TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.7, July 201
4, pp
. 5621 ~ 56
2
8
DOI: 10.115
9
1
/telkomni
ka.
v
12i7.415
2
5621
Re
cei
v
ed Au
gust 15, 20
13
; Revi
sed Ma
rch 2
0
, 2014;
Acce
pted April 9, 2014
Object Trackin
g
Based on Multiple Features Adaptive
Fusion
Jie Cao
1
, Lei
l
ei Guo*
1,2
, Jinhua Wang
3
, Di Wu
3
1
Colle
ge of Co
mputer an
d Co
mmunicati
on,
L
anzh
ou U
n
iver
sit
y
of T
e
chnol
og
y,
Lanz
ho
u 730
0
50, Chi
n
a
2
T
e
chnol
og
y & Rese
arch Ce
nter of Gansu
Manufactur
i
n
g
Information En
gi
neer
ing,
Lanz
ho
u 730
0
50, Chi
n
a
3
Colle
ge of Ele
c
trical an
d Information En
gin
e
e
rin
g
, Lanz
hou
Univ. of
T
e
ch,
Lanz
ho
u 730
0
50, Chi
n
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: 7455
41
228
@
qq.com
A
b
st
r
a
ct
Multipl
e
fe
atur
es fusi
on
bas
e
d
trackin
g
is o
ne
of t
he
most
active res
earc
h
in track
i
n
g
lit
e
r
ature, In
this p
aper,
a n
o
vel
ad
aptiv
e fusio
n
strategy
i
s
prop
os
e
d
for
mu
ltipl
e
fe
atur
es fusio
n
, b
a
se
d o
n
tw
o co
mmo
n
used fus
i
o
n
rul
e
s: prod
uct rul
e
an
d w
e
ig
hte
d
su
m rul
e
. T
h
is strategy e
m
ploys
particl
e fi
lterin
g techn
i
q
ue,
prod
uct rule a
nd w
e
ig
hted s
u
m r
u
le
are u
n
ifie
d into
a
n
ada
ptive fra
m
ew
ork th
roug
h
define
d
featu
r
es
distanc
e. In practice, the ne
w
fusion strategy s
how
s mor
e
robustn
ess than pr
oduct fu
sion a
nd w
e
ig
hted
sun rul
e
.
Ke
y
w
ords
: o
bj
ect tracking, pa
rticle f
ilterin
g, features d
i
stanc
e,
multi
p
l
e
features fusio
n
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
Targ
et tracki
ng is o
ne of
the co
re technol
o
g
y in compute
r
visi
on, it has
wi
despre
ad
appli
c
ation
s
i
n
human
-co
m
puter inte
ra
ction, su
rveill
ance, visual
servoing
and b
i
omedi
cal ima
ge
analysi
s
[1]. Tra
cki
ng b
a
sed on m
u
lti-fe
ature fu
sion
h
a
s the
r
efo
r
e
been
an a
c
tive re
sea
r
ch to
pic
for over a d
e
c
ad
e, in ord
e
r
to make the
tracking
m
o
re rob
u
st an
d
more
stable, i
t
fuses m
u
ltiple
feature
s
in
clu
ded
colo
r, ed
ge an
d motio
n
feature.
Th
e main
strate
gies
of feature fusio
n
h
a
ve
prod
uct
rul
e
and wei
ghted
su
m rule. Th
e
alg
o
rithm
s
prop
osed
by [
2
, 3] i
s
the
typical
exampl
e of
using product rule to multi-feat
ure tracki
ng. Further,
the probability
framew
ork of a combinati
o
n
tracking
alg
o
r
ithm p
r
op
osed by [4], the alg
o
rithm
s
ba
se
d o
n
prod
uct
rule
get rea
s
on
able
analysi
s
u
n
d
e
r thi
s
p
r
oba
bility framework.
On th
e
o
t
her
h
and, weighted su
m rule plays a very
importa
nt rol
e
in multi-fe
ature fusi
on
tracki
ng. Th
e
algorithm p
r
opo
se
d by [5] which used
weig
hted
su
m rul
e
to
app
roximate j
o
int
likeli
hoo
d. M
o
reove
r
, the
r
e a
r
e
som
e
f
u
sio
n
meth
od
s of
multi-feature, such as th
e algorith
m
of online
swit
chi
ng feature propsed by
[6], the algorith
m
of
online
switchi
ng tra
c
ker propo
sed
by [7
]. Likewi
se,
min max fu
si
on rule p
r
o
p
o
se
d by [8]
and
demo
c
ratic el
ectoral fusio
n
propo
se
d by [9].
In orde
r to co
mbine p
r
od
uct rule’s
advan
tage and
wei
ghted
sum ru
le’s adva
n
tag
e
,there
are two achie
v
ed ways at
pre
s
ent. The
first achieved
way is that timely swit
chin
g fusion
strat
egy
according
to different scen
ario
s.
Th
e al
gorithm
p
r
op
ose
d
by
[10]
is the
typical
example
of t
he
first a
c
hi
eved
way, thi
s
al
gorithm
e
s
timats
se
co
nd
order mom
e
nt of the
wei
ghted
sam
p
l
e
set
and computin
g its Frob
eni
us no
rm to d
enote ho
w fe
ature
s
are rel
i
able, and the
n
swit
ch the two
fusion rule
s in time. Obviously, if there are mo
re
sa
mples, this al
gorithm’
s
re
al
-time nee
d to test
and ve
rify. The se
co
nd a
c
h
i
eved way is t
hat ma
ki
ng
produ
ct rul
e
an
d wei
ghted
sum rul
e
unifie
d
into an
ad
ap
tive frame
w
o
r
k,
whi
c
h
adj
ust the
weig
hts
of produ
ct
rule and
the weig
hts of
weig
hted su
m rule in the
tracking
re
sul
t
s acco
rdin
g to adaptive fa
ctor. Th
e alg
o
rithm p
r
op
o
s
ed
by [11] is the typical example
of the se
con
d
achi
eve
d
way,this al
gorithm defin
es a ne
w fea
t
ure
uncertainty m
easure
m
ent
method to ad
just the
relati
ve contrib
u
tio
n
s of different
feature
s
.
Curre
n
tly information ent
ro
py theory has been
su
cce
s
sfully applied
to information
fusion
theory
[12
-
14
].This
pap
er prop
oses an adaptive
fu
si
on strategy b
a
se
d on info
rmation ent
ro
py
theory. ou
r
algorith
m
ma
ke
pro
d
u
c
t rule an
d
wei
ghted
sum
rule unifie
d
i
n
to an
ada
p
t
ive
frame
w
ork a
c
cordi
ng to
defined feat
ure
s
di
st
an
ce. An extensive num
ber of compa
r
a
t
ive
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5621 – 56
28
5622
experim
ents
sho
w
th
at the
propo
sed
al
gorithm
is
mo
re
stabl
e a
n
d
ro
bu
st than
prod
uct
rul
e
and
weig
hted su
m rule in the
obje
c
t trackin
g
.
In the re
st of
this pa
per,
we expl
ain th
e sh
ortcomin
gs a
nd o
u
r al
gorithm i
n
Se
ction 2.
Experimental
re
sults
and
analysi
s
a
r
e re
porte
d i
n
Sectio
n 3.
We
con
c
lu
de this
pap
e
r
in
Section 4.
2. Adap
tiv
e
Fusion Stra
teg
y
Based on Particle Filter
2.1. Adap
tiv
e
Fusion Str
a
tegy
The m
odel
o
f
prod
uct
rul
e
a
s
sume
s t
hat the fe
atu
r
e i
s
in
dep
e
ndent
of ea
ch othe
r.
Namely, ea
ch feature g
e
n
e
rate
s ind
epe
ndent ob
se
rv
ation, then th
e joint likelih
o
od of n features
can b
e
expre
s
sed a
s
:
1
1
(|
)(
|
)
n
ni
i
p
zz
x
p
z
x
(1)
Among the a
bove formul
a
s
:
i
z
is the ob
se
rvation of i-th
f
eature which is inde
pen
d
ent of
each other, x is the state of
estimated target. Eq
.(1) is
simple b
u
t including ri
ch inf
o
rmatio
n.
Another
com
m
on fusi
on rule is weight
ed su
m ru
le,
it is a very effective tool for the
compl
e
x of density estimat
i
on pro
b
lem.
i
t
s spe
c
ific fo
rm as follows:
12
1
(|
)
(
|
)
n
ii
i
p
zz
x
p
z
x
(2)
Among the
a
bove form
ula
s
:
i
sh
ows the
i-th feature
corre
s
p
ondin
g
to the weig
ht of
observation probability. the we
ight
s are
norm
a
lized to ensure
1
1
n
i
i
This p
ape
r p
r
opo
se
s a
n
a
daptive fusi
o
n
st
rategy which
com
b
ine
s
the a
d
vant
age
s of
prod
uct
rule
and
weig
hte
d
sum rule,
whi
c
h
can
ef
fectively solv
e the p
r
o
b
le
m of difficult
y to
disting
u
ish when the simil
a
r targ
et
close to the targe
t. This algorit
hm is ba
sed
on the fact: on the
one ha
nd, when the feat
ure
s
suppo
rt
to each ot
her, this illu
strate
s that the features
are
influen
ced by
a small de
g
r
ee of conta
m
ination,
the
n
usin
g prod
uct rul
e
ca
n improve tracking
accuracy. O
n
the oth
e
r
ha
nd, when
th
e
feature
s
d
on’t
su
ppo
rt to e
a
ch
othe
r, thi
s
illu
strate
s t
hat
the features
are i
n
fluen
ce
d by a g
r
e
a
t degree
of
co
ntamination, t
hen u
s
in
g
we
ighted
sum
rule
can mai
n
tain
the multi-mod
a
l of distributi
on, and supp
ress noi
se.
This
paper
will use the particle
filter tracker. T
h
eref
ore,
we can obtain
the sample’s
prob
ability assignm
ent. In
orde
r to
conv
enien
ce
of illustratio
n
, two
feature
s
a
r
e
denote
d
C
1
an
d
C
2
, the particl
es set ba
se
d on two features are
12
1
,(
|
)
,(
|
)
M
ii
i
i
xp
z
x
p
z
x
, M is the numb
e
r of
particl
es,
1
(|
)
i
p
zx
an
d
2
(|
)
i
p
zx
is the
wei
ghts
of pa
rticles
whi
c
h
obt
ained th
ro
ug
h define
d
likeliho
od mo
del ba
sed two feature
s
.
Defin
e
1.
if
1
(|
)
i
p
zx
and
2
(|
)
i
p
zx
is the p
a
rticle’
s
weig
hts of two fe
a
t
ures, th
en
we
build the mod
e
l of base
d
e
n
tropy as:
1
1
11
1
(|
)
(|
)
(
|
)
M
M
xx
pz
x
pz
x
p
z
x
1
2
22
1
(|
)
(|
)
(
|
)
M
M
xx
pz
x
pz
x
p
z
x
(3)
The inform
ation dista
n
ce o
f
feature
C
1
with feature
C
2
can b
e
define
d
as:
1
2
[(
|
)
]
[(
|
)
]
1
12
2
1
(,
)
[
(
|
)
]
l
o
g
i
i
pz
x
M
pz
x
i
j
dC
C
p
z
x
(4)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Obje
ct Tra
cki
ng Base
d on
Multiple Feat
ure
s
Adapti
v
e
Fusion
(Ji
e
Cao
)
5623
Then the pa
rt
ial cre
d
ible
co
efficient for
C
1
can be defin
ed as:
12
12
1
12
12
1-
(
,
)
(
,
)
1
()
1
(,
)
(
,
)
1
dC
C
i
f
d
C
C
Sup
C
dC
C
i
f
d
C
C
(5)
The inform
ation dista
n
ce o
f
feature
C
2
with feature
C
1
can b
e
define
d
as:
2
1
[(
|
)
]
[(
|
)
]
2
21
2
1
(,)
[
(
|
)
]
l
o
g
i
i
pz
x
M
pz
x
i
j
dC
C
p
z
x
(6)
Then the pa
rt
ial cre
d
ible
co
efficient for
C
2
can be d
e
fin
ed as:
21
21
1
21
21
1-
(
,
)
(
,
)
1
()
1
(,
)
(
,
)
1
dC
C
i
f
d
C
C
Sup
C
dC
C
i
f
d
C
C
(7)
Defin
e
2.
Th
e information
dista
n
ce b
e
twee
n featu
r
e
C
1
and
featu
r
e
C
2
can be
define
d
as:
12
21
[(
|
)
]
[
(
|
)
]
[(
|
)
]
[
(
|
)
]
12
12
2
2
11
(,
)
[
(
|
)
]
l
o
g
+
[
(
|)
]
l
o
g
ii
ii
pz
x
p
z
x
MM
pz
x
p
z
x
ii
jj
DC
C
p
z
x
p
z
x
(8)
Defin
e
3.
Th
e cre
d
ible
co
efficient between feature
C
1
and feature
C
2
ca
n be def
ined a
s
:
12
12
12
12
12
1-
(
,
)
(
,
)
1
(,
)
1
(,
)
(
,
)
1
DC
C
i
f
D
C
C
Sup
C
C
DC
C
i
f
D
C
C
(9)
Among the
above form
ul
as: in order to
prevent the denomi
n
ator is
zero,
equal
s
0.0001.
12
(,
)
Sup
C
C
reflect
s
the degree sup
p
o
rt
of features. Namely, whe
n
12
(,
)
Sup
C
C
is
relatively gre
a
ter, whi
c
h ill
ustrate
s
that featur
e
s
su
p
port ea
ch ot
her, then the
fusion re
sult
o
f
prod
uct rule
occupi
es a
n
importa
nt po
sition comp
ared with
weig
hted su
m rul
e
. On the ot
her
hand , wh
en
12
(,
)
Sup
C
C
is rel
a
tively smaller, which illustra
tes that features don’t support each
other, then th
e fusion resul
t
of weighted
sum rule
o
c
cupie
s
an imp
o
rtant po
sitio
n
comp
ared
with
prod
uct rule.
Acco
rdi
ng to
the sim
u
latio
n
re
sult
s of the ex
pe
rime
nts, thus we
think p
r
o
duct
rule
and
weig
hted su
m rule can be
unified into a
n
adaptive fra
m
ewo
r
k throu
gh the credibl
e coeffici
ent.
Defin
e
4.
Fra
m
ewo
r
k for a
daptive multi-feature fu
sion
can be d
e
fin
ed as:
1
1
1
1
1
[(
|
)
(
)
]
[
1
(
)
]
(
|
)
(|
)
(
)
n
n
i
i
n
i
i
n
n
ii
p
zx
U
x
S
u
p
C
C
p
zx
p
z
z
x
Sup
C
C
(10)
In Equation
(10
)
,
1
()
()
n
ii
i
i
Sup
C
Sup
C
, con
s
i
derin
g that
whe
n
the
weights of
particl
es
(the
value of feature’
s likeliho
od functi
on
)c
los
e
to
0
,
b
e
c
a
us
e
pr
od
uct r
u
le
mak
e
th
e
anothe
r feat
ure’
s
contri
b
u
tion be
com
e
sm
all for
these
pa
rticl
e
s, we affiliate the unif
o
rm
distrib
u
tion which i
s
dire
ctl
y
propo
rtional
to
this feature’s supp
ort to every feature
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5621 – 56
28
5624
3. Tracking
Algorithm Based on Pro
posed Fu
sio
n
Strateg
y
3.1. Particle Filter
Particle filter is a filtering method b
a
se
d on Mo
nte Carl
o an
d recursive Bayesian
estimation.
In
re
ce
nt years, it ha
s b
e
co
me a
n
effe
ctive tool fo
r ta
rget tra
c
king
unde
r n
o
n
-
lin
ear
or n
o
n
-
Gau
ssian
conditio
n
s [15, 16]. Th
e
detailed
de
scriptio
n
refer
pape
r [17], th
e pa
rticle
filter’s
principle
will not introduce
in this section .
In the algorit
hm achieving process, we
choi
ce
the
ellipse to descri
be the target’s state,
namely
,,
,
,
xy
x
y
xc
c
l
l
,
x
c
，
y
c
，
x
l
，
y
l
and
are the center coordinates of ellipse, the
long axis, the
sho
r
t axis an
d the defle
ction angl
e. Be
side
s, in the
particl
e filteri
ng techniqu
e, we
use
the
sim
p
l
e
st a
nd
mo
st co
mmonly fi
rst-ord
e
r lin
e
a
r
system
a
s
the
state tra
n
sition
mod
e
l
of
the particl
e filter.
3.2. Extrac
te
d Featu
r
es
Colo
r is
one
of the main f
eature
s
fo
r d
e
scribi
ng the
target, re
se
arche
r
have
m
ade a l
o
t
of study for target’
s
color
feature. A co
lor hi
stogram
prop
ose
d
by
[1] is used t
o
describe th
e
target’s
colo
r
feature, Its expre
ssi
on is:
1
()
(
(
,
)
)
c
B
c
u
hu
I
x
y
u
(11)
(,
)
I
xy
is the pixels
of candi
date
distri
ct,
c
B
repre
s
ent
s the lon
g
of color hi
st
ogra
m
.
Edge [18], a
s
an
other
efficient featu
r
e
des
cri
p
tor, can be u
s
e
d
here to
enh
a
n
ce th
e
power of col
o
r feature.
Pixels
(,
)
I
xy
are ev
enly extracte
d within th
e e
llipse, e
dge’
s stren
g
th
G
and dire
ction
a
ngl
e
are defin
ed a
s
:
22
(,
)
x
y
Gx
y
I
I
1
(,
)
t
a
n
(
)
y
x
I
xy
I
(12)
The ellip
se i
s
divided into f
our p
a
rt
s accordin
g to two
axes, for e
a
ch part, directi
on angl
e
is qu
antified
e
B
gra
de hi
sto
g
ram
s
, an
d t
hen fu
se the
edge’
s
stre
n
g
th inform
ation into e
a
ch
point, we
ca
n get the
weighted
gra
d
i
ent orie
ntation hi
stogra
m
of each
part. finally, the
histog
ram
s
of
four part
s
are combi
ned a
nd normali
ze
d.
For the de
scription of hist
ogra
m
, the Bhattach
aryya
coefficie
n
t is a popula
r
si
milarity
measure [19]. Consi
deri
n
g discrete
d
ensitie
s such
as two hi
stogra
m
s
mo
d
h
and
ta
r
h
, the
coeffici
ent is
defined a
s
:
mo
d
m
o
d
1
(,
)
1
(
)
(
)
B
ta
r
t
a
r
u
hh
h
u
h
u
(13)
On this ba
si
s ,the obse
r
ved
likeliho
ods
of
two feature
s
can b
e
define
d
as:
2
mo
d
(|
)
e
x
p
(
(
,
)
)
i
ii
t
a
r
pz
x
h
h
(14)
1,
2
i
, color i
s
the first feature,
ed
ge is the second feature.
3.3. Propose
d
Particle Filter Tra
cking
Algorithm
In s
e
c
t
ion 3.2, we introduce how to extrac
t features.
No
w, on the basi
s
of them
, the
detailed p
r
o
c
essing of imp
r
oved alg
o
rith
m is given.
(1) Initiali
zati
on: k=1, initialized particl
e
sets
1
,,
1
,
,
i
k
x
iN
N
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Obje
ct Tra
cki
ng Base
d on
Multiple Feat
ure
s
Adapti
v
e
Fusion
(Ji
e
Cao
)
5625
(2)
2,
,
f
kN
(a) Pr
edi
ction
:
()
1
~(
|
)
,
1
,
,
ii
kk
k
x
px
x
i
N
(b)
Two feat
ure
s
a
r
e em
powere
d
an
d
norm
a
lized:
Colo
r featu
r
e:
()
1
(
)
1
(|
)
ii
kk
p
zx
.
Edge feature:
()
2
(
)
2
(|
)
ii
kk
p
zx
(c) According
to Equation
(10) fusi
ng two feature
s
.
(d) Outp
ut: th
e target’s stat
e of
k
mome
n
t
can
b
e
cal
c
ulated by
the
weig
hted su
m
of
the
particl
es:
()
(
)
1:
1
[|
]
M
jj
kk
k
k
k
j
x
Ex
z
x
(e) A
c
cordin
g to the di
st
ribution
of p
a
rtic
le
s’ weig
hts
de
cidi
ng wheth
e
r re
sa
mple.
If
2
,
1
(1
)
2
N
tj
t
N
, then
()
1
,1
/
,
1
,
i
k
x
Mi
M
~
()
(
)
11
,,
1
,
,
jj
kk
x
jM
.Othe
r
wi
se, don’t d
eal.
4. Experimental Re
sults
and An
aly
s
is
First,
we set experim
ental para
m
eters
a
s
follo
ws
: the
initial po
sitio
n
of targ
et is given
manually, th
e nu
mbe
r
of
pa
rticle
s i
s
set to
200,
the u
n
iform
d
i
stributio
n
()
Ux
eq
uals
1
N
,
dire
ction
hist
ogra
m
B e
q
u
a
ls 1
8
,col
or
histog
ram B
equal
s 2
16.
The valu
e of
are sho
w
n
in
Table 1.
Table 1. The
Coeffici
ents o
f
Two Featu
r
es
corresponding
color
1
()
edge
2
()
Video sequence of expe
riment 1
90 30
Video sequence of expe
riment 2
90 40
Simultaneo
usly, in order to measure the trac
kin
g
erro
r, we define two measure mode.
2
ˆ
tt
AE
x
x
2
1
1
ˆ
T
tt
t
R
MS
E
x
x
T
(15)
AE
measu
r
e
s
each frame e
rro
r,
RMSE
m
easure
s
all frames e
r
ror.
4.1. Aircraft
Video
Experiment
1 use the video of model ai
rcra
ft, whi
c
h l
ength i
s
770 f
r
ames, The adaptive
coeffici
ents b
e
twee
n p
r
od
uct rule
and
weig
hted
su
n
rule
are
sho
w
n in
Figu
re
1, the
credi
ble
coeffici
ents of
feature
s
are
sho
w
n in Fig
u
re 2.
Figure 1. The
Adaptive Co
efficients of T
w
o
Fusio
n
Rul
e
s
Figure 2. The
Credi
ble Coe
fficients of
Features
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5621 – 56
28
5626
From Fig
u
re
1, we can
co
nclu
de that our
algo
rithm can adaptively
adjust the weights of
prod
uct
rule
and
weig
hte
d
su
m rul
e
according to
the ch
ang
e
of enviro
n
m
ent. Before
300th
frame, b
e
cau
s
e th
e e
n
viro
nment i
s
relat
i
vely simp
le, prod
uct rul
e
occupi
es
an
importa
nt
po
si
tion
in the trackin
g
. Ho
weve
r,
after 30
0th frame, be
ca
use the e
n
viro
nment i
s
rela
tively comple
x,
weig
hted su
m rule o
c
cupi
es an imp
o
rta
n
t position in
the tracking.
From
Figu
re
2, we
can
cl
e
a
r
ob
serve
th
at ou
r al
gorit
hm
can
adj
u
s
t the
weight
s of
two
feature
s
a
c
co
rding to th
e chang
e of env
ironm
en
t, Na
mely, our alg
o
rithm a
c
hiev
e the ada
ptive
cha
nge of fe
ature
s
’ wei
g
h
t
s within the
frame
w
or
k of
a unified fusion. Figure 3
sho
w
s p
a
rt
o
f
tracking
re
sul
t
s by the prop
ose
d
algo
rith
m, produ
ct rul
e
and weight
ed sum
rule.
Figure 3. Some Re
sults o
n
Experiment
1 by usi
ng :p
rodu
ct rul
e
(the first ro
w), sum rule (the
se
con
d
ro
w),
and propo
se
d
method (the
thir
d ro
w)
(Frames:2
81,28
3,500,51
8,64
9,770)
From
Figu
re
3. Beca
use
the chan
ge
of
illuminati
on is sm
all
and the
ba
ckgroun
d i
s
relatively sim
p
le from
first frame to
500t
h fram
e,
prop
ose
d
alg
o
rith
m, pro
d
u
c
t ru
le and
weight
ed
sum
rul
e
can
su
c
c
e
ssf
ully
t
r
ac
k t
h
e
t
a
r
g
et
.
Ho
wev
e
r,
as th
e target
moves
on, it i
s
o
c
clud
ed b
y
the tree,
as
sho
w
n
in fra
m
e 50
0 a
n
d
frame
51
8.
These ma
ke
it ch
allen
g
ing
for the t
r
a
cki
ng
algorithms
to follow the target. It c
an be s
e
en
that our al
gorith
m
over
come
s these difficulti
e
s
throug
hout th
e whole t
r
a
c
king p
r
o
c
e
ss,
whe
r
ea
s th
e
prod
uct
rule
and
weig
hted
sum
rule d
r
if
ts
away from th
e target d
u
ri
ng the tra
c
ki
ng proces
s a
nd finally loses the ta
rget
. This sho
w
the
stability and robu
stne
ss of
prop
osed alg
o
rithm.
4.2
．
Speak
er
Tracking
Experiment
2
is the
sp
ea
ker trackin
g
in
sma
r
t me
etin
g room, vid
e
o
length
i
s
5
0
0
frame
s
.
Tra
cki
ng difficultie
s are th
e rotation of h
ead an
d the o
ccl
usi
on of others sp
ea
ke
r.
First,
based
on e
qual
s
(1
5),
we
give t
he e
r
ror an
al
ysis. Be
sid
e
s, we
only
gi
ve the
frame
s
’ tra
cki
ng error b
e
tween 38
2-th frame to
402
-th frame when
the target is
occlud
ed.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Obje
ct Tra
cki
ng Base
d on
Multiple Feat
ure
s
Adapti
v
e
Fusion
(Ji
e
Cao
)
5627
Figure 4. The
X-axis Absol
u
te Erro
r
Figure 5. The
Y-axis Absol
u
te Erro
r
From Fig
u
re
4 and Figu
re
5, we can
co
nclu
de t
hat produ
ct rule i
s
sen
s
itive to noise, so
its AE is big
gest i
n
39
2-t
h
fram
e, mo
reover,
we
ig
h
t
ed sum
rule
isn’t
se
nsitive to n
o
ise, t
he
cha
nge of AE curve isn’
t obvious. T
he pro
p
o
s
ed
algorithm
combine
s
the
produ
ct rul
e
’s
advantag
e an
d weig
hted sum rule’
s
adv
antage, so its AE is smalle
st.
Table 2. Co
m
pari
s
on of RMSE Value
Fusion rule
RMSE
X Y
Weighted sum ru
le
0.2982
0.3129
Product rule
0.2936
0.2735
Our algo
rithm
0.2347
0.1970
From T
able
2. Comp
are
d
weighte
d
sum rule
, in
the X-axis,
our al
gorith
m
’s e
rro
r
redu
ce
d 0.06
35. in the Y
-
a
x
is, our
algo
ri
thm’s e
r
ror
re
duced 0.1
159
. Comp
are
d
prod
uct
rule,
in
the X-axi
s
, o
u
r
algo
rithm’
s e
r
ror re
du
ced 0.0
589,
i
n
the Y-axis,
our algo
rithm
’
s e
r
ror re
du
ced
0.0765.
So o
u
r
al
gorith
m
has better accuracy.
Fig
u
re 6
sho
w
s p
a
rt of tra
c
kin
g
re
sult
s by t
he
prop
osed alg
o
rithm, pro
d
u
c
t rule an
d weighted
sum rule.
Figure 6. Some Re
sults o
n
Experiment
1 by usi
ng: p
r
odu
ct rul
e
(the first ro
w), sum rule (the
se
con
d
ro
w),
and propo
se
d
method (the
third ro
w)
(Frames:5
3
,250,
392,50
0)
From Fi
gu
re
6, we
can
ob
serve
that wh
en t
he ta
rget’
s
he
ad o
c
clu
des
at fram
e
392, the
fusion
re
sults of produ
ct ru
le deviate the
cente
r
of target. Beside
s, the fusion
re
sults of wei
ght
ed
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5621 – 56
28
5628
sum
rule exi
s
t that som
e
particl
es b
e
g
in to tr
an
sp
ire. Ou
r algo
rithm can co
rre
ctly track
the
target. This
show the effe
ctiveness of propo
sed al
gori
t
hm.
5
．
C
o
nc
lu
s
i
on
After analyzi
ng the traditi
onal fu
sion
method i
n
clu
d
ing p
r
od
uct
rule
and
wei
ghted
su
m
rule,a
novel f
u
sio
n
alg
o
rith
m pro
p
o
s
ed
by this p
ape
r whi
c
h
unify the traditio
nal
method i
n
to
a
adaptive fusi
on frame
w
o
r
k acco
rdi
ng to adaptive
coefficient. In experim
ents,
the new fu
sion
strategy
sho
w
s mo
re
robu
stne
ss a
nd st
abilit
y than produ
ct fusion
and weighte
d
sun rul
e
.
On the other hand, in this pape
r, we
only
consi
d
er the features
of video, without
con
s
id
erin
g t
he featu
r
e
s
of audio. T
h
erefo
r
e,
we
will research
the fusi
on
strategy for fu
sing
audio a
nd vid
eo multi-featu
r
e in the next resea
r
ch.
Ackn
o
w
l
e
dg
ements
This work was su
ppo
rted
by the National
Scien
c
e
and Te
chn
o
l
ogy Support
Program
(No.
201
2BAF12B19
) ,the
Natio
nal
Na
tural S
c
ien
c
e
Fou
ndation
of Chi
na
(61
2630
31).
Th
e
Finan
ce Dep
a
rtment Fo
un
dation of Gan
s
u Provin
ce,
Chin
a(0
914Z
TB148).
Referen
ces
[1]
Z
h
iqi
nag
Ho
u, Cho
ngzh
ao
Ha
n.A surve
y
of visual
trackin
g
.
Acta Autom
a
tica Sinica.
20
06
; 3
2
(
4
)
: 60
3-
617.
[2]
Shen C
H
, Hen
gel A, Dick A.
Probab
ilistic
mu
ltipl
e
cue i
n
tegrat
io
n for particle b
a
sed t
r
ackin
g
.
Proc.
VIIth Digitl Image Com
puting: T
e
c
hniques and
Applicatio
ns, Melbourne, Austra
lia. 2003; I: 399-408.
[3]
LI PH, Chaumentte F.
Ima
g
e
cues fus-
ion
for ob
ject tracki
ng b
a
se
d o
n
p
a
rticle fi
lter
. Procee
din
g
s
o
f
the 3r
d Intern
a
t
iona
l W
o
rksh
o
p
o
n
Articu
late
d Motio
n
and
Deforma
ble
Objects. Pa
lma
de Ma
ll
orca,
Spai
n: Sprin
g
e
r
. 2004; 99-
107
.
[4]
Leic
h
ter I, Linden
ba
um M,
Rivli
n E.
A Proba
bil
i
stic framew
ork for co
mb
ini
ng tracki
ng al
gorith
m
s
.
Proceeding of
IEEE Conference
on
Computer Vision and Pattern
Rec
ognition, Washington, USA
.
200
4; (2): 445-
451.
[5]
Spen
gler
M, S
c
hiel
e B. T
o
w
a
rds ro
bust
mu
lti-cue
inter
g
rati
on
for
visual tracking.
In
te
rnati
o
n
a
l
Jou
r
nal
of Machin
e Visi
on an
d App
lica
t
- ions.
2003; (
14): 50-5
8
.
[6]
Coll
ins
R, LIU
Y.
On-lin
e se
lectio
n of d
i
scr
- imin
ative tra
cking fe
atures.
Procee
di
ng
o
f
Internatio
n
Confer
ence
on
Computer Vis
i
on,
Nice, F
r
anc
e. 2003: 3
46-3
52.
[7]
T
o
y
a
m
a
K, H
a
ger G. Increm
ental
focus
of attent
io
n for r
obust v
i
sio
n
-b
ased
trackin
g
. Internati
o
n
a
l
Journ
a
l Of Computer Visi
on. 199
9; 35(1): 45
-63.
[8]
Kittler J, Hatef M, et
al. On co
mbining classif
e
rs.
IEEE Trans.on PAMI
. 1998; 20: 22
6-23
8.
[9]
T
r
iesh J, Malsburg
C. Dem
o
cratic inte
g- ra
tion: self
orga
ni
zed i
n
t
egr
ation
of ad
aptiv
e c
u
e
. Neura
l
Co
mp
utation
. 2
001; 13: 2
049-
207
4.
[10]
Xi
ao
pin
Z
hon
g
,
Jianru
Xue,
Nan
n
in
g Z
h
e
n
g
, et al
. A
n
Ad
aptive Fus
i
o
n
Strateg
y
Bas
e
d Multi
p
le
Cu
e
T
r
acking.
Journal of Electro
n
i
cs & Informatio
n T
e
chon
lo
gy
. 200
7; 29(5): 10
17- 10
22.
[11]
Xi
n Gu, Haitao
W
ang, Lingfe
ng
W
ang, et al
. F
u
sing Multip
le F
eatur
es for
Object
T
r
acking Base
d
o
n
Uncerta
i
nt
y
M
e
asurem
ent.
Acta Autom
a
tica Sinica.
2
0
1
1
; 37
(5): 550-5
59.
[12]
Yong
De
ng,
D
ong
W
ang,
Qi
Li, et
al
. A
ne
w
meth
od t
o
a
nal
yz
e ev
id
enc
e co
nflict.
C
ontro
l
Th
eo
ry
&
Appl
icatio
ns
. 2
011; 28(
16): 08
39-0
844.
[13]
Ji-xian
g
Sun,
Hui-mi
n Shi,
H
ong-
qia
ng W
a
ng. T
he
T
heory Re
lative to E
n
trop
y
in Infor
m
ation F
u
si
on.
Chin
ese Jo
urn
a
l of Co
mp
uter
s.
2003; 26(
7): 796-
801.
[14]
Kull
back L
e
i
b
l
e
r div
e
rge
n
ce
[EB/OL]. [2010- 11-0
9
]. http://en.
w
i
ki
pe
dia.
or
g/
w
i
ki/KU
LLBA
C
K_l
e
ib
ler
_
diver
genc
e.
[15]
Gordon
NJ, S
a
lmo
nd
DJ, S
m
ith AF
M.
No
vel
appr
oac
h t
o
n
onl
in
ear/no
n
-ga
u
ssia
n
ba
yesia
n
stat
e
estimation.
IEE Proceed
in
gs o
n
Rad
a
r an
d Si
gna
l Processi
n
g
. 1993; 1
40(2)
: 107-11
3.
[16]
Gao Y, Yuan
GW
,
Z
hou H, et al. Object
D
e
scriptor C
o
m
b
ini
ng C
o
l
o
r a
nd Detecti
o
n
. TELKOMNIKA
Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2013; 1
1
(6): 3
072-
307
9.
[17]
Indah Ag
ustie
n
, Siradj
udd
in,
M
Rahmat W
i
d
y
anto, T
Basaru
ddi
n. Pa
rticle F
ilter
w
i
t
h
Gaussi
an
W
e
ighti
ng for
Huma
n T
r
acking.
T
E
LKOMN
I
KA Indon
esia
n Jour
nal
of E
l
ectrical
Eng
i
n
eeri
n
g
. 2
012
;
10(4): 80
1- 80
6.
[18]
Dala
l N, T
r
igg
s
B.
Histogra
m
s of
orie
nted
grad
ients for
hu
ma
n d
e
tecti
on.
Proce
e
d
i
n
gof the IEE
E
Computer Societ
y
Confer
enc
e on Com
puter
Vision and P
a
ttern Re
cognition.
San Diego, USA: IEEE.
200
5; 886-
893.
[19]
Perez P, Vermaak J, Blake A
.
Data fusion f
o
r visua
l
tracki
ng w
i
th particl
es
. Proceeding of the IEEE.
200
4; 92(3): 49
5-51.
Evaluation Warning : The document was created with Spire.PDF for Python.