TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.1, Jan
uary 20
14
, pp. 558 ~ 5
6
4
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i1.3381
558
Re
cei
v
ed
Jun
e
5, 2013; Re
vised July
1
1
, 2013; Accept
ed Augu
st 6, 2013
Face Tr
acking Based on Particle Filter with Multi-
feature Fusion
Zhiy
u
Zhou*
1
, Dichong Wu
2
, Xiaolong
Peng
1
, Zefei
Zhu
1
, Chuan
y
u
Wu
1
, Jinbin Wu
1
1
Z
hejia
ng Sci-T
e
ch Un
iversit
y
,
Hangz
ho
u 310
018, Ch
in
a
2
Z
hejia
ng Un
iv
ersit
y
of F
i
n
anc
e & Economics
,
Hangzh
ou 3
1
001
8, Chi
n
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: zhouzh
i
yu
19
93@
163.com
A
b
st
r
a
ct
T
r
aditio
nal
par
ticle filter c
a
n
n
o
t acco
mmo
d
a
t
e to the e
n
vir
o
n
m
e
n
t of bac
kgrou
nd i
n
terf
erenc
es,
illu
min
a
tion v
a
riatio
ns an
d o
cclusio
ns. This
pap
er pres
ent
s a face tracki
ng
meth
od w
i
th fusio
n
of col
o
r
histogr
a
m
,
co
n
t
our
feat
ures a
nd grey
mod
e
l
base
d
on parti
cle fi
lter. F
i
rst, it bro
u
g
h
t in
co
ntour fe
atures
as
the
mai
n
cu
e
of multip
le fe
atures w
hen
tra
cking th
e
face
w
i
thout stabl
e
color
histo
g
ra
m. T
h
e
n
, as p
r
ior
infor
m
ati
on w
a
s neg
lecte
d
in
traditi
ona
l p
a
rtic
le filter
, this
pa
per e
m
ploy
ed
GM(1,1) mo
de
l
to yiel
d pr
opo
sa
l
distrib
u
tion, su
ch that the pr
o
posa
l
distri
buti
on w
oul
d be
ar
a hi
gher
ap
pr
o
x
imatio
n to po
sterior pro
b
a
b
il
ity.
F
i
nally,
in th
e i
m
p
o
rtanc
e sa
mp
lin
g step, s
a
mpli
ng w
a
s c
o
rresp
ond
ed to
the p
a
rticle w
e
ig
ht in c
a
se o
f
th
e
particl
e de
gra
datio
n. T
he e
x
peri
m
e
n
ts sh
ow
that our meth
od o
u
tper
forme
d
the pr
evio
us w
i
th more
accuracy
an
d
flexi
b
ility,
pa
rticularly
u
nde
r the c
ond
itio
n of c
o
lor
b
a
ckgrou
nd
inter
f
erences,
dras
tic
illu
min
a
tion var
i
atio
ns an
d co
mp
lete occ
l
usi
ons.
Ke
y
w
ords
: F
a
ce tracking, col
o
r histogr
a
m
, contour
feat
ures
, particle filter, GM(1,1) mod
e
l
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
Particle filter
[1-3] is
wildly
use
d
recentl
y
, having sol
v
ed do
zen
s
of trackin
g
issue
s
. To
overcome the
impacts of fa
ce ro
tatio
n
, complexion int
e
rferen
ce an
d partial o
ccl
usio
n, Jianp
o
G
[4] propo
se
d
a face tracki
ng metho
d
with the cue
s
of colo
r an
d sha
pe ba
se
d
on pa
rticle fi
lter.
Hui T
[5] put
forth a parti
cle
filte
r
ing
method
wi
th
the fusi
on
of
col
o
r an
d t
e
xture fe
atures
,
descri
b
ing th
e face features with the
distin
ctive en
vironme
n
tal adapta
b
ility of weighted
color
histogram and rotated composite
wav
e
lets. To c
ope with interferences
of illumination and pose
variation
s
, Ju
an W [6] p
r
e
s
ented a n
e
w f
a
ce trackin
g
method
with the cue
s
of co
lor an
d conto
u
r.
All these me
thods h
o
ld some rob
u
st
perfo
rm
an
ce,
neverthele
ss, yet
they never handl
ed the
con
d
ition of complete o
c
cl
usio
ns. As th
e tradition
al particl
e filter i
gnored the g
u
idan
ce effe
ct of
prio
r informat
ion up
on the
prop
osal di
stribution,
sim
u
l
a
ting the p
o
sterior
pro
babi
lity distributio
n
woul
d be very difficult. Haitao Y [7] expl
oited t
he searching ability of par
ticle swarm optimization
throug
h a
no
nliner an
d n
o
n
-Ga
u
ss
mult
imode
dist
rib
u
tion. Ming
qi
ng Z
[8] empl
oyed hi
story
state
estimation
as the prior i
n
fo
rmation to yield the
propo
sal di
stributio
n. Yet all the method
s abo
ve
failed to solve
the interfere
n
ce i
s
sue
s
in a poor
colo
r e
n
vironm
ent.
In orde
r to bo
ost the relia
bi
lity of
face tra
cki
ng,
this p
a
per treats
col
o
r info
rmatio
n as th
e
first cue a
nd
conto
u
r
as th
e se
co
nd
cue
,
in addition
with an imp
r
ov
ed GM
(1,1
)
model to yeil
d the
prop
osal di
stribution. In the
importan
c
e
sampling
pro
c
ess, sam
p
lin
g is a
c
cordi
n
g to the parti
cle
weights to all
e
viate the particl
e degradation, thus the
accuracy
will
be enhanced
further m
o
re.
2. Particle Filter
w
i
th Mul
t
i-Feature Fusion
Traditio
nal p
a
rticle filte
r
merely em
pl
oy
s the
col
o
r a
s
the
single featu
r
e
,
which
posse
sse
s
a
fairly po
or
a
c
cura
cy. In t
h
is
pape
r,
co
lor
featu
r
e
s
, conto
u
r
fe
atu
r
es
a
s
well as
GM(1,1
) mod
e
l will be mel
t
ing all togeth
e
r. A
nd then
an ada
ptive pro
c
ed
ure wil
l
be pro
c
e
edi
ng
by snatchi
ng
a dominate
d
value each time in fluc
tuat
ed weig
hts. F
o
llowin
g
are
major di
stinct
ive
feature
s
used
in this pape
r, they will fuse
together to p
e
rform
an ad
aptive face tracking.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
TELKOM
NIKA
Vol. 12, No
. 1, Janua
ry 2014: 558 – 5
6
4
559
2.1. Color Fe
ature
s
The color
sp
ace of
HSV dra
w
ing
cle
a
r
bou
nd
s bet
wee
n
bri
ghtn
e
ss and
col
o
r, it is
invulnerable
to illuminatio
n variation
s
. In this
paper, to minimize the impa
cts of lumin
ance
comp
one
nt V, a quantizati
on pro
c
e
s
s wi
ll be perfo
rm
e
d
with model
8X8X4. Whe
n
cal
c
ulatin
g the
colo
r hi
stog
ram, wei
ghts impo
sed
by discrepa
nt
influen
ce fa
ctor of
parti
cl
es,
which h
a
ve
alleviated the
influen
ce of
backg
rou
nd i
n
formatio
n o
n
to the pixel
s
locate
d in e
d
ges. T
h
is a
r
ti
cle
selects Epanechnikov Kernel func
tion as the weight,
whi
c
h
could
i
m
prove the
reliability of
color
distrib
u
tion.
Suppo
se th
e
colo
r hi
stog
ra
m in candi
da
te are
a
is
)
(
y
p
, weighted
col
o
r histog
ram
of face model
is
1,
ˆˆ
{}
uu
m
qq
, and then the similarity function
()
y
betwee
n
face m
odel an
d
can
d
idate mo
del will be def
ined a
s
:
1
ˆ
ˆˆ
ˆ
()
[
(
)
,
]
(
)
m
u
u
yp
y
q
p
y
q
(1)
2.2. Conto
u
r Features
Face
contou
rs a
r
e th
e ve
ry discri
minativ
e f
eatures a
s
well, it’s
re
asonabl
e to
ch
o
o
se
the
conto
u
rs as t
he se
con
d
cue. This pa
p
e
r em
ploy
s a
method of shape context to extract th
e
conto
u
r fe
atu
r
es, i
n
itially p
r
opo
se
d by B
e
longi
el S
et
al. [9]. Firstly, Can
n
y ope
ra
tor is
appli
ed
to
acquire the face
edge, and
a unifo
rm sampling will be
processed
af
te
r it. For each feature i
n
t
he
sampli
ng
co
ntours, to bu
ild pola
r
co
o
r
dinate
sy
ste
m
in their
center p
o
sitio
n
. Suppo
se
one
coo
r
din
a
te of
sel
e
cte
d
feat
ure
is (x
0
, y
0
), anoth
e
r feat
ure’
s
coo
r
di
n
a
te
in i
denti
c
al contou
rs is (x,
y), then the polar coo
r
dinat
e (r,
) is cal
c
ul
ated. The co
o
r
dinate of
on
e
single featu
r
e in conto
u
rs
is a
bout th
e
vector len
g
th an
d a
ngle
histo
g
ra
m
with relation
to anoth
e
r fe
ature. A
stati
s
tic
numbe
r of tot
a
l features l
o
cated i
n
e
a
ch cell
of
rele
vant pola
r
co
ordin
a
te
will then b
e
cou
n
ted
up. After a
n
o
rmali
z
atio
n tran
sition, the
face
histog
ra
m
H
s
F
and th
e
candid
a
te obj
e
c
t hi
stogram
H
s
O
will be
kno
w
n. Then, Chi
-
aqua
re me
asure will b
e
used to descri
b
e the simila
rity
s
d
:
2
11
()
()
0.5
*
()
()
ij
nK
oF
s
ij
ik
oF
hk
h
k
d
hk
h
k
(2)
Whe
r
e
()
i
o
hk
is
the can
d
idate co
ntours
histo
g
ram,
()
j
F
hk
denotes
face mod
e
l hi
stogram.
2.3. Proposa
l
Distributio
n Gener
a
te
d w
i
th GM(1,1) Model
The traditio
n
a
l GM(1,1
) m
odel sele
cts t
he first
value
of original
seque
nce as t
he initial
value, whi
c
h
lacks
of rat
i
onality as
well as the
o
re
tical ba
si
s to
some
deg
re
e. A minimum
indicator fu
nction is
set to
determi
ne th
e initial
value
in [10], havi
ng imp
r
oved
the accu
ra
cy
of
GM(1,1
)
gen
erally. Th
rou
gh a
solution
of first-o
r
de
r
homog
ene
ou
s e
quatio
n, the time
re
sp
on
se
seq
uen
ce at the mome
nt k+1 is:
(1
)
(
0
)
((
1
)
)
ak
uu
xx
e
aa
（
k+
1
）
(3)
Whe
r
e
is t
he adj
ustm
e
n
t coef
fici
ent
.
Then th
ro
ugh a
re
cov
e
ry of first-orde
r
regressive, the predi
ction v
a
lue of
GM(1,1) model will
be updated as:
(0
)
(
1
)
(
1
)
ˆ
ˆˆ
()
=
(
)
-
(
1
)
xk
xk
xk
(4)
And the cova
rian
ce of pa
rticle filter is:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
Face T
r
a
c
kin
g
Based o
n
Particle Filte
r
with Multi-fea
t
ure Fu
sion (Zhiyu Z
hou
)
560
()
(
0
)
(
0
)
1
ˆˆ
ˆ
(
(
))
(
(
))
N
ii
i
T
kk
k
k
i
wx
xk
x
x
k
(5)
Finally
, the grey predi
ction
model ge
ne
ra
tes the propo
sal di
stributio
n, which is:
(0
)
1
ˆ
ˆ
(|
,
)
(
(
)
,
)
ii
kk
k
k
qx
x
z
N
x
k
(6)
2.4. Face Tra
cking Metho
d
Here are the major
stage
s
of computatio
n in our meth
od:
1)
T
o
cal
c
ulate
the histo
g
ra
ms
H
s
F
from front, side
an
d ba
ck
part
in huma
n
fa
ce.
Th
e
gene
rated o
u
t
puts will be p
r
eserve
d in the related ve
ctor
.
2)
T
o
sele
ct
a rectan
gle win
dow and co
mpute
t
he
weighted
col
o
r histog
ram
of face in t
hat
wind
ow a
s
the matchi
ng
template.
Then the thre
shol
d of color simila
rity
T
1
will be
cal
c
ulate
d
.
3)
T
o
initiali
ze
N particl
es i
n
the center
of fa
ce
area, the initial weight
of
each particle will
be
set as 1/N.
After that, configure the len
g
th of GM(1,1) model, m=5, and engen
derin
g the
initial GM(1,1
) model.
4)
T
o
d
enote th
e ren
e
wed p
o
sition
of ea
ch parti
cle, a
c
cording to
propo
sal di
strib
u
tion of the
GM(1,1
) mod
e
l. In the previous m fram
es, t
he positi
on of particle
s
is to be updated ba
sed
on Gau
s
s dynamic m
odel.
5)
T
o
count
the colo
r
hi
stogra
m
s
in the
cen
t
er win
d
o
w
of each
ren
e
we
d parti
cle
s
. Each featu
r
e
is
to be c
o
mpared to the fac
e
template, f
o
llowing
with
a crite
r
ion of
simila
rity betwee
n
them.
6)
T
o
u
pdate
the
weig
ht of ea
ch p
a
rti
c
le o
n
the gr
oun
d o
f
similarity val
ue.
T
o
sort th
e pa
rticle
s
judgin
g
by ea
ch weight val
ue.
7)
T
o
pi
ck u
p
the maximum
weig
ht, decid
ing wh
ether it
s value withi
n
the given thre
shol
d
T
2
.
The value
of thre
shold
T
2
rest
s
wit
h
s
i
milarit
y
of
contours hi
sto
g
ram
d,
T
2
=0.3*d. After
capturing the
new position, GM(1,1) model will be updated.
8)
In the followi
ng sta
ge of
resam
p
ling, 1/
3
of the sort
ed pa
rticle
s
will be
dupli
c
ated twi
c
e,
while the re
m
a
inde
rs
will duplicate only once.
The du
plicatin
g pro
c
ess will never cea
s
e until
the sum of pa
rticle
s app
ro
a
c
he
s N.
3. Experimental Re
sult a
nd
Analy
s
is
The first experime
n
t proceed
s by a
face
tra
cki
ng
sequen
ce p
r
ovided by Stanford
University
. In this task, 100 particle
s
are sele
ct
ed to condu
ct the contra
st work.
The contra
st
results of
refe
ren
c
e [6], [8] and o
u
r m
e
th
od are sho
w
n
in Figu
re 1. I
n
the tra
d
ition
a
l parti
cle filte
r
,
particl
es deg
eneration o
c
curs frequ
ent
ly
, thus targe
t
s will lo
se
focu
s e
a
sily
with the
simil
a
rity
interferen
ce
s of color
.
Th
e conto
u
rs cue is in
corp
orated into t
he parti
cle
s
in referen
c
e [6],
amelio
rating robu
stne
ss wh
en encounte
r
i
ng the com
p
lexion interference in the backgroun
d.
The
obje
c
ts in
re
feren
c
e [8]
will get lo
st
easily
wh
e
n
encounte
r
ing
the simila
rit
y
interfere
n
ce of
compl
e
xion.
With a m
e
rg
e of contou
rs as th
e se
co
nd cue in
ou
r method, pl
u
s
with
sto
r
ag
e of
side
and
ba
ck face info
rm
ation in a
d
va
nce. O
u
r m
e
thod mai
n
tain
s a
rob
u
st tra
cki
ng, even
when
the colo
r feature
s
are not
evident and a
drasti
c ro
tati
on of face occurs no
w and
then. In addition
with grey information to yield the propo
sal distri
butio
n
,
the
trouble
of particle de
grad
ation will
be
alleviated.
Th
e erro
r contrast di
a
g
ram
of four m
e
tho
d
s m
entione
d
above in
x a
nd y directio
n
is
sho
w
n a
s
Fig
u
re 2, the defi
n
ition of error is:
||
100%
ro
r
pp
e
p
(7)
Whe
r
e
p
r
den
otes the
real
positio
n of hu
man face, an
d
p
0
is the o
b
serve
d
value
of face po
siti
on
with related track
i
ng methods
,
▽
p
r
d
enot
es the auth
e
n
t
ic displ
a
cem
ent of two adj
ace
n
t frame
s
.
The second experiment is performed wit
h
dr
asti
c illum
i
nation variati
ons.
The illumination
cha
nge
s app
ear a
s
a stron
g
interfere
n
ce
factor onto
the colo
r
,
the traditional pa
rticle filter fails to
deal this pro
b
lem rea
s
on
ably
,
thus easily losing focus in some face tra
cki
ng environ
ment.
The
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
TELKOM
NIKA
Vol. 12, No
. 1, Janua
ry 2014: 558 – 5
6
4
561
RGB sp
ace is employed in
referen
c
e [6], performi
ng well in a colo
r-stable
situatio
n. But once the
illumination
chang
es d
r
a
s
tically
, the tra
c
king ef
fe
ct
would be im
pai
red b
adly
.
As a fusion of the
color cue proposed in
referenc
e [8], the tracking
ef
fect woul
d still
be weakened once the col
o
r
V
a
rie
s
a
c
utel
y
.
The tra
cki
n
g
re
sult of fo
ur
meth
od
s mentione
d ab
ove is sho
w
n
in Figure 3,
and
the error
cont
rast dia
g
ram in x and y direction is
sho
w
n as Figu
re 4.
(a)
Re
sults of
traditional pa
rticle filter
(b)
Re
sults of
refere
nc
e [6]
(c
) Re
sult
s of refere
nc
e [8]
(d)
Res
u
lts
with our method
Figure 1. Con
t
rast re
sult
s o
f
four
method
s (F
rame 3
2
, 86, 98, 137)
(a) T
r
a
cki
ng
error in di
re
ction x
(b) T
r
a
cki
ng
error in di
re
ction y
Figure 2. Erro
r cont
ra
st dia
g
ram of fou
r
method
s
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
Face T
r
a
c
kin
g
Based o
n
Particle Filte
r
with Multi-fea
t
ure Fu
sion (Zhiyu Z
hou
)
562
(a)
Re
sults of
traditional pa
rticle filter
(b)
Re
sults of
refere
nc
e [6]
(c
) Re
sult
s of refere
nc
e [8]
(d) Res
u
lts
with
our
method
Figure 3. Tra
cki
ng re
sult
s with dra
s
tic ill
um
ination va
riations (Fra
m
e
27, 30, 46, 49)
(a) T
r
a
cki
ng
error in di
re
ction x
(b) T
r
a
cki
ng
error in di
re
ction y
Figure 4. Error c
ont
rast diagram
with drastic illumination variations
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
TELKOM
NIKA
Vol. 12, No
. 1, Janua
ry 2014: 558 – 5
6
4
563
(a)
Re
sults of
traditional pa
rticle filter
(b)
Re
sults of
refere
nc
e [6]
(c
) Re
sult
s of refere
nc
e [8]
(d)
Res
u
lts
with our method
Figure 5. Tra
cki
ng re
sult
s with occlu
s
io
ns (F
ram
e
40
, 43, 46, 48 )
The third exp
e
rime
nt tracks the fa
ce wi
th compl
e
te
occlu
s
ion
s
. T
he traditio
nal
particl
e
filter perform
s a barely sa
tisfacto
ry face tracki
ng. A
s
is propo
se
d in referen
c
e [6], comple
xio
n
and contou
rs information
are ju
st too b
a
rren to fo
cu
s the targ
ets.
Tradition
al GM(1,1
) mod
e
l is
applie
d in
referen
c
e
[8] to
yield
pro
p
o
s
al di
strib
u
tion
, enjoying
ce
rtain
rob
u
stn
e
ss in
several
ca
se
s of
occlusio
ns.
This pap
er prese
n
ts a
n
im
pro
v
ed GM
(1,1
) mod
e
l, which up
date
s
th
e
predi
ction
to
alleviate th
e pa
rticle
de
grad
ati
on. T
he exp
e
rim
e
ntal outp
u
ts
have p
r
oved
ou
r
method
s
with
a hig
her ro
b
u
stne
ss an
d
efficien
cy
. Figure
5
sho
w
s the
re
sults
of face t
r
a
cki
ng
with occlu
s
io
ns.
4. Conclusio
n
This pap
er p
r
esents
a particle
filter
fa
ce
tra
cki
ng
wit
h
data fu
sio
n
of col
o
r hist
ogra
m
,
conto
u
r hi
sto
g
ram. And a
merge of G
M
(1,1) m
odel
is attached
to yield prop
osal di
strib
u
tion,
approximatin
g the
re
sults to aut
h
entic po
sterio
r
probability di
st
ribution
mo
re cl
osely. It is
demon
strated
throu
gh sev
e
ral
trackin
g
tasks
t
hat th
e
new metho
d
coul
d
solve t
he p
r
oble
m
s
o
f
stron
g
col
o
r interfe
r
en
ce
s, d
r
a
s
tic ill
umination
va
riation
s
and
co
mplete
o
ccl
usi
o
n
s
, a
nd it
outperfo
rm
s the previo
us
with better tra
cki
ng ro
bu
stn
e
ss and hi
gh
er co
mputatio
nal efficien
cy.
Ackn
o
w
l
e
dg
ment
This
work i
s
sup
porte
d by
Zhejian
g
Provincial
Natu
ral Scie
nce
Found
ation o
f
China
(No. LY1
3
F0
3001
3, R11
1
0502
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
Face T
r
a
c
kin
g
Based o
n
Particle Filte
r
with Multi-fea
t
ure Fu
sion (Zhiyu Z
hou
)
564
Referen
ces
[1]
Agustie
n
Sir
a
d
j
ud
din I, R
ahm
at W
i
d
y
anto
M
,
Basarud
d
i
n
T
.
Particle F
ilter
w
i
t
h
Ga
ussia
n
W
e
ighti
n
g
for Huma
n T
r
acking.
T
E
LKO
M
NIKA T
e
leco
mmu
n
icati
on
Co
mp
uting E
l
e
c
tronics a
nd C
ontrol
. 20
12
;
10(4): 80
1-8
0
6
.
[2]
Ke
yan L, Yu
nh
ua L, Sha
nqi
n
g
L, Lia
ng T
,
L
e
i W
.
A ne
w
para
lle
l particl
e
filter face tracking meth
o
d
base
d
on h
e
ter
oge
ne
ous s
y
stem.
Journa
l of Real-T
i
m
e Ima
ge Process
i
ng
.
2012; 7(3): 1
5
3
-16
3
.
[3]
Suche
ndr
a MB
, Xi
ngzh
i
L. Int
egrate
d
d
e
tecti
on a
nd
track
i
n
g
of multi
p
l
e
fa
ces usi
ng
parti
cle filter
ing
and
optic
al fl
o
w
-
b
as
ed e
l
asti
c matchin
g
.
C
o
mputer V
i
sio
n
and
Image
U
ndersta
ndi
ng
. 200
9;
11
3(6):
708-
725.
[4]
Jian
po G, Yuji
an W
,
Hao Y,
Z
hen
ya
ng W
.
Particle f
ilt
er fa
ce trackin
g
usi
ng co
lor sh
ap
e
histogr
am a
s
clues.
Jour
nal
of Imag
e an
d Graphics
. 20
07
; 12(3): 466-4
7
3
.
[5]
Hui T
,
T
i
ngzhi
S, San
y
ua
n Z
,
Bing H. F
a
ce
tr
acking
alg
o
ri
thm combi
ng c
o
lor
and te
xtur
e featur
e
s
base
d
on p
a
rti
c
le filter.
T
r
ans
action of Bei
j
i
n
g Institute of Techn
o
lo
gy
. 201
0; 30(4): 46
9-4
73.
[6]
Juan W
,
Yan
x
i
a
J, Caiho
ng T
.
F
a
ce tracking
base
d
on p
a
rti
c
le filter usi
ng
color h
i
stogra
m
and cont
ou
r
distrib
u
tions.
Opto-Electro
nic Engi
neer
in
g
. 2012; 39(
10): 32
-39.
[7]
Haita
o
Y, Fuxi
Z, Haiqian
g
C
.
Face tr
acking
based o
n
ad
a
p
tive PSO particle filter.
Ge
om
a
t
i
cs an
d
Information Sci
ence of W
u
h
a
n
University
. 20
12; 37(4): 4
92-
495.
[8]
Ming
qin
g
Z
,
Z
h
ili
ng
W
,
Z
ong
hai
C. Vis
ual
trackin
g
a
l
gor
ithm b
a
sed
o
n
gre
y
pred
iction
mod
e
a
n
d
particl
e.
Contro
l and D
e
cisi
on
.
2012; 2
7
(1): 5
3
-57.
[9]
Mod G, Belongiel S, Malik J.
Shap
e c
ontext
s
en
abl
e
efficie
n
t retriev
a
l
of s
i
milar
sh
apes
.
Procee
din
g
s
of IEEE Conference o
n
Com
p
uter Visio
n
an
d
Pattern Reco
g
n
itio
n. 200
1: 723-7
30.
[10]
Junfen
g L, W
e
nzha
n D.
Res
e
arch o
n
the am
elior
a
tin
g
GM(1,1) Mod
e
an
d
Its Applicati
on i
n
the p
o
w
e
r
qua
ntit
y
mo
de
li
ng of sha
ngh
ai
cit
y
.
Systems
Engi
neer
in
g T
heory an
d Practi
cal
. 200
5; 25(3
)
: 140-14
4.
Evaluation Warning : The document was created with Spire.PDF for Python.