TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.6, Jun
e
201
4, pp. 4274 ~ 4
2
8
2
DOI: 10.115
9
1
/telkomni
ka.
v
12i6.512
0
4274
Re
cei
v
ed
No
vem
ber 1
1
, 2013; Re
vi
sed
De
cem
ber 3
1
,
2013; Accep
t
ed Jan
uary 2
4
, 2014
Resear
ch of Background Segmentation Method in
Sports Video
Shen Li*
1
, Hou Lihong
2
1
Harbin Me
dic
a
l Univ
ersit
y
, H
a
rbi
n
163
31
6 Chin
a
2
Beihu
a
Univ
er
sit
y
, Jili
n 13
20
13, Chi
n
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: shenl
i_hm
u
@
16
3.com
A
b
st
r
a
ct
Hu
ma
n
moti
o
n
i
m
a
ge se
g
m
entatio
n is th
e
first step
in th
e
process
of hu
ma
n
moti
on
an
alysis; it is
low
-
level
proce
ssing p
a
rt of the visua
l
an
alysi
s of
human
mo
vement. T
he pr
ocessi
ng res
u
lt
s of the stage
of
qua
lity directly
affects the progress of
the fol
l
ow
-up w
o
rk, the seg
m
e
n
tatio
n
result has e
n
ormous i
n
flue
n
c
e
on th
e fin
a
l r
e
sults h
u
man
o
f
mov
e
me
nt a
nalysis.
Hu
ma
n
motio
n
i
m
ag
e se
gmentati
o
n rese
arch
is
an
importa
nt g
oal
similar
to th
e v
i
sual
perc
epti
o
n
of th
e hum
an
, wh
i
c
h
ma
ke the
comp
u
t
e
r
fe
el
th
e
m
o
veme
nt
of hu
ma
n bo
d
y
, and
make
computers e
a
s
i
er to un
derst
a
nd the fe
eli
n
g
s
of sports be
havi
o
r. Devi
ati
o
n
infor
m
ati
on ba
sed on
Ga
us
sian mixture
mo
de
l
to jo
i
n
the c
h
ro
ma
and
bri
ghtn
e
s
s
of b
a
ckgro
u
nd
seg
m
e
n
tatio
n
meth
od is
not
only su
itab
le for ordi
nary st
a
t
ic scene, but
also it is su
ita
b
le for the s
p
e
c
ia
l
envir
on
me
nt such as reflecti
ve of
ice, fu
zzy shadow
, mo
ving o
b
ject refl
ection. Accord
i
ng to exp
e
ri
ment
s
results, it is show
n that the pr
opos
ed a
l
gor
ithm h
a
s strong
robustn
ess.
Ke
y
w
ords
: image se
g
m
e
n
tation, motio
n
an
alysis, sports vi
deo,
metho
d
, researc
h
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
As
kno
w
n
tha
t
amon
g the
e
n
vironm
ental
informat
io
n, h
e
ld ve
ry la
rge
propo
rtion
of
visual
informatio
n, while the dy
namic vi
sual
information
is the mai
n
com
pon
en
t. And a lot of
meanin
g
ful visual info
rmati
on is involve
d
in t
he movement. The rese
arch of b
a
ckgroun
d of the
human b
ody movement im
age segme
n
tation is to ma
ke the co
mp
uter have the
function
s whi
c
h
are
simil
a
r to
the vi
sual
p
e
rception
of t
he p
e
rs
on,
the
com
puter ca
n fe
el the
view within
the
human b
ody target, and
ca
n unde
rsta
nd
people'
s spo
r
ts be
havior o
f
the compute
r
, and al
so ca
n
tak
e
the appropriate ac
tion.
At the same time, the backgroun
d se
g
m
ent
ation m
e
thod not onl
y can be u
s
e
d
in the
human b
ody
target, but
also it can
be widely u
s
ed in other
area
s, su
ch
as lice
n
se pl
ate
recognitio
n
, video
surveill
a
n
ce, et
c. But due to the
na
ture of the
al
gorithm
s, the
y
are limited
to
the ap
plicatio
n envi
r
onm
e
n
t, at the
cu
rrent
re
sea
r
ch
statu
s
and
tech
nolo
g
y le
vel, we
still
can't
find out a ge
n
e
ral al
gorith
m
, which is a
p
p
licabl
e to
any
external e
n
vironment
(lighti
ng
condition
s,
cameras, anti
-
interf
erence
ability, etc.) and
applicaple to all kinds
of motion segmentation under
compl
e
x ba
ckgroun
d, thus in orde
r to solve so
m
e
limited co
nditi
ons o
r
some
appli
c
ation u
nder
the backg
rou
nd of the mo
vement of the image
b
a
ckgrou
nd segm
entation probl
em is still be
the
main topics
of the res
e
arc
h
in this
field [1].
Huma
n motio
n
image
seg
m
entation is
an indi
spe
n
sable pa
rt of human motio
n
analysi
s
,
and it plays a
very importa
nt in kinem
atic analy
s
is. B
a
se
d on this,
a large
numb
e
r of schol
ars in
the field h
a
ve don
e a l
o
t
of usef
ul work. The
comm
only se
gme
n
tation alg
o
rith
m incl
ude
s t
he
following sev
e
ral ways:
1. The finite difference time
domain
2. The ba
ckg
round differen
c
e
3. The se
gme
n
tation metho
d
based on b
a
ckgroun
d m
odel
4. The metho
d
based on
statistical mo
de
l
5. Other meth
od
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Re
sea
r
ch of Backgroun
d Segm
entation Method in
Sports Vide
o (Shen Li
)
4275
Figure 1. The Ba
ckgroun
d Sub
s
tra
c
tion Result of Movement Emulator
There a
r
e ma
ny other m
e
thod
s in the
m
o
tion
dete
c
tio
n
.
Such as m
o
tion
vecto
r
method,
it is suitable
for multidimensional
changing
circum
st
ances, a
n
d
can elimin
ate
the ba
ckgro
u
n
d
pixels, an
d
can ma
ke
obje
c
t in
ce
rtain d
i
rectio
n of m
o
tion more p
r
o
m
inent, but th
e motion ve
ct
or
method
can
n
o
t accurately segm
ent the obje
c
t. the
EM algorith
m can e
s
tabl
ish cl
assification
Gau
ssi
an mi
xture mod
e
l
for ea
ch
pixel, the mod
e
l
can
be a
u
tomatically u
p
d
ated, an
d can
adaptively to cla
ssify ea
ch
pixel as the b
a
ckgr
oun
d, the sha
d
o
w
or
spo
r
ts outlo
o
k
[2-5].
Although
ma
ny schol
ars
h
a
ve cond
ucte
d a l
a
rge
nu
mber of u
s
ef
ul research,
b
e
ca
use of
the com
p
lexity and irre
gul
arity of huma
n
movement,
make it difficult to study throu
g
h a u
n
ified
way. Many
method
s a
n
d
model
s a
r
e
too sim
p
le
and n
o
t wid
e
ly use
d
, or too comple
x to
appli
c
ation
in
pra
c
tice. Th
e main
existin
g
proble
m
s
of ba
ckgro
u
n
d
of h
u
man
motion im
ag
e
segm
entation
technol
ogy rese
arch at
prese
n
t are a
s
follows [6]:
a. There a
r
e
no ge
neral
method
s in t
he
ba
ckgrou
nd of the h
u
m
an bo
dy m
o
vement
image
s
segm
entation, it is
usu
a
lly
ca
rri
e
d
out un
de
r the conditio
n
of spe
c
ific en
vironme
n
tal, and
so
it shoul
d desi
gn corre
s
po
ndin
g
se
gmentation
method.
Su
ch
as du
ring the
moveme
nt
of
human, it is n
o
t obscured, backg
rou
nd i
s
rela
tively s
i
mple, s
t
ationary c
a
mera, etc
.
b. The
segm
entation of h
u
man m
o
ve
ment, und
er t
he complex
b
a
ckgroun
d, d
ue to the
influen
ce on
surro
undi
ng
environ
ment,
and it ofte
n
powe
rle
s
s to outsid
e
int
e
rferen
ce. When
pede
stria
n
s wea
r
clothe
s
whi
c
h col
o
r
is simila
r
t
o
the
b
a
ckg
r
oun
d
colo
r or cha
nge
s of
backg
rou
nd l
i
ght is big, it
is difficult to
s di
scern the moving o
f
the human
body from t
he
backg
rou
nd.
c. The si
ze
o
f
human bo
d
y
contou
r, the colo
r of th
e clothe
s a
n
d
the texture
with th
e
dre
ss a
nd a
ppea
ran
c
e
chara
c
te
risti
c
s, such a
s
the weathe
r ch
ange
s, and chang
es of other
external conditions, it has
very strong variability.
d. The singl
e image proce
s
sing is
difficult
to achieve reliab
l
e segm
enta
t
ion of
backg
rou
nd, and the
r
efore
it must
be a
dopted
with seque
nce ima
ge
processin
g
, but it need
s to
store
an
d p
r
o
c
e
s
s large
a
m
ount of
dat
a, and
re
al-t
i
m
e pe
rforma
nce
of the
sy
stem i
s
difficult to
guarantee.
e. The
ba
ckgrou
nd
of th
e hum
an
bo
dy moveme
n
t
image
s
seg
m
entation te
chn
o
logy
involves the
kno
w
le
dge
of many
subj
ects, incl
udi
ng
comp
uter visi
on, imag
e p
r
oce
s
sing, p
a
ttern
recognitio
n
, a
r
tificial intellig
ence, how to
combi
ne
the
m
is the re
se
arch qu
estio
n
.
In the curre
n
t
resea
r
ch stat
us a
nd the t
e
ch
nical leve
l, the
reali
z
at
ion of ge
neral ba
ckgro
u
n
d
se
gmentati
on
sy
st
em i
s
n
o
t
realit
y
,
so
how t
o
sol
v
e som
e
lim
ited co
nditio
ns u
nde
r th
e ba
ckgro
u
n
d
of
appli
c
ation
b
a
ckgroun
d se
gmentation
p
r
oble
m
will still
be th
e m
a
in
topics
of the
resea
r
ch in
th
is
field.
2. Segmenta
tion Algorith
m Resea
rch
In the study, the main co
n
t
ent of the bac
kgro
und of the huma
n
bo
dy movement
image
segm
entation
algorith
m
s,
inclu
d
ing int
e
rframe
difference an
d ba
ckgro
und
differen
c
e, o
n
lin
e
singl
e G
a
u
ssi
an m
odel
an
d meth
od
of the G
a
u
ssi
an
mixture m
o
d
e
l a
r
e i
n
trod
u
c
ed.
The
onli
ne
singl
e Ga
ussian mo
del a
nd the
ba
sic idea
of the
Gau
s
sian
m
i
xture mod
e
l
and
re
alizat
ion
method a
r
e also the p
r
e
s
ented. After analysi
s
of
algorithm an
d the experime
n
tal results, the
con
c
lu
sio
n
s a
r
e obtain
ed th
roug
h co
mpa
r
iso
n
.
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: 4274 – 4
282
4276
Huma
n moti
on image b
a
c
kgro
und
se
gmentation i
n
recent yea
r
s ha
s be
en
got more
attention an
d
it is one
of the forefront o
f
the di
re
ction
,
it is the com
b
ination
of m
odern biol
ogi
cal
mech
ani
cs a
nd com
puter vision tech
nology, it
has a very broad a
nd im
portant field
of
appli
c
ation of
it in the smart su
rveillan
c
e, hum
a
n
compute
r
interaction, motio
n
analysi
s
a
nd
virtual reality, and other fiel
ds.
Although in t
he pa
st ten
years, p
eopl
e hav
e d
one
a lot of wo
rk
on the p
r
oblem;
however, so
far, there
is no effe
ctive algor
ithm
can b
e
ap
plied to mo
vement ima
ge
segm
entation
unde
r vari
ou
s envi
r
onm
e
n
ts. The
re
se
arch
work of
the hum
an
body movem
ent
image ba
ckg
r
ound
segm
en
tation is prop
ose
d
for so
m
e
spe
c
ific a
p
p
licatio
n pro
b
l
em, the further
study of algorithm remain
s
to be need
ed
[7].
In the pra
c
tical appli
c
ation
system
and
inte
re
sted ta
rget motio
n
i
n
the sce
ne
can
be
divided into four combin
ations, the first
combin
atio
n
is actually a
static scene
with static o
f
came
ra
-targe
t, the proce
s
sing meth
od is the stat
ic i
m
age processing meth
od,
which i
s
wid
e
ly
applie
d in face recognitio
n
, iris re
co
gniti
on, the
se
con
d
is the co
mb
ination of stat
ionary came
ra
and moving t
a
rget, this is a very important dy
nami
c
scen
es, the pro
c
e
ss g
e
nerally in
clud
es
moving targ
et detection,
cl
assificatio
n
, tracking
a
nd b
ehavior
und
erstandi
ng,
it is
mainly used f
o
r
early warni
n
g
,
secu
rity mo
nitoring
event
s. The
third is the combin
ation of moving ca
mera a
nd
static targ
et, it is mainly used in robot
vision
navigation, electroni
c maps ge
ne
rat
e
d
automatically, and 3
D
sce
ne un
derst
an
ding, etc. th
e fourth
com
b
ination i
s
with the movin
g
came
ra
and
moving ta
rget, it can
be u
s
ed
in
moving targ
et detectio
n
in the
com
p
lex
circum
stan
ce
, such as m
o
nitoring
and
control
system
in satellite o
r
on the plan
e
.
In the study, i
t
mainly di
scu
s
ses th
e targ
et movement
unde
r the
condition
of st
ationary
cam
e
ra
and
ho
w to
reali
z
e the re
sea
r
che
s
on the divisio
n
of human m
o
ve
ment on the p
a
rticul
ar
scen
e [8].
The proba
bility density model an
d onl
ine sin
g
le G
aussia
n
ba
ckgroun
d mo
del and
Gau
ssi
an mi
xture mod
e
ls are i
n
trod
uced. Interfa
c
e
d
differe
nce
method i
s
through th
e current
image
with a
d
jacent fram
e differen
c
e t
o
get mo
ving
target p
r
o
s
p
e
cts. Ba
ckground
differen
c
e
method is u
s
ually sele
cted
empty sce
ne
image as a b
a
ckgroun
d im
age of movin
g
target
s.
Backgroun
d
model
metho
d
is the
key
to t
he ba
ckground
ima
ge descri
p
tion
m
odel of
backg
rou
nd
model; it is the basi
s
of the
backgroun
d motion se
gm
entation prospect
s
.
Backgroun
d model mai
n
ly includ
es
sin
g
le-m
ode
sta
t
e and multi
m
odal two ki
nds, the
former mod
e
l
each ba
ckg
r
oun
d pixel
colo
r di
stri
b
u
t
ion is
con
c
e
n
trated, it ca
n use a
sing
le
distrib
u
tion
p
r
oba
bility mo
del to
de
scri
be the
di
stri
bution,
while
the latte
r
model
is mo
re
disp
ersed, an
d need to u
s
e
more di
stribu
tion prob
abilit
y model to describe.
In many appli
c
ation
scena
rios, such as
ripple
s
on the
surfa
c
e
of the
water, the
swaying
bran
ch
es
an
d waving fla
g
s
, pixel value
s
of them p
r
ese
n
ts the m
u
ltimodal feat
ure
s
. The m
o
st
comm
only used pro
bability
density mod
e
l in describ
i
ng the scene
backgroun
d colo
r distri
but
ion
prob
ability density functio
n
is Gau
s
sia
n
distri
b
u
tion
, [9], as describ
ed in bel
ow onli
ne si
ngle
Gau
ssi
an ba
ckgroun
d mod
e
l and Ga
ussi
an mixtur
e m
odel are both
belon
g to the model.
3. Impro
v
ed
GMM Ba
ckgr
ound Segme
n
ta
tion Algo
rithm
There is no g
eneral metho
d
on backg
ro
und se
gme
n
tation of the human bo
dy movement
image
s, it is usually
carri
ed o
u
t un
der the lim
ited
con
d
ition, a
n
d greatly infl
uen
ced
by t
h
e
surro
undi
ng
environ
ment,
usually it is powe
r
le
ss to interferen
ce, and only
can d
e
si
gn t
h
e
corre
s
p
ondin
g
segm
entati
on method fo
r a parti
cula
r environ
ment.
In
the study, the
ro
ad sp
e
ed skating
tra
i
ni
ng
gro
und
s und
er th
e
scenari
o
of th
e
human
body after th
e
cha
r
a
c
ter
mo
tion obje
c
t de
tection a
nd
segmentatio
n, on the b
a
si
s
of the cla
s
sic
of
the Gau
s
sian
mixture mo
d
e
l, brig
htne
ss deviati
on a
n
d
the info
rma
t
ion su
ch
a
s
colo
r d
e
viation
are im
proved
, and the
ne
w imp
r
oved
algorith
m
can
be a
pplied
to the
spe
c
ial
scene,
sh
ad
ow,
reflectio
n
and
glare an
d it can get goo
d segmentatio
n effect [10-1
3
].
3.1. Chara
c
teristic Analy
s
is of the Ice
En
v
i
ronment
Und
e
r compl
e
x scen
e
ba
ckgroun
d se
g
m
entation,
the important p
r
erequi
site is
to know
the cha
r
a
c
teristics of the scene.
Only
in this way, it can elimin
a
t
e or red
u
ce
the impact
of
backg
rou
nd
segm
entation
of stat
ic st
ate and oth
e
r facto
r
s. T
herefo
r
e, alt
houg
h many
of
seq
uen
ce
im
age
of movin
g
target d
e
te
ction
are
stu
d
ied und
er
st
atic backg
ro
u
nd, but as the
compl
e
x ba
ckgroun
d
scen
e segme
n
tation, it is
ne
ce
ssary to
co
nsider th
e dyn
a
m
ic b
a
ckg
r
ou
nd.
The video is taken by the mono
cula
r
came
ra
wh
i
c
h is the train
i
ng video of skater
with h
i
gh
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Re
sea
r
ch of Backgroun
d Segm
entation Method in
Sports Vide
o (Shen Li
)
4277
spe
ed in the
static ro
ad, n
a
me
ly that the came
ra p
o
sition on
the ro
ad outsi
de be
nd alignm
ent of
spe
ed
skatin
g athlete
s
. T
h
rou
gh th
e
compa
r
ison
of
the ima
ge vi
deo
of differe
nt enviro
n
me
nts,
due to the chara
c
te
risti
c
of the ice ro
ad, this
sce
n
a
rio h
a
s the
followin
g
feature
s
re
stri
cts the
segm
entation
effect:
1. The change of bac
kground illumination
2. The static
obje
c
ts in the
backgroun
d
3. The sh
ade
d area
4. Reflectio
n
area
5. Strong refl
ective ice
6. The interfe
r
en
ce of ice a
r
ea
7. Other noi
se
S
t
r
ong r
e
f
l
e
c
t
i
v
e
R
e
f
l
e
c
t
e
d
i
m
a
g
e
S
h
a
dow
Figure 2. Col
o
r Image in th
e Environme
n
t of Skating Rin
k
In this enviro
n
ment, thro
u
gh the ba
ckg
r
oun
d segme
n
tation algo
rit
h
ms, it ca
n segment
the moving target
s, remo
ve the static backgr
oun
d, but the segmentation result
s are n
o
t
s
a
tis
f
ac
tory, s
h
adow, reflec
tion and reflec
tive r
egi
on
have not got the corre
c
t pro
c
e
ssi
ng, they
seri
ou
sly affect segm
entati
on. Based o
n
the parti
cul
a
rity, scenari
o
, it makes the i
m
provem
ent on
the existin
g
a
l
gorithm
s. T
h
roug
h th
e
stu
d
y, it ca
n
b
e
found th
at at
hletes on
i
c
e
refle
c
tion
a
r
ea
,
the col
o
r valu
e ch
ang
e is b
i
gger th
an th
e ice
colo
r va
lues, b
u
t it has sm
alle
r values
of the col
o
r
deviation. Th
e athlete'
s
sh
ado
w re
gion
colo
r is dee
p
e
r tha
n
the
surface
colo
r, the brig
htne
ss of
the two large
r
deviation d
egre
e
s. So
we can
add th
e two ki
nd of
information t
o
the algo
rith
m
based o
n
mix
t
ure G
a
u
ssi
a
n
mod
e
l to di
stingui
sh
sh
a
ded a
r
ea,
refl
ection
area, a
n
d thu
s
imp
r
o
v
e
the segm
enta
t
ion as sho
w
n
in Figure 2.
3.2. Relate
d Defini
tions
The m
odel
is propo
sed
un
der the
RGB
col
o
r s
pace, as sho
w
n i
n
Figu
re
3, a
s
to the
singl
e pixel in
the image I,
[(
)
,
(
)
,
(
)
]
ii
i
i
RG
B
rep
r
e
s
ent p
r
i
m
e sp
ot of b
a
ck In RGB space
.
[(
)
,
(
)
,
(
)
]
ii
i
i
XX
R
X
G
X
B
rep
r
e
s
ent va
lue of the cu
rre
nt image
pixels.
i
X
and
i
r
e
pr
es
e
n
t
s
brightn
e
ss d
e
v
iation and
colors deviatio
n
. The lin
ear
pass th
rou
gh
the ori
g
in a
n
d
the line
i
o
is
calle
d chrom
a
line,
colo
r
deviation
i
CD
is t
he minim
u
m
distan
ce
from
point
i
X
t
o
c
h
r
o
m
a
l
i
n
e
i
o
.
Figure 3. Mathematical Mo
del of Brightn
e
ss Di
st
ortion
and Ch
rom
a
ticity Distortio
n
in the RGB
C
o
lor
Sp
ac
e
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: 4274 – 4
282
4278
Brightne
ss deviation BD is a scalar va
lue
，
we defin
e
2
()
ii
i
BD
X
B
D
, when the
value of
()
i
BD
is m
i
nimum, the
i
BD
is the
deviatio
n
of the
colo
r, when
the b
a
c
kgro
und
pixel is
same
as th
e
current pixe
l,
1
i
BD
when b
a
ckgroun
d pixe
l is less b
r
ig
ht than cu
rre
n
t pixel
1
i
BD
, when ba
ckg
r
oun
d pixel is bright than current pixel,
1
i
BD
.
Colo
r deviati
on CD is defi
ned t
he di
sta
n
ce from the
point
i
X
to a straight line
i
o
,tha
t
is the distan
ce betwe
en th
e point
i
X
and
i
o
,
Colo
r deviatio
n
is given by formul
a (1
):
ii
i
i
CD
X
B
D
(1)
Color deviation
i
CD
and the Bri
ghtne
ss d
e
viation BD ca
n be cal
c
ul
ated
through:
22
2
m
i
n
(
()
()
(
(
)
(
)
(
)
(
)
)
i
i
ii
i
i
i
i
ii
BD
X
R
BD
R
X
G
B
D
G
X
B
BD
B
(2)
22
2
m
i
n
(
()
()
(
(
)
(
)
(
)
(
)
)
i
i
ii
i
i
i
i
ii
CD
X
R
BD
R
X
G
B
D
G
X
B
BD
B
(3)
The followi
ng
two ca
se
s ch
roma d
e
viatio
n (3-4) in
stea
d of using the
following formula:
a) If you
m
i
x, die type
B point
s
h
i
gher value
for the
bla
c
k
colo
r of
the
cloth
()
(
)
()
0
ii
i
RG
B
as
in
f
i
BD
in orde
r to avoid this p
r
oble
m
()
(
)
()
0
ii
i
RG
B
22
2
((
)
(
)
(
)
ii
i
i
C
D
XR
X
G
XB
(5)
b) When th
e moving targ
et
s in the
cu
rre
nt image pixe
l is clo
s
e to t
he ori
g
in, it always i
s
divided i
n
to b
a
ckgroun
d pi
xels. Be
cau
s
e all
of the
co
lor lin
e p
a
ss t
h
rou
gh th
e o
r
igin, cl
ose to
the
origin
of pixels is
close to any a col
o
r
line,
through
the color deviati
on it will
be wrong pixel
points. In ord
e
r to avoid th
is problem, the line
color
deviation an
d
the deviation
brightne
ss can
validate pixel
deviation of
the si
ze
of the ch
ro
m
a
ticit
y
and b
r
ightn
e
ss. Among
them, un
der the
environ
ment
of an ice ri
nk, the ch
roma
deviation va
l
ue can
refle
c
t the variatio
n of ch
rom
a
ticity
on i
c
e, th
e a
t
hletes'
refle
c
tion region
o
n
the
ic
e u
s
u
a
lly ha
s
smal
ler val
ues of
the pixel
col
o
r
deviation, the
intro
d
u
c
tion
of ch
rom
a
tic
deviation
co
mpared to
reflection
re
acti
on
can
inhi
bit th
e
light cha
nge
s. Shadow i
s
gray, deg
ree
of bright
n
e
ss is lo
w, thro
ugh calculati
ng it can
det
ect
wheth
e
r
cha
nge
regio
n
i
s
the
sha
d
o
w
. The
r
efo
r
e
,
base
d
on
this ide
a
, se
gmentation
of
chromati
city and b
r
ightn
e
ss deviatio
n
u
nder t
he h
u
m
an bo
dy imag
e ba
ckground
enviro
n
ment
is
good.
3.3. Gaussia
n
Mixture M
odel Algorithm
Thro
ugh
a l
o
t of experi
m
ents, it is
sh
o
w
n th
at among th
e
nume
r
ou
s b
a
ckgroun
d
segm
entation
method
s, the
effect of the
origin
al
meth
od of the G
a
u
ssi
an mixture
model i
s
go
o
d
,
the moving t
a
rget in th
e video
can b
e
detecte
d,
but
there exi
s
ts
some i
n
terfe
r
ence re
gion.
The
sha
ded
area
and th
e refle
c
tion
are
a
a
r
e moving
wit
h
the m
o
ving
target, it
ca
n be
mista
k
e
as
pro
s
pe
ct g
oal
s, an
d al
so
can influ
e
n
c
e t
he effe
ct of
d
e
tection. At t
he
same
time
, the reflectio
n
of
the ice i
s
st
rong than th
e
comm
on sit
uation.
Th
rou
gh the b
r
ight
ness d
e
viation and d
e
viation
informatio
n
can
re
move
ch
rom
a
ticity sh
ado
ws
, reflection
s an
d
supp
re
ss
relate
d strong
reflectio
n
. We put the bri
ghtne
ss d
e
viation of
the
Gau
ssi
an mi
xture model
and colo
r de
viation
informatio
n in
the o
r
igin
al a
l
gorithm, th
us the o
r
ig
in
al a
l
gorithm
of th
e Ga
ussia
n
mixture mo
de
l is
improve
d
.
Und
e
r the
RGB spa
c
e, it
prop
ose Ga
u
ssi
an mi
xture
model for
ea
ch pixel in th
e
image of
the thre
e col
o
r
cha
nnel
s
R, G, B re
sp
ectively,
and three col
o
r chann
els of
RGB
of
ea
ch pixel
can
be
cal
c
ul
ated, the B d
i
stributio
n of t
he Ga
ussia
n
mixture mo
d
e
l
R
B
G
B
B
B
also cab be
set
up. Then through the
s
e B
distrib
u
tions,
the mean pa
ramete
rs
of
R
B
G
B
B
B
can be
cal
c
ulated
according
to
the d
e
finitio
n
an
d
cal
c
ul
ation formula
of bri
ghtne
ss d
e
viation
i
BD
and
color
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Re
sea
r
ch of Backgroun
d Segm
entation Method in
Sports Vide
o (Shen Li
)
4279
deviation
i
CD
. Finally, based
o
n
the
cu
rrent
B distri
bution
of
pixels an
d
the mat
c
hing
situation
of
R
B
G
B
B
B
, chro
ma d
e
viation, devi
a
tion of bri
g
htnes
s info
rmation, ea
ch
pixel point
can
be
cla
ssif
i
e
d
as f
o
llow
s
:
a. If in the color spa
c
e mo
del RGB, Th
e value of the cu
rre
nt pixel is not sa
m
e
as the B
distrib
u
tion (
R
B
G
B
B
B
) of the three
mixture mod
e
ls of the
on
e pixel togeth
e
r, then it is
divided
into as the ba
ckgro
und p
o
i
n
t.
b. If In the co
lor
spa
c
e
mo
del RGB, Th
e value
of the
cu
rrent pixel
is n
o
t sa
me a
s
the B
distrib
u
tion (
R
B
G
B
B
B
) of the th
re
e mixture m
o
dels
of the o
ne pixel tog
e
t
her, value
of colo
r
deviation CDi
>Th_
CD, the
n
it is divided as the target area p
a
rt.
c. If in the space of the
RGB colo
r m
odel, t
he current pixel col
o
r value
s
is n
o
t match
with B
di
stri
bution
(
R
B
G
B
B
B
) of
three
mixed
model to
geth
er,
col
o
r
de
gree
value
of
pa
rtial
differential CD < Th
CD, a
nd the brig
htn
e
ss
BD<T
h_
BD, It is taken as as the
shado
w
area.
d. If in the
space of
the
RGB
col
o
r m
odel, th
e
current pixel
col
o
r valu
es is n
o
t match
with B
di
stri
bution
(
R
B
G
B
B
B
) of
three
mixed
model to
geth
er,
col
o
r
de
gree
value
of
pa
rtial
differential CD <
T
h
CD, and
th
e
b
r
igh
t
ness
B
D
>T
h_BD, It i
s
ta
ken
a
s
as th
e up
sid
e
sha
dow
area.
Among them,
Th _CD represe
n
ts deviat
i
on thre
sh
old
and Th
_BD repre
s
e
n
ts b
r
i
ghtne
ss
deviation th
re
shol
d. The
p
o
st-p
ro
ce
ssin
g to re
move
discrete
noi
s
e and
the n
o
i
s
e of th
e target
area
whi
c
h is too small, and r
eali
z
e foreground area t
r
eatment, f
illi
ng holes, and other functions
are
reali
z
ed
throug
h m
a
thematical
morp
hol
o
g
y operators ex
pan
sion,
corrosi
o
n, op
en
ing
operation a
n
d
clo
s
in
g op
e
r
ation, et
c. In ord
e
r
to
achi
eve these fun
c
tion
s an
d im
prove
unde
r t
h
e
con
d
ition of complex move
ment image
detectio
n
a
ccura
cy, and re
duce the cost of the corre
c
t
segm
entation
,
increa
se
th
e effectiven
e
s
s of th
e
foll
ow-up
work.
Figure 4
an
d
Figu
re
5
are
the
segm
entation
result
s of this algo
rithm. We defin
e limitation value
al
o
, when
the
mi
n
i
BD
,
22
2
(
(
)
(
)
(
(
)
(
)
()
()
)
ii
i
i
i
i
i
CD
X
R
R
X
G
G
X
B
B
(6)
Figure 4. Source
Colo
r Ima
g
e
Figure 5.
Re
sults of Subtra
ction on the B
i
nary
Image
4. Experiment Re
sults
In ord
e
r to ve
rify the validity of the algo
ri
th
m in this
st
udy, we p
e
rfo
r
m a va
riety of video
simulatio
n
experim
ents, at
the same time make
the related com
pari
s
on
with Gau
ssi
an av
erag
e
method a
nd
method of th
e Gau
ssi
an
mixture mod
e
l
. Experiment
a
l environ
m
e
n
t are the Pe
ntiu
m
IV 3.0GHz
CPU, 512MB
RAM, and Vi
sual
c++ 6.
0
simulatio
n
pl
atform. The
data sa
mple
s fo
r
experim
ents
mentione
d ab
ove are t
he n
a
rration of training athlete
s
taken by monocular
came
ra
and the f
r
am
e rate i
s
2
5
frame
s
p
e
r
se
con
d
, the
si
ze of ea
ch fra
m
e in the vid
eo imag
e is
576
X720 pixel
s
of colo
r im
ag
es. Paramete
rs
of the al
go
rithm in th
e selecte
d
a
s
th
e value
s
in t
he
Table 1.
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: 4274 – 4
282
4280
Figure 6. Interface of Expe
riment Syste
m
Table 1. Ch
oi
ce of the Parameter
Parameter
Name
S
y
mbol of Para
m
eters
Value of Expe
riment
Number of
Gaus
sian Distribution
K
3
Learning rate
0.2
Brightness Boundaries Under
De
viation
al
o
0.75
Colour deviation
threshold
Th_CD
40
Brightness deviation threshold
Th_BD
0.8
(a) T
he num
b
e
r 12 fra
m
e
(b) T
he num
b
er 15 fra
m
e
(c) The n
u
mb
er 17 fra
m
e
(d) Th
e num
ber 19 frame
Figure 7. Source
Colo
r Ima
g
e
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Re
sea
r
ch of Backgroun
d Segm
entation Method in
Sports Vide
o (Shen Li
)
4281
(a) T
he num
b
e
r 12 fra
m
e
(b) T
he num
b
er 15 fra
m
e
(c) The n
u
mb
er 17 fra
m
e
(d) T
he num
b
er 19 fra
m
e
Figure 8. Re
sult of Subtraction on Binary
Image wi
th the Method of
Mixture of Ga
ussian
s Mod
e
l
(a) T
he num
b
e
r 12 fra
m
e
(b) Th
e num
ber 15 frame
(c) The n
u
mb
er 17 fra
m
e
(d) T
he num
b
er 19 fra
m
e
Figure 9. Re
sult of Subtraction on Binary
Im
age with the Method of
the Improved
Gau
ssi
an
Mixture Mode
l
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: 4274 – 4
282
4282
In orde
r to validate the result
s, the impr
ove
d
add
brightne
ss deviation an
d chroma
deviation of the Ga
ussia
n
mixture mod
e
l of
ba
ckg
ro
und
segm
ent
ation algo
rith
m are
adopt
ed.
Thro
ugh
com
pari
s
on
of th
e
expe
rimental
re
sult
s a
nd t
he exp
e
rim
e
n
t
al data, it
ca
n be
foun
d th
at
the experime
n
tal results o
f
the algorithm is su
p
e
rio
r
to other methods, and t
h
is algo
rithm
is
suitabl
e for t
he ima
g
e
s
segmentatio
n
of huma
n
bo
dy movemen
t
in the ice e
n
vironm
ent with
c
o
mplex backgr
o
und.
5. Conclusio
n
In the study, it mainly analyze
s
the r
oad spee
d skating trainin
g
unde
r ba
ckgroun
d
segm
entation
of human m
o
vements, on
the basi
s
of
the cla
s
sic of
the Gau
ssi
an
mixture mod
e
l,
it improves th
e model throu
gh addi
ng bri
ghtne
ss d
e
viation and
col
o
r deviation.
The
new imp
r
oved
algo
rit
h
m
can
be
a
pplied
to the
sp
eci
a
l
scen
e, and
can
g
e
t goo
d
backg
rou
nd
segmentatio
n
effect. From
the expe
rim
e
n
t
al re
sults of the alg
o
rithm
comp
ared
with
other
com
m
o
n
ba
ckgroun
d se
gme
n
tation alg
o
rithm,
the segme
n
tation effect
of the algo
rit
h
m
prop
osed in the pape
r is the best, it not only can
re
alize the dyn
a
mic segm
e
n
t
ation of athletes
pro
s
pe
ct info
rmation, an
d
also
can ef
fectivel
y rest
rain the
sha
dows, strong
light, and
mist
importa
nt it can ide
n
tify the su
rfa
c
e reflection. So
it i
s
foun
d that i
n
this
com
p
le
x ice sce
n
a
r
i
o
s,
our alg
o
rithm
has g
ood effe
ct and ha
s th
e better pe
rfo
r
man
c
e.
Referen
ces
[1]
Christo
pher
W
r
en, Ali Azarba
ye
ja
ni, P fi
nder: Real-
T
ime
T
r
acking
of theHuma
n
bod
y.
IEEE
T
r
ansactio
n
s o
n
Pattern Ana
l
ysis and Mac
h
i
neIntel
lig
enc
e.
199
7;19(
7): 78
0~
795.
[2]
O Javed, K Shafique, M Shah.
A hierarc
h
ic
al. ap
proac
h to
robustb
ackgro
und s
ubtractio
n usin
g col
o
r
and gr
adi
ent in
formati
on.
Wkshp. onMoti
on a
nd Vid
eo
C
o
mputin
g IEEE. 2002: 22~2
7
.
[3]
YZ
Hsu, HH Nag
e
l, G Reb
e
rs. Ne
w
L
i
ke
liho
od T
e
st Methods for
C
ha
nge D
e
tectio
n
in Imag
e
Sequ
enc
e.
Co
mp
uter Visi
on, Graphi
cs, an
dIma
ge Proc
essi
ng.
198
4; (26).
[4]
PL R
o
sin, T
Ellis.
I
m
ag
e
Differenc
e T
h
res
h
old
Strategi
es
and
Sha
dow
D
e
tection.
Pr
oc
essin
g
of th
e
British Mach
in
e Visio
n
Conf
e
r
ence. 19
95.
[5] D
Gutchess.
A
back
g
rou
nd
mode
l i
n
itia
li
z
a
ti
on
alg
o
rith
m fo
r vide
o surv
eil
l
ance
. Pr
oc. of
the 8th
IEE
E
Int'
l Conf. on Computer Vis
i
on
Vancouv
er. 20
01: 733~
7
40.
[6]
PW
Po
w
e
r, JA
Scho
on
ees.
U
ndersta
ndi
ng
b
a
ckgro
und
mix
t
ure
mo
dels
fo
r foregr
oun
d s
e
g
m
e
n
tatio
n
.
Proc. of IVCNZ
. 2002: 26
7~
27
1.
[7]
D Hal
l
, J Nasci
mento, P Ri
bei
ro, E Andra
de.
Co
mp
ariso
n
of
target detecti
o
n
al
gorith
m
s us
ing ada
ptive
backgr
oun
d mode
ls.
Internati
ona
l w
o
rksh
op
on
Performa
nc
e.
[8]
G Bailo, M Bariani,
P Ijas, M Raggio.
Back
grou
nd Esti
ma
tion w
i
th Ga
us
sian
Distri
butio
n for I
m
a
g
e
Seg
m
e
n
tatio
n
, a
fast appr
oac
h.
IEEE International Workshop on M
easur
ement S
y
stems
for Hom
e
land
Securit
y
, C
ontr
aba
nd D
e
tectio
n and Pers
on
al
Safet
y
Orlan
d
o
, F
L
, USA, 2005:29~
3
0
.
[9]
R Cucch
iara,
C Grana, M Piccardi, A Prati. Detect
ing m
o
ving o
b
jects, g
hosts an
d sha
d
o
w
s i
n
vid
e
o
streams.
IEEE
Trans on Pattern Anal
and Machine Intell.
20
03; 25(1
0
): 133
7~
134
2.
[10]
Zhilia
ng Wa
ng
, Jian Gao, C
hua
n
x
ia J
i
a
n
, Yu Ce
n, Xin
Che
n
. OLED
Defect Inspe
c
tion S
y
st
e
m
Devel
opm
ent
throug
h In
dep
end
ent C
o
mp
one
nt An
al
ysi
s
.
T
E
LKOMNIKA Indo
nes
ia
n Jo
urna
l
o
f
Electrical E
ngi
neer
ing
. 2
012;
10(8): 23
09-
23
19.
[11]
Ming-H
u
w
i
Hor
ng. Multi
l
ev
el
Minimum
Cros
s Entr
op
y Ima
ge T
h
resho
l
di
n
g
usi
ng Artific
i
al Be
e C
o
lo
n
y
Algorit
hm.
T
E
LKOMNIKA Indones
ian J
ourn
a
l of Electrica
l
Engi
neer
in
g
. 2013; 11(
9): 522
9-52
36.
[12]
Aghlm
and
i D,
F
aez K. A
u
tomatic S
egme
n
tati
on
of Gl
ottal Sp
ace fro
m
Vide
o Ima
g
e
s Bas
ed
o
n
Mathematic
al
Morph
o
lo
g
y
an
d the h
oug
h T
r
ansform.
Internatio
nal J
our
n
a
l of Electric
al
and C
o
mp
ute
r
Engi
neer
in
g (IJECE)
. 2012; 2(
2): 223-2
30.
[13]
Jun
Lai, M
e
i
Xi
e. Autom
a
tic
Segm
entati
o
n
fo
r Pu
lmon
ar
y V
e
sse
ls i
n
Plai
n T
horacic
CT
Scans
.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2012; 1
0
(4): 7
43-7
51.
Evaluation Warning : The document was created with Spire.PDF for Python.