIR under differe
nt
numbe
r of pa
rticle
s (from 5
0
to 1500), th
e tracking
a
c
curacy is in
creased with in
cre
a
si
ng pa
rti
c
le
numbe
r, (b
) the com
putati
onal cost is fitting wi
th qua
dratic b
a
sed
on numb
e
r of
samplin
g, the
c
o
s
t
fit with fu
nc
tion f(x)=
a
3
*
x
3
+a
2
*
x
2
+a
1
*
x
+a
0
.
3. Neural Netw
ork Auxilia
r
y
Particle Filtering
SIRBF algorit
hm involves t
he sim
u
lation
of mu
ltiple p
a
rticle
s hyp
o
theses, in
sce
ne to high
dimen
s
ion
a
lity of mod
e
l-b
a
se
d tracking
, the real
-tim
e pe
rforman
c
e remain
s
ch
alleng
e [14].
Th
e
solutio
n
s in [
4
] propo
se
s a
GPU-a
c
cele
rated PF for
3
D
visual tra
cking appli
c
atio
n, in [15] aim to
lowe
r the
dim
ensi
onality of
the p
r
oble
m
, and [1
6] hav
e bee
n mo
dified to mi
nimize the n
u
mbe
r
of
particl
es m
e
e
t
ing to redu
ce
the computat
ional cost.
Con
s
trai
ned t
he count
s of
particl
es with
low-
level hy
pothe
se
s is t
he effective
resol
u
tion
f
o
r l
o
we
r t
h
e
com
put
at
i
o
n
a
l
cost
f
o
r S
I
RB
F
,
howev
er,
a cert
ai
n
num
ber of
sam
p
l
e
s a
r
e
st
ill
necessa
ry to captu
r
e the severe dyna
mi
c motion du
e to the broad
n
e
ss of the wh
ole state spa
c
e,
so,
simply d
e
c
re
ase the
nu
mber of pa
rti
c
le
s inev
itabl
e severe
dete
r
iorate the t
r
a
cki
ng a
c
cu
ra
cy.
Herein
we
p
r
esent
a m
e
thodol
ogy u
s
i
ng
con
s
tr
ain
ed p
a
rti
c
le
s-weig
hts
pai
r to a
pproxim
ate
estimation
th
e po
steri
o
r P
D
F, an
d the
deterio
ra
te
d perfo
rman
ce
cau
s
e
s
by
mi
nimize
d
pa
rti
c
le
s
wa
s com
pen
sated via aide
d by radial ba
sis fun
c
tion n
eural n
e
two
r
k (RBF
NN).
Accordi
ng to the Equat
ion (2) the target po
sition
will be optim
a
l approximated by
sele
cting and
simul
a
ting statistically
rel
e
vant
su
bsets with
en
oug
h lar
ger num
bers, ho
weve
r,
numbe
r-co
nst
r
aine
d pa
rticl
e
sub
s
et
s will
dire
ctly
leadi
ng to deteri
o
rated er
ro
rs
caused by ab
sent
particl
es. Th
e
r
efore, the de
sire
d ta
rg
et position
sho
u
ld
be given by:
)
(
)
(
ˆ
)
(
ˆ
*
t
t
t
par
X
X
X
(4)
Whe
r
e
t
par
)
(
X
refers
to dete
r
iorate
d erro
r
cau
s
es
by Minimi
zed
pa
rticle
s.
In this pa
p
e
r a
method to
co
mpen
sate fo
r the erro
rs
t
par
)
(
X
is prop
osed to i
m
prove th
e e
s
timation
accura
cy
of the SIRBF
with lowe
r pa
rticle
s by inco
rpo
r
ati
ng the
RBFNN into the stat
e e
s
timation stag
e. As
appe
ars Fi
gu
re
3, the
RB
FNN i
s
em
be
d into
re
sa
m
p
ling-ba
sed
Bayesian
filtering
to ove
r
come
he hig
h
com
putational
bu
rden
cau
s
ed
by a large
n
u
m
ber
of pa
rticle
s. The
alg
o
rithm of
neu
ral-
Bayesian
re
sampling filteri
ng (NB
R
F
)
co
nsi
s
ts of
syst
em model, SIR sam
p
ling, state estimatio
n
.
The RBF
N
N we cho
s
e
n
with one hi
dden laye
r whi
c
h is the
most widel
y sprea
d
architectu
re t
y
pe, and th
e
activation fun
c
tion of th
e hi
dden
nod
es i
s
cho
s
en
to b
e
a
radial
ba
sis
function, the
output of each hidde
n neu
r
on with Ga
ussian b
a
si
s fun
c
tion is d
e
fi
ne
d as:
)
-
exp(
1
2
1
1
i
i
b
n
G
C
(5)
2
/
1
2
1
1
2
/
1
1
1
1
)
(
]
)
)(
[(
k
i
i
i
g
c
T
T
T
G
C
G
C
G
C
(6)
Com
p
utational cos
t
/s
Particle
nu
m
b
e
r
/n
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Seve
re-Dyna
m
ic Tracking
Problem
s Ba
sed o
n
Lo
wer Particle
s Re
sam
p
ling (Xu
ngao Zh
ong
)
4735
Whe
r
e
n
m
R
G
is the
input
sam
p
le
s
set a
nd
m
n
R
1
C
is t
he
central ve
ctor
with th
e i
t
h eleme
n
t
denote
d
a
s
1
m
i
R
g
,
1
m
i
R
c
,respe
ctively;
1
i
b
is the th
re
shold of ith
n
euro
n
in th
e
hidde
n
layer. Th
e
ou
tput of the
ne
twork i
s
th
e li
near
su
m
of t
he o
u
tputs of
the
hidde
n
n
euro
n
s.
So, t
h
e
comp
en
satio
n
for the errors for targ
et position e
s
tima
tion is app
rox
i
mated as:
2
1
1
2
)
(
i
j
n
j
j
par
b
n
w
t
X
(7)
Whe
r
e
2
i
w
and
2
i
b
are the weig
hts and thre
shol
d of output layer, respe
c
tively.
(t
)
z
N
1
i
i
1)
-
(t
}
{
x
N
1
i
i
(t)
}
~
{
x
)
(
ˆ
*
t
X
}
,
{
)
(
)
1
-
(
1)
-
(
t/t
t
t
f
W
X
1
T
)
(
t
V
)
(
t
W
)
1
-
(
t
X
}
,
{
)
(
(t
)
)
(
t
t
h
V
X
N
1
{1/
N
}
N
1
i
i
(t
)
}
~
{
w
)
z
,
~
/
(
~
~
)
:
(1
)
1
-
:
(1
)
(
)
(
t
i
t
i
t
i
t
x
x
q
x
N
i
i
t
i
t
w
x
1
)
(
)
(
}
,
{
)
(
)
(
t
X
)
(
k
X
)
(
t
par
X
)
(
ˆ
t
X
Figure 3. The
Structure of Neu
r
al
-B
ayesian Re
sam
p
li
ng Filterin
g (NBRF
)
4. Experimental Re
sults
To test the empiri
cal pe
rfo
r
man
c
e of ou
r tracking a
p
p
r
oa
ch, we col
l
ected severa
l motio
n
seq
uen
ce
s th
at involve se
vere dyn
a
mic in vario
u
s scen
ario
s,
whi
c
h in
clu
d
ing
the low-fra
m
e-
rate, sudd
en
dynamic cha
nge
s, and
th
e su
dde
n
dy
namic chan
g
e
s
with d
o
wn
sampli
ng vid
eos.
All the expe
ri
ments we
im
plemente
d
to
co
mpa
r
e
th
e
tra
cki
ng
perf
o
rma
n
ce
of
o
u
r tracke
r, i.e
.,
NBRF a
nd SIRBF with different count of
particle
s
.
Scena
rio 1 lo
w-fram
e-rate
vide
o:
We ha
ve first invest
igated ho
w th
e numb
e
r of
particl
e
affects the tracking p
e
rfo
r
manc
e of SIRBF and
ho
w ca
n ou
r pr
opo
sed
NBRF to improve
the
tracking
p
r
e
c
ise
even
wit
h
lo
wer p
a
rti
c
le
s. In
thi
s
experim
ents,
the te
st vid
eos with
ran
dom
smooth
motio
n
, whi
c
h a te
nnis
run
n
ing
on the floo
r,
we b
egin
by giving the
co
mpari
s
o
n
of the
tracking p
e
rf
orma
nce of SIRPF and NBRF on
a
low-fra
m
e-rate video squa
sh, whi
c
h is
downsample
d
by keepin
g
one
fram
e
in every 1
0
frame
s
fro
m
a ori
g
inal vi
deo. In o
r
de
r to
qualitatively e
v
aluating the
impact of th
e
cou
n
t of
pa
rticle
s, we te
st
SIRBF with
s
a
mple
s 10
0 a
n
d
1000
scena
ri
os, and the
sampl
e
fram
es of re
sult
are
sho
w
n in
Figure 4. In
Figure 4
(
a
)
, the
perfo
rman
ce
of SIRBF wit
h
lower
100
particl
es is
b
ad, even lo
st
the tra
cki
ng
due to the
ab
rupt
motions
cau
s
ed by
seve
re
frame
d
r
op
pi
ng, whe
r
ea
s
whe
n
the
pa
rticles in
cre
a
si
ng to
100
0 th
e
result of SIRBF (Figure 4(b)) i
s
accu
rat
e
ly tracki
ng the ball after frame 1
9
th(o
ri
ginal fram
e 190),
but this at the co
st of large numb
e
r o
f
parti
cle
s
, while our
NBRF method (Fi
gure 4
(
c)) ca
n
effectively de
alt with this d
i
fficulty using
only lower
1
00 sample
s,
and
kee
p
ing
more
accu
rat
e
ly
perfo
rmed th
an SIRBF through
out the seque
nce.
We the
n
pe
rform a q
uantit
ative comp
ari
s
on of tracki
ng accu
ra
cy betwe
en SIRBF and
NBRF with di
fferent particl
es to further
verify
that th
e use of RB
FNN d
o
e
s
he
lp on a simil
a
rly
test video. The com
p
a
r
iso
n
is ba
sed o
n
t
he position
error in pixel
s
, as follows:
)
,
(
)
,
(
i
g
i
g
i
p
i
p
i
y
x
y
x
e
, i = 1,2,3
(8)
W
h
er
e
)
,
(
i
p
i
p
y
x
is th
e e
s
timation
value of ta
rg
et po
sition
by SIRBF o
r
NBRF,
)
,
(
i
g
i
g
y
x
is
the
grou
nd
-truth positio
n.
Tra
c
king error
i
s
eval
uated
as the differen
c
e between th
e mea
s
u
r
em
ent
positio
n and
the positio
n
estimat
ed
by the SIRBF or NB
RF. Th
e
trackin
g
pe
rforma
nce of
the
SIRBF and t
he NB
RF
with differe
nt p
a
rticle
s a
r
e
compa
r
ed i
n
Figure 5, the
positio
n e
rro
r of
NBRF
is
app
arently lo
we
r
than that
of S
I
RBF with th
e
sam
e
p
a
rticl
e
s, a
s
sho
w
n
the result wit
h
100 pa
rticle
s.
On the other hand, it is worth n
o
te that all five e
x
perime
n
ts d
e
mon
s
trate o
u
r
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4731 – 4
739
4736
NBRF p
e
rfo
r
med high
er
accuracy tha
n
SIRB
F even with lower pa
rticle
s, and the SIRBF
perfo
rmed
sa
me as
NBRF
but co
st more
than 10 ti
mes numb
e
r of p
a
rticle
s. We b
e
lieve that the
improve
d
tra
c
king
perfo
rm
ance of NB
RF is mai
n
ly
d
ue to the p
r
o
posal RBF
N
N to comp
en
sate
the errors during the constrain part
icl
e
s,
to cover the possible moti
on uncertainty
will be a
direct
solutio
n
to i
m
prove
the t
r
ackin
g
p
r
e
c
i
s
e,
me
an
whil
e,
the RBFNN can
be well
de
crea
se the
comp
utationa
l cost via by constrain the n
u
mbe
r
of part
i
cle
s
and
kee
p
ing the well perfo
rman
ce
s.
Figure 4. Tra
cki
ng Perfo
r
mances of th
e Two T
r
a
c
ke
rs on L
o
w-fra
m
e-rate Vide
o (green targ
et
positio
n, red trackin
g
re
sult
), (b) SIRBF
with 100
p
a
rti
c
le
s, (c)SIRB
F
with 1000 p
a
rticle
s, (d
) o
u
r
NBRF
with 1
00 parti
cle
(a)
(b)
Figure 5. Tra
cki
ng Error
wi
th Different Pa
rticle
Con
d
itions, (a
) SIRB
F, (b) NB
RF
Scena
rio 2 b
oun
ced ball
with dyn
a
m
i
c chang
es
: T
o
further qu
a
litatively evaluate th
e
tracking p
e
rfo
r
man
c
e of ou
r NBRF, whe
r
e we
are
sh
own o
u
r exp
e
rime
nts a
r
e
a boun
ce
d table
ball that back and force st
ruck the floor
with su
dde
n dynamic
cha
nge
s. The un
expecte
d mot
i
on
dynamic m
a
kes the tra
c
kin
g
task
rathe
r
hard by an a
c
curate m
o
tio
n
model. Ou
r experime
n
ts
to
illustrate th
e
prop
osed
NBRF ap
proa
ch
can
effect
ive
l
y deal with t
h
is difficulty only usi
ng lo
wer
particl
es.
0
5
10
15
20
25
30
0
20
40
60
80
100
120
1
00 s
a
m
p
l
e
s
4
00 s
a
m
p
l
e
s
6
00 s
a
m
p
l
e
s
8
00 s
a
m
p
l
e
s
1
000 s
a
m
p
l
e
s
0
5
10
15
20
25
30
0
20
40
60
80
100
120
50 s
a
m
p
l
e
s
60 s
a
m
p
l
e
s
70 s
a
m
p
l
e
s
80 s
a
m
p
l
e
s
100 s
a
m
p
l
e
s
Fr
am
e index
Erro
r
(
p
ixels
)
Erro
r
(
p
ixels
)
F
r
am
e i
nde
x
(c
)
(a)
(b)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Seve
re-Dyna
m
ic Tracking
Problem
s Ba
sed o
n
Lo
wer Particle
s Re
sam
p
ling (Xu
ngao Zh
ong
)
4737
Sample fram
es
are
shown in Fi
gure 6
,
with
10
0
sa
mples,
our NBRF meth
od
(Fig
ure
6(c)) succe
ssfully tracked t
he bo
un
cing
pingp
ong th
ro
ugho
ut the
se
quen
ce.
Note
that, even wi
th
1000
sam
p
le
s, SIRBF me
thod (Fig
ure 6(b
)) p
e
rf
o
r
m
ed poo
r tra
c
king a
b
ility, experie
nci
n
g
a
signi
fi
cant
drif
t of the target
obje
c
t. Mo
re
over, SIRBF
(Figure 6
(
a
)) f
a
iled to
tra
c
k
the table
ball
in
most fram
es
usin
g the nu
mber of samp
les lo
wer 1
5
0
.
The frame
-
by
-frame
comp
arison
of the
positio
n
er
ro
r in
p
i
xe
ls
for
th
o
s
e tw
o tr
ac
ke
rs
is
sho
w
n i
n
Fig
u
re
7. It ca
n
be
seen
that
com
p
a
r
ed
with the SIRB
F tra
cki
ng
re
sult, ou
r met
hod
NBRF i
s
mo
re clo
s
e
r
to the groun
d-truth po
siti
on, it means th
at with the aid
of RBFNN t
h
e
performanc
e
of NB
RF is
better than the SIRBF,
du
e t
o
the
RBF
NN ca
n im
prove
the robu
stne
ss
of NBRF for
severe dyna
mi
c ch
ang
es.
Figure 6. Tra
cki
ng Perfo
r
mances of th
e Tw
o T
r
a
c
ke
rs O
n
Dynami
c
Ch
ang
es Vi
deo (g
re
en
target po
sitio
n
, red tra
cki
n
g
result), (b
) SIRBF
with 1
50 parti
cle
s
, (c) SIRBF with
1000 pa
rticle
s,
(d) o
u
r NB
RF
with 150 pa
rticle
s.
Figure 7. The
Trackin
g
Position Erro
r in Pixels
of the SIRBF and th
e NBRF
with 100,50
0 ,100
0
particl
es, respectivel
Scena
rio 3 sudde
n
d
y
nam
ic chan
ge
s wi
th
low-fram
e-rate vide
o:
T
he
fi
nal experiment
is
to qualitativel
y evaluate t
he tra
c
king
perfo
rman
ce
of the
NBRF and
SIRB
F on
a synt
hetic
seq
uen
ce tha
t
involves the severe
abru
p
t motion of the obje
c
t ca
u
s
ed by fram
e
s
incon
s
e
c
uti
v
e
and lo
w-fram
e-rate video.
In this experi
m
ent, the sa
me 100
sam
p
les a
r
e u
s
e
d
for NB
RF
and
0
5
10
15
20
25
30
35
0
50
100
150
200
250
SI
RB
F
NBR
F
0
5
10
15
20
25
30
35
0
50
10
0
15
0
20
0
25
0
SI
R
B
F
NBR
F
0
5
10
15
20
25
30
35
0
50
100
150
200
250
SI
R
B
F
NB
R
F
Fr
am
e index
F
r
am
e i
nde
x
Erro
r
(
p
ixels
)
Erro
r
(
p
ixels
)
Erro
r
(
p
ixels
)
(a)
(b)
(c
)
Fr
am
e index
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4731 – 4
739
4738
SIRBF. The resul
r
of sam
p
le frames are illustra
ted in
Figure 8. It is obs
erved that our approach
can effe
ctively track the
obje
c
t throug
hout t
he se
q
uen
ce even
the ball bou
nce
d
ba
ck
with
sud
den
dyna
mic
cha
nge
s.
On th
e oth
e
r h
and,
SI
RBF freq
uentl
y
lose
the t
r
ack a
nd
poo
rly
perfo
rm on this seq
uen
ce
due to the large motion un
certai
nty. It mean
s that the perform
an
ce
of
our NB
RF is
better than th
e SIRBF, and it also
illustrate the neura
l
network pla
y
s an impo
rta
n
t
role i
n
error
compensation to
improve the NB
RF t
r
acking
abilit
y, experim
ents proved that
our
prop
osed
NB
RF with i
n
te
nsive tra
c
kin
g
perfo
rm
a
n
ce for la
rgely
uncon
strain
e
d
abrupt mot
i
on
even with le
ss numb
e
r of p
a
rticle
s.
Figure 8. Co
mpare the Tr
acking
Re
sult
s of SIRBF and NBRF wi
t
h
the smae
10
0 Particle
s on
the
Sudden
Dyna
mic ch
ang
es
with Lo
w-fra
m
e-rate video
(gre
en targ
et position, re
d tracking
re
sul
t
)
For the sake of test our propo
sed tra
cker’s
rob
u
stn
e
ss fo
r differe
nt numbe
r of particle,
many othe
r t
r
acking
expe
ri
ments with
th
e video
sa
me
as Fig
u
re
8,
the succe
s
sful tra
c
king
rat
e
with two tra
c
ker comp
are
in Ta
ble
1,
whi
c
h
demo
n
s
trated
that t
he tra
c
king
perfo
rman
ce
o
f
SIRBF is sen
s
itive to the numbe
r of part
i
cle
s
, it is
worth noting that if particle
s
less 50 the SIRBF
will lo
st the tracking
ability, and
with m
o
re tha
n
10
00
parti
cle
s
the
su
cce
ssful t
r
acking
rate
will
towards sta
b
ility 83%. Whi
l
e our NB
RF
only with lo
wer 1
00 p
a
rticl
e
s p
e
rfo
r
med
good t
r
a
cki
n
g
whi
c
h the sa
me as SIRB
F with 1000
particl
es.
SO
, we can
kin
d
ly gets co
n
c
lu
sion that
our
method i
s
a
robu
st tra
cker no
matter with lo
wer
particl
es
co
ul
d be inte
nsi
v
e tracking t
he
uncon
straint motion.
Table 1. The
Successful T
r
acking
Rate (×10
0%)
of SIRBF and
NBRF with the
Different Particl
e
s
on Sudde
n Dynamic Chan
ges
with Lo
w-frame-rate Vi
deo
method
The numbe
r of p
a
rticles
50
100 300 500 1000
>1000
SIRBF
lost
0.41 0.57 0.65 0.81
0.83
NBRF
0.66
0.87 0.96 0.96 0.96
0.96
5. Conclusio
n
In this pape
r,
a method wi
th neural n
e
twork auxilia
ry
seqe
ntial importan
c
e
re
sampling
Bayesian
filtering (SIRBF
) has bee
n
p
r
e
s
ente
d
to
im
prove th
e pe
rforman
c
e fo
r
robu
st trackin
g
with lower pa
rticle
s. The SIRBF can e
n
l
a
rge the
sam
p
ling varia
n
ce to cover the
possible mot
i
on
uncertainty,
however, the
high
co
mput
ational b
u
rd
e
n
cau
s
ed
by a la
rge
nu
mber of pa
rti
c
le
s
often make
s
the SIRBF infeasibl
e
for real-tim
e ap
pli
c
ation
s
. Therefore, a nov
el method was
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Seve
re-Dyna
m
ic Tracking
Problem
s Ba
sed o
n
Lo
wer Particle
s Re
sam
p
ling (Xu
ngao Zh
ong
)
4739
prop
osed wit
h
RBFNN merge tog
e
the
r
with
resa
mpling
-
ba
sed
Bayesian filtering, and the
RBFNN h
a
s
useful to
imp
r
ove
the
t
r
a
c
king preci
s
e even
with
l
e
ss pa
rticle
s. Many
compa
r
e
experim
ents
illustrate that
our
proposed NBRF
with
robust tracking performance for largely
uncon
strain
e
d
ab
rupt m
o
tion only
req
u
ire lower
num
b
e
r of
parti
cle
s
com
pare
with SIRBF, on
the
other ha
nd, o
u
r metho
d
also with strong
adaptive for d
i
fferent numb
e
r of parti
cle.
Ackn
o
w
l
e
dg
ements
This wo
rk i
s
sup
porte
d
by Nationa
l Scien
c
e
F
ound
ation
of Chi
na
und
er
Gra
n
t
6130
5117.
Referen
ces
[1]
SY Chen. Kalman Filter fo
r
Robot Vision: A Survey
.
IEEE Transactions on Industrial Electronics
.
201
2; 59(1
1
): 4409-
442
0.
[2]
G Gualdi, A Pr
ati, R Cucc
hi
ar
a. Multistag
e
P
a
rtic
le W
i
ndo
ws for F
a
st and
A
ccurate Object Detection.
IEEE Transactions on Patter
n
Analys
is and M
a
chi
ne Intel
lig
e
n
ce
. 201
2; 34(
8): 1589-
16
04.
[3]
Rafae
l
Mu
ñoz-
S
ali
nas, R
Me
d
i
na-C
a
rn
icer, F
J
Madr
id-C
uev
as. A C
a
rmo
na
-Po
y
ato.
Peo
p
l
e
detecti
on
and track
i
n
g
w
i
th multip
le ster
eo
cam
e
ras us
ing particl
e
filte
r
s.
Jo
u
r
n
a
l
o
f
Vi
su
al
C
o
mmun
i
ca
tio
n
and
Imag
e Re
pres
entatio
n
. 200
9; 20(5): 339-
35
0.
[4]
J Antho
n
y
Br
o
w
n,
D
a
vid
W
Caps
on. A
fra
m
e
w
ork f
o
r 3
D
mod
e
l-b
a
se
d
visual
tracki
ng
usin
g a
GPU-
accel
e
rated pa
rticle
filter.
IEEE Transactions
on Visuali
z
at
ion and Comput
er Draphics
. 2012; 18(
1)
:
68-8
0
.
[5]
L Bai, Y W
a
n
g
. Road track
i
ng usi
ng p
a
rti
c
le
fi
lters
w
i
th
partitio
n
sampl
i
ng a
nd a
u
x
il
ia
r
y
vari
ab
les.
Co
mp
uter Visi
on an
d Ima
ge
Und
e
rstand
in
g
. 2011; 1
15(1
0
): 1463
–1
471.
[6]
W
Kun, W
Ji
e, S
T
ao.
Improve
m
ent of fission bo
ot
strap particl
e filterin
g
. 10th Internati
o
n
a
l
Confer
ence
on
Electronic Me
asurem
ent & Instru
ments (ICEMI). Chengd
u.
201
1; 3: 94-10
0.
[7]
Z
Khan, T
Balch, F
.
Dellaert.
An MCMC-bas
ed particle
fi
lt
er
for tracking
multipl
e
i
n
teracti
ng targ
ets
. In
Proc. ECCV. Pragu
e. 200
4; 3024: 27
9-2
90.
[8]
K Choo, D Fleet.
Peopl
e trac
king
usin
g hy
b
r
id Mont
e Car
l
o
fi
lterin
g
. In Proc. ICCV. Vancouver, BC.
200
1; 2: 321–
3
28.
[9]
Y Atchad, JS
Rose
nthal. On
ada
ptive M
a
rk
ov cha
i
n M
ont
e Car
l
o
alg
o
rit
h
ms.
Bernoulli
.
200
5; 1
1
(5):
815-
828.
[10]
GO Roberts,
JS Rosenth
a
l.
Coupl
in
g an
d er
go
dicit
y
o
f
adaptive M
a
rkov cha
i
n Monte Carl
o
algorithms.
Jou
r
nal of ap
pli
ed
prob
abi
lity.
200
7; 44(2): 45
8-4
75.
[11]
Xi
uzh
uan
g Z
h
ou, Yao
Lu, J
i
w
en
Lu, Ji
e Z
hou. Abru
pt Motion T
r
ackin
g
Via Inte
nsiv
el
y Ad
aptiv
e
Markov-Ch
a
i
n
Monte Car
l
o S
a
mpli
ng.
IEEE Transactions on image processing
. 20
12; 21
(2): 789-8
01.
[12]
F
Liang, C L
i
u, RJ Carrol
l
. Stochastic
ap
pro
x
i
m
ation i
n
Mont
e Carl
o comp
utation.
Jo
urna
l of Amer
ic
a
n
Statistical Association.
200
7; 102(4
77): 30
5-3
20.
[13]
M Isard, A Blake. Co
nde
ns
ation-c
ond
itio
n
a
l de
nsit
y pr
o
pag
atio
n for vi
sual tracki
ng.
Internatio
na
l
Journ
a
l of Co
mputer Visi
on
. 1
998; 29(
1): 5-2
8
.
[14]
AC Sa
nkara
n
a
r
a
y
an
an, A Sr
i
v
astava, R
Ch
ella
pp
a. Alg
o
rit
h
mic a
nd
arch
itectural
optim
i
z
ations
for
computati
o
n
a
ll
y efficie
n
t parti
cle filterin
g.
IEEE Transactions on Im
age P
r
ocessing
. 200
8; 17(5): 737-
748.
[15]
Y W
u
, J
Lin,
T
Hua
ng. A
nal
yz
i
ng
an
d C
aptur
i
ng Artic
u
lat
ed
Han
d
Moti
on
i
n
Imag
e S
e
q
u
ences.
IEEE
T
r
ansactio
n
s o
n
Pattern Ana
l
ysis and Mac
h
i
ne Intell
ig
ence
.
2005; 2
7
(12):
191
0-19
22.
[16]
M.Bray
, E Koller-Meier, L
Va
n Gool. Smart
particle filtering
for high-dim
ensional track
i
ng.
Co
mp
uter
Visio
n
and I
m
a
ge Un
dersta
ndi
ng
. 200
7; 106(
1): 116-1
29.
Evaluation Warning : The document was created with Spire.PDF for Python.