TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.5, May 2014, pp
. 3521 ~ 35
2
8
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i5.3539
3521
Re
cei
v
ed
Jun
e
18, 2013; Revi
sed
De
ce
m
ber 11, 201
3; Acce
pted
De
cem
ber 3
1
,
2013
A Coarse-to-Fine Human Body Segmentation Method
in Video
Yingna Deng
*, Wenqing
Wang
Xi
’a
n Univ
ersit
y
of Posts and
T
e
lecommunic
a
tions,
We
i
g
uo
R
o
ad
, X
i
’
a
n
,
C
h
i
n
a
,
86
-0
29
-8
81
66
461
*Corres
p
o
ndi
n
g
author, e-ma
i
l
:den
g
y
in
gn
a@
126.com
A
b
st
r
a
ct
Hu
ma
n bo
dy precis
e seg
m
e
n
tation is d
i
fficult bec
a
u
se of inter-occl
usi
on
w
hen there are
multi
p
l
e
hu
ma
n b
odi
es
in vi
de
o. A c
oarse-to-fi
ne s
e
g
m
e
n
tati
o
n
meth
od
is pr
o
pose
d
. In co
a
r
se seg
m
entati
on,
hu
ma
n sha
pe
mo
de
ls are
us
ed to g
e
t hu
ma
n
’
s p
o
sitio
n
a
n
d
coars
e
reg
i
o
n
. T
he hu
man
mo
de
ls w
i
th va
riant
scale
an
d
post
u
re
are
constr
ucted w
i
th
he
a
d
, torso,
and
l
egs. F
o
r
eac
h
hu
ma
n
body,
i
t
s corresp
on
di
ng
hu
ma
n sha
pe
mo
de
l is obt
ai
ned
by mod
e
l
match
i
n
g
, and
by w
h
ich hu
man p
o
sitio
n
is
obtai
ne
d rou
g
h
l
y.
Hu
ma
n prec
is
e cont
our is
o
b
tain
ed
in fi
ne
seg
m
e
n
tati
o
n
by curv
e ev
ol
ution w
i
th
initi
a
l co
ntour
obt
ain
e
d
from coars
e
seg
m
e
n
tatio
n
. Experi
m
ent re
sults s
how
that the propose
d
meth
o
d
cou
l
d
segment hu
man
obj
ect precise
l
y.
Ke
y
w
ords
: Hu
ma
n sha
pe
mo
del, lev
e
l set, b
a
yesi
an
mod
e
l,
segmentati
on
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
Cro
w
d p
r
e
c
ise segm
entati
on is impo
rta
n
t fo
r human tracking a
nd reco
gnition. Howeve
r,
whe
n
there a
r
e multiple pe
destri
a
n
s
, hu
man obje
c
t p
r
eci
s
e
segm
e
n
tation is difficult be
cau
s
e
of
inter-occlu
s
io
n.The mo
del
s which hu
m
an bo
dy
se
g
m
entation b
a
s
ed
on
coul
d
be cl
assified
into
dynamic mo
del a
nd
app
eara
n
ce m
o
d
e
l. Dynam
i
c
model
s
a
r
e
made up of obje
c
t’s
m
o
ving
spe
ed, directi
on and
othe
r dynamic fe
a
t
ures [1
, 2].
Appea
ran
c
e
model
s are
made u
p
of color,
sha
pe, p
o
siti
on a
nd oth
e
r
featu
r
e
s
which
re
pre
s
e
n
t obje
c
t's a
ppea
ran
c
e.
Elgammald
and
Rama
nan
used color
and
positio
n information to co
n
s
tru
c
t obje
c
t
model
s un
der the assum
p
tion
of obje
c
t ente
r
ing the
cam
e
ra vie
w
lon
e
l
y [3, 4]. and
in this
con
d
ition, ea
ch o
b
je
ct’s
regio
n
was
obtaine
d by
ke
rnel
de
nsity estimation
.
Zhe m
odifi
ed Elga
mmal
d
’s m
e
thod,
and
used E
M
evolution fo
r
human
segm
entation [5].
Wu
and S
a
p
p
mad
e
up
h
u
man
body p
a
rt cl
assifications
with boo
sting
algorithm [6,
7]. Lu emplo
y
ed a coa
r
se
to fine method to seg
m
e
n
t human bo
dy
from photo
s
[8].
Huma
n shap
e model
s co
ud provide p
o
sition, a
r
ea,
postu
re, an
d som
e
othe
r rou
gh
informatio
n; it mean
s that based on h
u
m
an
shape
model
s, hum
an obje
c
t co
u
l
d be se
gme
n
ted
roug
hly. Ho
wever, getting
pre
c
ise conto
u
r of ea
ch
ob
ject is
ne
ce
ssary for h
u
ma
n obje
c
t preci
s
e
segm
entation
.
In this pap
er, a
metho
d
ba
sed
on
b
o
th hum
an
shape
mod
e
l
and l
e
vel
set
is
prop
osed. Hu
man obje
c
ts'
numbe
r, rou
g
h
a
r
ea
an
d
shape
mo
dels are o
b
taine
d
thro
ugh
initi
a
l
segm
entation
,
then o
n
th
e ba
si
s of i
n
itial s
egme
n
tation, ea
ch
obje
c
t’s pre
c
ise
conto
u
r i
s
obtaine
d by curve evolutio
n usin
g level set.
2. Coars
e
Segmenta
tion
Bas
e
d on Hu
man Shape
Model
2.1.
O
v
er
v
i
e
w
o
f
the
M
e
thod
Huma
n b
ody
is m
ade
up
of
hea
d, tarso,
arm
s
a
nd l
e
g
s
, an
d the
hu
man
sha
pe
chang
es
regul
arly, as
sho
w
n in
Fig
u
re 1
and Fi
g
u
re 2. So, 7
human
sh
ape
model
s with
different po
st
ure
are consturct
ed
cons
isting of ellipses
to
simulate
hu
man walki
ng
inclu
d
ing
3 front views an
d 4
side vie
w
s, a
s
sh
own in Figure 3.
Suppo
se th
e
obje
c
t re
gio
n
dete
c
ted
b
y
bac
kg
rou
n
d
su
btra
ction
is d
enote
d
as I,
θ
is
human
shap
e mod
e
l, th
e be
st seg
m
entation
re
sult
could
b
e
obtain
ed
by estimatin
g
the
maximum of posterior probability,
as shown in Equation (1).
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
0
46
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3521 – 35
28
3522
*
ar
g
m
ax
(
|
)
Θ
θ
P
θ
I
θ
(1)
Whe
r
e
θ
={
n
,{
M
1
,
M
2
,…,
M
n
}},
n
is the
numbe
r of
h
u
man
obje
c
t
s
,
M
i
s
the
huma
n
mo
del
para
m
eter, d
e
fined
a
s
M
={
h
,
height
,
l
}, an
d ea
ch fa
ctor
pre
s
ent
s he
a
d
po
sition, hu
man h
e
ight a
n
d
human p
o
se sep
a
rately.
Bas
ed on Bayes
i
an theory
, we k
n
ow that,
(|
)
(
|
)
(
)
P
θ
IP
I
θ
P
θ
(2)
Whe
r
e
P
(
θ
) i
s
the
object prior
probability,
P
(
I
|
θ
) is th
e simil
a
rity of
model
and
fo
reg
r
ou
nd
regi
on.
Suppo
se ea
ch obje
c
t app
eara
n
ces
with equal p
r
io
r pr
ob
ability, the segm
entat
i
on re
sult is
θ
*
whi
c
h maximi
ze the simila
rity
P
(
I
|
θ
).
Figure 1.
Side
Vie
w
o
f
H
u
ma
n
Wa
lk
ing
Figure 2.
Fro
n
t View of Hu
man Walkin
g
2.2 Head Candidates
2.2.1. Get He
ad Candidate Region
w
i
th Region Se
arch
For a
poi
nt h
in fore
gro
und
, let it as the
uppe
r left corner to
build
a
re
ctangl
e wit
h
heig
h
t
o
f
o
b
j
ec
t a
n
d
w
i
d
t
h
o
f
1
/
3
h
e
i
g
h
t. Se
c
t
ion
2
.
3
d
e
sc
r
i
be
s
ho
w
to
g
e
t th
e
h
e
i
g
h
t
of h
u
m
a
n
s
h
ape
Fi
g
ure 3.
Human Pose Models
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Coarse
-to-Fine Hum
an
Body Segm
e
n
tation Metho
d
in Video (Yi
ngna
Den
g
)
3523
model. Th
e p
r
oba
bility of head
can
d
idat
e point
r
(
h
) i
s
the propo
rtio
n of pixels i
n
foreg
r
ou
nd a
nd
the rect
angle,
that is:
()
ob
j
re
c
S
rh
S
(3)
Based
on E
q
uation
(3),
ob
ject he
ad
ca
n
d
idate
s
could
be o
b
taine
d
by setting
a t
h
re
shol
d
T
, as shown i
n
Figure 4(c). Howe
ver, the head
ca
ndi
dates with hi
gh probab
iliti
e
s
are not all
true
head
candid
a
t
es, so o
b
ject
edg
e an
d h
e
ad
sho
u
lde
r
matchin
g
i
s
a
p
plied to
elimi
n
ate mo
st of t
he
false candi
da
tes [9].
2.2.2. Edge Bas
e
d He
ad
Shoulder Ma
tching
The h
ead
sh
oulde
r
sha
p
e
is alm
o
st
u
n
ch
ang
ed
while hu
man i
s
walki
n
g,
so a h
ead
sho
u
lde
r
mo
del is
co
nstructed to
eval
uate the h
e
a
d
ca
ndid
a
tes,
as
sho
w
n i
n
Figure 4. Fo
r a
point h
in
ed
ge ima
ge, th
e si
milarity b
e
twee
n it
a
n
d
the
hea
d
should
er
mod
e
l is defin
ed
as
follows
:
00
g
()
(,
,
)
(
)
(
)
m
v
Ph
Gvv
δ
nv
n
v
N
(
4
)
Whe
r
e
v
i
s
t
he ne
are
s
t p
o
int of edg
e i
m
age to th
e
head
sh
ould
e
r
mod
e
l in th
e
dire
ction th
at the
tes
t
line points
to,
v
0
i
s
th
e inter
s
e
c
rion
of te
stline
a
nd h
ead
sho
u
lder
mo
del
contour,
G
(
v
,
v
0
,
δ
)
is a
Ga
ussian
model,
n
0
(
v
)
is th
e no
rmal
dire
ction
of p
o
int
v
,
and
n
m
(
v
)
i
s
the
di
rectio
n of
test
line wh
ere p
o
i
nt
v
is
,
N
is the numb
e
r of
test lines in h
ead shoul
der
model.
The metho
d
of get head candid
a
te
s i
s
shown in Figu
r
e
5.
Figur
e 4. Hea
d
Shoulde
r M
odel
(
a
)
(
b
)
(
c
)
(
f
)
(
e
)
(
d
)
Figure 5. The Process of Getti
ng Head Candidates. (a)
a video
fr
ame, (b) object region, (c)
probability of
head candidate, (d) object
edge image, (e) head shoul
der model, (f) head
candidates
(
b
l
ack
p
oints
)
.
Tes
t
line
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3521 – 35
28
3524
2.3. Human
Shape Mode
l Height Esti
mation
2.3.1. Foot Point Estimati
on
The hei
ght o
f
object in i
m
age i
s
relat
ed to the di
stance
betwee
n
obje
c
t an
d
came
ra
optical
cente
r
, so, the hu
man mod
e
l height
could
be estimate
d from the h
ead point
h
by
estimating th
e foot point under the a
s
su
mption
that all the human
s are all ad
ult, that is:
11
g
fh
fh
x
x
yH
y
(5)
Whe
r
e
H
is t
he hom
ograp
hy matrix betwee
n
the pla
nes
of feet a
nd hea
ds a
r
e
located, a
n
d
it
coul
d be
esti
mated
with le
ast squa
re
s
method [1
0]. (
x
f
,
y
f
) i
s
the
e
s
timated fo
ot point coo
r
din
a
te,
and (
x
h
,
y
h
) i
s
the head p
o
in
t coordi
nate.
2.3.2. Homo
grap
y
Matrix Estimation
Homo
graphy
matrix
con
n
e
ct h
u
man
h
ead f
r
om fo
o
t
points.
Give
n
n
(>=4
) h
e
a
d
poi
nts
and their
co
rresp
ondi
ng fo
ot points
12
(,
)
ii
mm
,the homog
ra
phy
matrix
H
co
ul
d be estimate
d
.
Let
12
3
456
78
1
,,
,,
,,
hh
h
Hh
h
h
hh
, and it
coul
d be writt
en as a vect
or
1
2
34
56
7
8
1
(
,
,,
,,
,
,
,
)
h
h
hh
hh
h
h
h
,
given a
pair of m
a
tch
i
ng p
o
ints
11
1
[,
,
1
]
T
mx
y
and
22
2
[,
,
1
]
T
mx
y
, 2 li
nea
r
eq
uation
s
a
bout
h
co
uld
be
obtaine
d, as
sho
w
n in Equ
a
tion (6
) and
Equation (7).
11
2
1
2
1
2
(
,
,
1
,0
,0
,0
,
,
)
x
yx
x
x
y
h
x
(6)
11
2
1
2
1
2
(
0
,0
,0
,
,
,
1
,
,
)
x
yy
x
y
y
h
y
(7)
There are 8 u
n
kn
own num
bers in
H
, so
4 pairs of matchin
g
point
s are ne
ede
d to get
H
,
besi
d
e
s
, the matchin
g
poi
nts sh
ould n
o
t
be coline
a
r.
2.4. Similarity
bet
w
e
e
n O
b
ject Region
and Shape
Model
Whe
n
shap
e
model
cove
rs obj
ect
regi
on, the
simil
a
rity betwee
n
imag
e
I
an
d sh
ape
model
is defined a
s
Equat
ion (8
).
10
1
0
01
01
()
(|
)
λ
N
λ
N
PI
θ
α
e
(8)
whe
r
e
α
is
a co
nstant i
ndep
e
ndent of
.
N
10
is the
num
b
e
r of
pixels t
hat are in
obj
ect regio
n
but not in
sh
ape mo
del,
N
01
is the num
ber of
pixels
that are i
n
sh
ape mo
del b
u
t not in obj
e
c
t
regio
n
.
λ
10
is a coefficient de
p
ende
nt on the
prob
ability that a pixel is i
n
obje
c
t regi
o
n
but not in
sha
pe m
odel,
and
λ
01
is a
coe
fficient de
pen
dent o
n
the
p
r
oba
bility that a pixel
is no
t in obj
ect
regio
n
but in sha
pe mod
e
l.
2.5. Human
Shape Mode
l Matching
Suppo
se th
e
r
e
are
N
s h
ead
ca
ndid
a
tes
obtain
ed
from
se
ction
2.2, the
tru
e
obj
ect
numbe
r
n
me
et the conditi
on of
n
≤
N
s.
The solut
i
o
n
θ
*
whi
c
h m
a
ximized
the p
r
obability
P
(
I
|
θ
) is
obtaine
d by iteration.
For a he
ad candid
a
te poin
t
h, the pose l in
object mo
del is obtai
ne
d by Equation
(9).
17
:
()
ar
g
m
a
x
(
)
,
Fj
Mj
jj
j
S
Pl
S
lP
l
(9)
Whe
r
e
S
M
is
model a
r
ea,
S
F
is foregrou
nd are
a
cove
red by the model.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Coarse
-to-Fine Hum
an
Body Segm
e
n
tation Metho
d
in Video (Yi
ngna
Den
g
)
3525
The ste
p
s of finding
solutio
n
θ
* whi
c
h m
a
ximized p
o
st
erio
rs a
r
e a
s
follows:
(1) Initialization
,
let
θ
*={
N
s,{
M
1
,
M
2
,…,
M
Ns
}},
M
i
={
h
i,
heig
h
t
i
,
l
i
}, the similarity between
model an
d ob
ject regi
on
P
(
I
|
θ
*) co
uld be
comp
uted through Equ
a
tio
n
(8).
(2)
For all the Ns head candi
d
a
tes, put off the cu
rrent obj
ect
t
(
t
=1,
2
,
…
,
N
s
),
let
*
,
cur
t
θθ
θ
1,
tt
M
, c
o
mpute the s
i
milarity
P
(
I
|
θ
cur
) agai
n.
(3) If
P
(
I
|
θ
cur
) >
P
(
I
|
θ
*), let
θ
*
=
θ
cur
, els
e
,
ke
ep
θ
* un
ch
an
ged, and retu
rn to step (3).
3. Fine Segmenta
tion Based on Lev
e
l Set
3.1.
O
v
er
v
i
e
w
o
f
the
M
e
thod
The contou
r
is initialize
d
based on th
e
coa
r
se area
of each
obje
c
t obtaine
d from initial
segm
entation
,
and the pre
c
ise re
gion i
s
obtained th
rough
cu
rve e
v
olution with
level set, that is
let
N
-1 level sets
to ac
hieve
N
regi
on
s segmentatio
n.
Suppo
se th
e
image i
s
sep
a
rated
into
N
regio
n
s by
N
-1 cu
rve
s
with
out re
gion
overlap, l
e
t
the ith evolu
t
ion curve i
s
expre
s
sed
as
(,
)
,
[
1
,
.
.
.
1
]
ur
i
γ
st
i
N
,whe
re
arc
len
g
th
[0
,
1
]
s
,
t
is
evolution tim
e
, the clo
s
e
curve set expressed
as
1
1
(,
)
:
[
0
,
1
]
ur
N
i
i
γ
st
co
ul
d se
parate th
e image
Ω
into
R
regi
on
s, as sho
w
n i
n
Equation (1
0).
11
2
1
2
1
1
2
1
12
1
,
,
.
..,
...
,
...
,
...
u
u
ru
u
ru
u
ru
u
ru
u
r
uuuu
r
u
u
ru
u
r
uuuu
r
NN
cc
c
c
c
c
NN
γγ
γ
γ
γ
γ
γ
γ
γ
R
R
R
R
R
R
RR
R
R
RR
R
(10
)
Whe
r
e
u
u
r
i
γ
R
and
uu
r
i
c
γ
R
indicate in
sid
e
and
outsi
d
e
re
gion
se
p
a
rately, zero
level set
s
nu
mber
N
-1 i
s
human o
b
je
cts numb
e
r o
b
tained fro
m
ini
t
ial segme
n
ta
tion in this pa
per.
3.2. Realiza
t
i
on of Huma
n
Bod
y
Segmenta
tion
w
i
th Lev
e
l Set
Suppo
se the i
m
age is
2
:
IR
, the energy is defi
ned a
s
:
(,
)
(
,
)
(
)
(
)
rr
r
u
r
RE
C
i
E
γ
uE
γ
uE
γ
E
γ
(11
)
Whe
r
e:
1
12
1
12
11
2
1
2
1
22
2
12
...
...
(
,
)
(
()
)
(
()
)
.
.
.
(
(
)
)
ur
c
cc
c
cc
c
c
γ
γγ
γ
N
γγ
γγ
γ
γ
γ
γ
N
Ri
R
n
RR
R
R
R
RR
R
R
R
R
E
γ
u
λ
Ix
u
d
x
λ
Ix
u
d
x
λ
Ix
u
d
x
(12)
1
1
()
r
r
i
N
E
γ
i
E
γμ
ds
(13
)
1
2
1
()
(
|
|
1
)
2
ur
ur
N
C
ii
i
v
E
γγ
dx
(14
)
Whe
r
e
,
1
,
...
,
1
,
,
1
,
..
.,
ru
r
iR
i
γγ
iN
u
u
iN
is the
pixel average
value,
,
1
,
2
,
...
,
i
λ
iN
is
wei
ght
value, and
μ
is weig
ht value above 0.
So, the huma
n
obje
c
t pre
c
i
s
e segme
n
tation is to find the minimum
energy
(,
)
r
E
γ
u
.
Suppo
se th
e
ze
ro
level
set acco
rdin
g
to t
he
regi
on
en
circled
by
evolution
cu
rve
sets
,1
,
.
.
.
,
1
ur
i
γ
iN
is
{
(
,
)
0,
1
,
2,
.
.
.
,
1
}
i
xy
i
N
, and the level set functi
on is defin
ed
as follo
ws:
(
,
)
0
(
)
(
,
)
0
(
)
(,
)
0
(
)
r
r
r
i
i
i
i
i
i
xy
x
I
n
s
i
d
e
γ
xy
x
γ
x
y
x
O
ut
s
i
de
γ
(15)
Suppo
se H(x) is Heavi
s
ide
functi
on, an
d it is defined a
s
follows:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3521 – 35
28
3526
1,
0
()
0,
0
if
x
Hx
if
x
(16
)
The re
gion in
dicative fun
c
tions a
r
e defin
ed as follo
ws:
1
1
2
2
1
12
1
1
12
1
1
()
[1
(
)
]
(
)
...
...
[
1
(
)
]
r
r
r
rr
r
c
cc
c
N
N
R
R
γ
R
R
γ
R
γ
N
Ri
R
γ
R
γ
R
γ
i
χχ
H
χχ
χ
HH
χχ
χ
χ
H
(17
)
So, the curve
evolution eq
uation coul
d
be expre
s
sed
by Equation (18
)
.
((
)
)
(
,
)
(
)
((
)
)
(
,
)
(
)
..
.
(
(
)
)
(
,
)
(
)
u
rr
r
u
rr
r
u
rr
r
1
1
1
2
2
1
11
11
1
2
2
2
22
22
2
2
2
1
11
11
1
Φ
Φ
Φ
N
R
R
N
NN
RN
N
N
Ix
u
μ
k
γ
xt
v
γ
k
t
Ix
u
μ
k
γ
xt
v
γ
k
t
Ix
u
μ
k
γ
xt
v
γ
k
t
(18)
Whe
r
e
i
R
u
is the
averag
e valu
e of region e
n
c
ircled by curve
i
, and
()
i
x
is d
e
fined a
s
:
11
2
1
1
1
1
(
,
)
0
(,
)
0
(,
)
0
22
1
1
22
(
,
)
0
(,
)
0
(,
)
0
2
(,
)
0
2
(,
)
0
Φ
()
(
(
)
)
()
(
(
)
)
()
()
.
.
.
(
(
)
)
(
)
(
)
..
.
(
)
(
)
+
(
(
)
)
ii
i
i
i
N
i
x
t
xt
xt
iN
N
Ni
iR
x
t
R
x
t
x
t
Rx
t
Rx
t
xI
x
u
χ
xI
x
u
χ
x
χ
x
Ix
u
χ
x
χ
x
χ
x
χ
x
Ix
u
χ
(
,
)
0
(,
)
0
(,
)
0
22
1
(
)
(
)
.
.
.
(
)
(
)
xt
xt
xt
iN
N
x
χ
x
χ
x
χ
x
(19)
Whe
r
e
(,
)
i
xt
χ
sat
i
sf
ie
s
:
if
if
(,
)
0
(,
)
0
((
,
)
)
(
,
)
0
1
(
(,
)
)
(,
)
0
i
i
xt
i
i
xt
i
i
χ
Hx
t
x
t
χ
Hx
t
x
t
(20)
3.3. Cro
w
d
S
e
gmentation
Step
w
i
th L
e
v
e
l Set
The crowd se
gmentation steps are as follows:
(1)
Level set initialization
Suppose there are
N
hum
an shape mo
dels, for each object, the
initial curve is
a circle
which let the center of obje
c
t model as cent
er point, and let one tenth of human shape model a
s
radius. So, th
e ith
level set
is defined as Equation (21):
(2)
Update the le
vel sets based on Equation
(18) sequenti
a
lly.
(3)
If
the evolu
t
io
n is unfinished, return to step 2.
(,
0
)
0
(
,
)
(,
0
)
0
(
,
)
(,
0
)
0
(
,
)
ii
i
ii
i
ii
i
x
i
f
d
x
ce
nte
r
r
a
dius
x
i
f
d
x
c
e
n
t
er
r
a
di
us
x
if
d
x
cen
t
er
r
a
dius
(21)
4. Results and Analy
s
is
In order to te
st the effectiveness of pr
o
posed metho
d
, videos of CAVIAR and Campu
s
are tested
se
parately. Campus i
s
a vid
eo sho
o
t by
the authors th
emselves, as
shown
in Figu
re 6-
7. CAVIAR is
an open video database, and the experim
ental results as shown in Figure 8-9.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Coar
se
-to-
Fine Hum
an
Body Segm
e
n
tation Metho
d
in Video (Yi
ngna
Den
g
)
3527
(a)
(b)
(c)
(d)
Figure 6. Object segmentation re
sult of
Campus1 frame 720#. (a)
shows the original video
frame.
(b) sho
w
s the object models obtai
ned. (c) shows the object region
segmented by
level set. (d)
shows the segmentation re
sult of reference [3]
(a)
(b)
(c)
(d)
Figure 7. Object Segmentation
Result of
Campus1 Fra
m
e 725#.
(a) shows the original video
frame. (b) shows the object
models obtained. (c) sho
w
s the obj
ect region segme
n
ted by
level
set. (d) shows the segmentati
on result o
f
reference [3]
(a)
(b)
(c)
(d)
Figure 8. Object Segment
ation Result of
CAVIAR
Sho
p
Assistant2Cor Fram
e 231
#. (a) shows the
original video
frame. (b) shows the object m
odels. (c) shows the obj
ect region se
gmented by
level set (d) shows the seg
m
entat
ion result of reference [3]
(a)
(b)
(c)
(d)
Figure 9. Object Segment
ation Result of
CAVIAR
Sho
p
Assistant3Cor Fram
e 171
#. (a) shows the
original video
frame, (b) shows the object m
odels, (c) shows the obj
ect region se
gmented by
level set, (d) shows the segment
ation result of reference [3]
In order to tes
t
the
accura
cy
of proposed
method, we
define t
he ac
cura
cy rate as follows:
tur
e
ob
j
N
r
N
(22)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3521 – 35
28
3528
Where
N
ture
is the pixel number of being s
egmented correctly for all objects, and
N
obj
is
the pixel
number of all
objects. Table 1 shows the accura
cy rate.
Table 1. Human Object Segmentation Accuracy Rate
Metho
d
Accurac
y
rate
Prop
osed
meth
o
d
in this
pape
r
85.6
%
Metho
d
in
Ref
e
re
nce[3]
74.2
%
The propo
sed
method coul
d segment
crowd obje
c
t precisely, howe
v
er, when the
color
of
human varie
s
frequently, the method is not suitabl
e as well. So, a method tolerate to color
variation should be researched in the fu
ture.
5. Conclusio
n
When there
a
r
e multiple hu
man objects,
it is
difficult to
segment ea
ch body pre
c
isely.
A
coarse-to-fine
segmentatio
n met
hod is
proposed in t
h
is pape
r. In
coarse segm
entation
step,
a
Bayesian estimation based
object initial
segment
ation was done by
shape model
matching, from
which, human
’s position, height, and posture are
obtained roughly. In
fine segmentation step,
the
precise regio
n
of each object was obtai
ned through
curve evolution with level
set. Experime
n
tal
results
show that the proposed meth
od could
se
gment cro
w
d
object pre
c
isely. For curve
evolution of e
a
ch obje
c
t, it
is expressed
by onl
y one level set, so when there
a
r
e some
different
colors in obje
c
t cloth, the segmentat
ion result is not sa
tisfied, so,
mu
ltiple level sets for one object
could be considered in the future.
Ackno
w
l
e
dg
ements
Project supp
orted by the
ShaanXi Province
Education Departme
nt Project (11JK0990)
and National Science Foun
dation of Chin
a (61100165
)
Referenc
es
[1]
R Ya
ng, Q Z
h
en
g.
Mu
lti-
movi
ng pe
op
le det
e
c
tio
n
fro
m
bi
no
cul
a
r
s
e
q
u
e
n
c
e
s.
Intern
ati
o
n
a
l
Conf
er
enc
e
on Ac
o
u
stic, S
p
e
e
c
h
a
nd S
i
g
n
a
l
Pro
c
ess
i
ng.
Hon
g
K
o
n
g
.2
0
03 V
o
l.
2:2
97-
3
00.
[2]
JG Brosto
w
,
R Cipolla.
Un
sup
e
r
vi
se
d
Ba
yesi
an
De
te
cti
on o
f
In
dep
en
den
t Mo
tion
in
Cro
w
d
s
.
IE
EE
Com
put
er S
o
c
i
et
y C
o
nfer
enc
e
o
n
C
o
m
put
er
Visi
o
n
a
n
d
P
a
ttern
R
e
co
g
n
iti
o
n. N
e
w
Yo
rk.
2
0
0
6
;
1: 5
9
4
-
60
1.
[3]
AM El
gamm
a
l,
LS D
a
v
i
s.
Pr
ob
ab
il
istic fr
a
m
ew
ork for
se
g
m
ent
in
g p
e
o
p
l
e u
n
d
e
r Occ
l
u
s
io
n.
T
he 8t
h
IEEE Conference on Computer Vision.
Vancouver. 2001;
2:145-
152.
[4]
D R
a
m
a
n
an,
DA F
o
rs
yth.
Fi
nd
in
g an
d tracki
ng
pe
op
le
from
th
e bo
ttom
up
.
IEEE
C
onf
ere
n
c
e
o
n
Com
put
er Vi
si
o
n
a
nd P
a
tter
n
Rec
o
g
n
it
io
n. M
a
d
i
so
n. 2
0
0
3
; 2
:
467
-4
74.
[5]
Z
Li
n, SD
Lar
r
y
,
D Doermann, et.al.
An in
te
ra
cti
v
e
ap
p
r
oa
ch
to
po
se
-assi
s
ted
and
ap
pe
a
r
a
n
ce
-b
ased
se
gmen
ta
tion
of h
u
ma
n
s
.
W
o
r
ksho
p
on Int
e
ra
ctive
Com
put
er
Visi
on.
Ri
o d
e
Jan
e
i
r
o. Br
azi
l
.
20
07,
1-8
[6]
W
u
Bo, Ram
Nev
a
ti
a. Det
e
c
t
ion
an
d track
i
ng of mu
lti
p
l
e
,
parti
al
l
y
occ
l
u
d
e
d
hu
ma
ns b
y
B
a
yes
i
a
n
comb
i
nati
o
n of
ed
ge
let
b
a
se
d
part
det
ecto
rs
.
Intern
ati
o
na
l Jour
n
a
l of
C
o
mput
er
V
i
si
on.
20
07;
27
(2):
2
4
7
-
266
[7]
B Sapp, A T
o
shev, B T
a
skar.
C
a
scad
ed
mode
l
s
fo
r
a
r
ti
cu
la
te
d
po
se
e
s
timati
o
n
. E
u
ro
p
e
a
n
Co
nfer
en
ce
on C
o
m
put
er V
i
si
on. H
e
r
a
kl
io
n
.
201
0; 2:
40
6.
[8]
H L
u
, G F
ang,
X S
h
a
o
,
X L
i
. S
egm
ent
in
g h
u
m
an fr
om p
hot
o i
m
ag
es b
a
s
ed
o
n
a c
oar
se
–to-f
i
ne sc
h
e
me.
IEEE T
r
ansact
i
ons
on s
y
stem
s, man,
and c
y
ber
net
ions
–P
ar
t B: Cy
bere
ntic
s. 2012; 42(
3
):
889-899.
[9]
Zhao T
ao, Ram Nev
a
tia. B
a
yesian
hum
an s
egm
ent
at
ion in
cro
w
ded sit
uat
ions. IEEE Com
put
er Soc
i
et
y
Co
nfer
en
ce
on
Com
put
er Vi
si
o
n
a
nd P
a
tter
n
Rec
o
g
n
it
io
n.M
a
d
i
so
n. 2
0
0
3
; 2
:
459
-4
66.
[10]
Z
he LS Dav
i
s,
D
D
o
e
rma
nn, D
D
e
me
nth
o
n
.
H
i
e
r
archi
c
a
l
pa
rt-temp
l
a
t
e
ma
tchi
ng
for
huma
n
de
te
cti
on
an
d se
g
m
ent
ati
on.
T
he 1
1
th I
n
tern
ati
o
n
a
l
Co
n
f
ere
n
ce
on
Co
mput
er Vi
si
on.
Ri
o d
e
Ja
ne
ir
o. 20
07; 1
-
8.
[11]
AR M
ans
o
u
ri,
A Mit
i
ch
e,
C
Vaz
q
u
e
z. M
u
lti
r
e
g
i
o
n
com
petit
io
n: A
l
e
v
e
l s
e
t
e
x
te
ns
i
on
of r
e
g
i
o
n
comp
etit
io
n to
mult
ip
le r
e
g
i
o
n
ima
ge
pa
rtiti
o
ni
ng.
Co
mput
e
r
Visi
o
n
a
nd I
m
ag
e U
n
d
e
rst
a
n
d
i
n
g
. 2
006
;
10
1(3
)
:1
37-
15
0
.
Evaluation Warning : The document was created with Spire.PDF for Python.