TELKOM
NIKA
, Vol.13, No
.3, Septembe
r 2015, pp. 1
006
~10
1
3
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v13i3.1772
1006
Re
cei
v
ed Ma
rch 2
0
, 2015;
Re
vised July
22, 2015; Accepted Augu
st
1, 2015
Adaptive Background Extraction for Video Based
Traffic Counter Application Using Gaussian Mixture
Models Algorithm
Ra
y
m
ond Su
tjiadi*
1
, Endang Set
y
ati
2
, Resma
n
a
Lim
3
1
Departme
n
t of Computer En
g
i
ne
erin
g,
Institut Informatika Indon
esia,
Jl. Pattimura N
o
. 3, Suraba
ya
601
89, East
Ja
va, Indon
esia,
Ph./F
ax: +
6231
-734
637
5/73
49
324
2
Departme
n
t of Information T
e
chno
log
y
, Sek
o
la
h T
i
nggi T
e
knik Sura
ba
ya,
Jl. Ngag
el Ja
ya T
engah 73-
7
7
, Suraba
ya
60
284, East Java
, Indonesi
a
, Ph
./F
ax: +
6231-5
027
92
0/504
15
09
3
Departme
n
t of Electrical En
gi
neer
ing, Petra
Christia
n Un
ive
r
sit
y
,
Jl. Si
w
a
l
ankert
o
121-
13
1, Suraba
ya
60
236,
East
Java, Indones
ia, Ph./F
ax: +
623
1-29
83
442/8
4
9
256
2
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: ra
y
m
o
n
d
@
ik
ado.ac.i
d
1
, end
ang
@stts.edu
2
, resmana@
pe
tra.ac.id
3
A
b
st
r
a
ct
T
he bi
g citi
es i
n
the w
o
rl
d a
l
w
a
ys face the
traffi
c ja
m. T
h
i
s
prob
le
m
is c
ause
d
by t
he i
n
creas
in
g
nu
mb
er of vehi
cle fro
m
time to time an
d the
incre
a
se
of ve
hicle
is not a
n
ti
cipate
d
w
i
th the dev
elo
p
ment
of
ade
qu
ate n
e
w
road s
e
ctio
n. One i
m
p
o
rt
ant
aspect i
n
th
e traffic mana
ge
ment
co
ncept
is
the n
eed
of tra
ffic
dens
ity d
a
ta of
every
roa
d
s
e
ction.
T
her
efor
e, the
pur
pos
e
of this
p
a
p
e
r is
to an
aly
z
e
t
he possi
bi
lity of
opti
m
i
z
at
ion
o
n
the use of
vide
o f
ile rec
o
rded fro
m
CC
T
V
camer
a
for visual o
b
serv
ation a
nd to
ol
for
counti
ng traffic
dens
ity. T
he used
metho
d
i
n
this p
a
p
e
r is
ada
ptive
back
g
rou
nd extr
acti
on w
i
th Gaussi
an
Mixture Mod
e
ls
algor
ith
m
. It is expected to b
e
the alte
rn
ativ
e soluti
on to g
e
t traffic density data w
i
th a quit
e
ade
qu
ate accu
racy as one of
aspects for dec
ision
makin
g
pr
ocess in the tra
ffic engin
eeri
n
g
.
Ke
y
w
ords
:
traffic ma
na
ge
me
nt syste
m
, traffic de
nsity
c
ounter, ada
ptiv
e
b
a
ckgro
un
d extraction, gau
ssia
n
mixtur
e mod
e
ls
Copy
right
©
2015 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
The big
citie
s
in the
wo
rl
d always fa
ce the traffic
jam at high
way. This p
r
o
b
lem is
cau
s
e
d
by th
e in
cre
a
si
ng
numbe
r
of ve
hicle
from
ti
m
e
to time
an
d
the in
crea
se
of vehicl
e i
s
not
anticip
ated wi
th the development
of ade
quate ne
w ro
ad se
ction. In
orde
r to solve the proble
m
,
prop
er traffic
manag
eme
n
t con
c
ept i
s
n
eede
d to pre
v
ent and solve traffic jam
probl
em. The
r
e
are
many d
e
velopme
n
ts i
n
the traffic
ma
nagem
ent
co
nce
p
t, su
ch
a
s
the
poli
c
y o
f
Tran
sp
ortati
on
Agenci
e
s of
Surabaya, East Java, Indon
es
i
a
. Transportatio
n
Agencie
s o
f
Surabaya has
impleme
n
ted
the installati
on of co
untd
o
wn time
r on
some traffic lights an
d the install
a
tio
n
of
Clo
s
ed Ci
rcui
t Television (CCTV
)
at so
me crossr
oad
s that are co
n
necte
d to Surabaya Intellig
ent
Traffic
Sys
t
em (SITS) [1].
Another im
po
rtant a
s
pe
ct i
s
the n
eed
of tr
affic de
nsit
y data on ev
ery ro
ad
se
ction. The
data are ne
eded fo
r bet
ter mana
ge
ment on so
me roa
d
sections, su
ch
as to man
a
g
e
the
duratio
n of traffic light and
the appoint
ment of polic
e officers to manag
e the traffic, especi
a
lly at
rush hours
[2].
The d
a
ta is
currently obtai
ned
with ma
n
ual me
th
od b
y
appointin
g the office
rs directly to
the roa
d
se
cti
on with lo
w d
a
ta accu
ra
cy. Moreove
r
, it can b
e
do
ne
with ele
c
tro
n
i
c
devices,
su
ch
as the
sen
s
o
r
with highe
r d
a
ta accu
ra
cy. This sen
s
o
r
can b
e
the el
ectri
c
al
sen
s
or (e.g. by u
s
ing
spe
ed
gun
se
nso
r),
me
cha
n
ical
sen
s
or t
hat pla
n
ted i
n
the
asphalt,
or
com
b
inatio
n of th
e u
s
e
of
electri
c
al
an
d
me
cha
n
ical
sen
s
o
r
s.
The
usage
of
se
nso
r
m
odel
wa
s te
sted
b
y
Public Works
Agency
on
Soekarn
o
-Hatt
a
Street, Ba
n
dung,
We
st
Java, Indon
esi
a
an
d it
wa
s
named
PLAT
O
(Peng
hitung
Lalu Linta
s
O
t
omatis/Auto
m
atic Traffic Cou
n
ter) [3].
Unfortu
nately
,
the usage
o
f
electri
c
al an
d me
chani
cal
sen
s
o
r
s i
s
o
ften disturbe
d by its
appli
c
ation
a
nd in
stallatio
n
at
roa
d
se
ction.
T
he u
s
e of sen
s
o
r
ca
n be effective
when
t
he
vehicle
s
are
on the prope
r row a
nd the
r
e must
be a
vehicle to run
on one row
at one time [4]. It
is u
s
ually a
p
p
lied in fo
rm
of gate
syste
m
like th
e
toll
gate. When i
t
is not a
pplie
d, the counti
ng
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Adaptive Ba
ckgroun
d Extraction for Vid
eo Ba
se
d Tra
ffic Counte
r
…
(Ra
y
m
ond Sutjiadi)
1007
pro
c
e
ss
will
not be a
c
curate. Mo
reo
v
er, not all road
s
can b
e
applie
d wi
th that metho
d
con
s
id
erin
g the limitation of
width of a road sectio
n.
In this
pape
r,
the po
ssibilit
y on the
usa
ge of
CCTV
came
ra
will
be a
nalyze
d
for the
cou
n
ter of tra
ffic den
sity. T
hus, it
can
be
optimi
z
ed
for man
ual vi
su
al ob
se
rvatio
n by th
e officers
and the auto
m
atic stati
s
tical data o
n
traffic de
ns
ity. This data
can be on
e of factors to make
deci
s
io
n fo
r
managi
ng th
e
du
ration
of traffic light
or o
t
her type
s
of
traffic m
anag
ement. Be
sid
e
s,
this
method will solve problem
s
on some counti
ng method on traffic
densit
y
like
what
was
explained b
e
f
o
re, so th
e achieved data i
s
expe
cted to
be more a
ccurate.
The
system
use
d
a
daptiv
e ba
ckg
r
oun
d extra
c
ti
on
with G
a
u
ssi
a
n
Mixture
Mo
dels.
Thi
s
method i
s
th
e an
swer to
minimize error that
cau
s
e
d
by ch
angin
g
of ba
ckgro
und
con
d
ition
in
certai
n time, like gradu
al illumination
ch
ange
s, addi
ti
on of sha
d
o
w
object, and o
t
her sto
p
moving
objec
t [5]. Gauss
i
an Mixture Models
is
a further
development from Running Gaus
s
i
an Average
method that
prop
osed by
Wren et. al. [6]. This
alg
o
rithm has
ca
p
ability to cou
n
t the avera
g
e
value of pixel
s
an
d process it as
Gau
s
si
an mod
e
ls
i
n
real time
upd
ating from
ea
ch vide
o fram
es
without con
s
u
m
ing more co
mputer
resou
r
ce
s.
2. Bac
k
grou
nd
Extrac
tion
The b
a
ckg
r
o
und extractio
n
is a
meth
od to
sepa
rate or
dete
c
t the obje
c
t
(a
s the
foreg
r
ou
nd)
a
nd ba
ckgroun
d of input fra
m
e of a vide
o [7]. Like
wh
at is sho
w
n o
n
Figu
re 1, th
e
video is recorded from
CCTV cam
e
ra in
stalled
at cert
ain road
secti
on a
nd it i
s
sent to
controll
ing
room by u
s
in
g netwo
rk
co
nne
ction. The
video is re
co
rded into AVI file format.
Figure 1. Video Re
co
rdin
g
Techni
que
In this ca
se,
the term foreg
r
ou
nd is t
he obje
c
t of obse
r
vation.
The backg
round i
s
anothe
r obje
c
t that does not belong
to the obj
ect of obse
r
vation. In another word,
The
backg
rou
nd i
s
the static
obje
c
t that does
not cha
nge freq
uentl
y
and foreground is dyn
a
m
ic
obje
c
t that moves from o
n
e
point to ano
ther in ce
rtain
sequ
ential frames [8].
T
h
e
p
u
r
p
o
s
e
o
f
b
a
c
k
g
r
o
u
n
d
e
x
tr
ac
tio
n
pr
o
c
e
ss is to
de
te
c
t
th
e e
x
iste
n
c
e o
f
for
e
g
r
o
und
obje
c
t that be
come
s th
e target of dete
c
ti
on. In th
is
re
search, the fo
regro
und
obje
c
ts a
r
e ve
hicl
es
that move on certain
road
sectio
n and
backg
rou
nd
obje
c
ts are a
nother o
b
je
cts that belong
to
road
situation
,
like asp
halt, trees, traffic
si
gns, pa
rked
cars, and oth
e
r stati
c
obje
c
ts.
After the foregro
und
obje
c
t is o
b
taine
d
, the obje
c
t
can
be p
r
o
c
e
s
sed b
a
se
d on the
appli
c
ation, such
as the
counting o
n
d
e
tected
obj
e
c
ts, the mea
s
urem
ent of o
b
ject
width, a
nd
cla
ssifi
cation of
vehicle.
3.
Gaussi
an Mixture M
odel
s
Gau
ssi
an Mi
xture Mod
e
ls is the type
of
den
sity
model with some co
mpo
nents of
Gau
ssi
an fu
n
c
tion
s. Thi
s
algorith
m
i
s
good
en
oug
h
to
sup
port
b
a
ckgroun
d
extraction
p
r
o
c
ess
sin
c
e it has th
e reliability on
the chan
ge o
f
weather a
n
d
conditio
n
of obje
c
t detecti
on rep
eatedly
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 1006 – 10
13
1008
With this alg
o
rithm, every
pixel in a image fram
e
is modelled
into K of Gaussia
n
distrib
u
tion. In this case, K is the num
be
r of us
e
d
Ga
u
ssi
an di
strib
u
tion mod
e
l fro
m
3 to 5. Every
Gau
ssi
an m
o
del rep
r
e
s
ent
s the
differe
n
t
pixel colou
r
. The
scala
r
value i
s
u
s
e
d
on
graysca
l
e
image, whil
e RGB imag
e u
s
e
s
vecto
r
value.
The choo
sin
g
on num
be
r of use
d
m
odel de
pen
d
s
on the
co
nsid
eratio
n o
f
image
resolution, th
e perf
o
rma
n
ce of co
mpute
r
sy
stem
, an
d co
mplexity of model
ba
ckgro
und. M
o
re
numbe
r of model on every
pixel can ca
use mo
re
ad
aptive backg
round extra
c
ti
on pro
c
e
s
s si
nce
more
col
our
comp
one
nts
can
be mo
de
lled into eve
r
y pixel. However, it must
be compe
n
sa
ted
with syste
m
reso
urce that will be u
s
ed,
es
p
e
ci
ally wh
en image
re
solution is bi
g enou
gh.
In other
words, this
al
gorithm will m
odel every appeari
ng col
o
ur on
every pixel
at
one
time t {X1, ..., Xt} on a im
age frame th
at is mo
delle
d into Ga
ussi
an di
stributio
n K by usi
ng
the
same initiali
zation parameter. Pr
obabilit
y from pixel
v
a
lue
based on
the previous colours can
be
formulated as follows
[9]:
P
X
∑
w
,
∗
η
X
,
μ
,
,
Σ
,
(1)
Whe
r
e,
P(X
t
)
= Probability of colour value at time t.
K
= Num
ber of
Gau
ssi
an di
stribution o
n
every pixel.
w
,
= Estimation
on wei
ght fro
m
i-th Gau
ssi
an distri
butio
n at time t.
µ
,
= Mean valu
e from i-th Ga
ussian di
strib
u
tion at time t.
∑
,
=
Covariance matrix from
i-th Gauss
i
an at time t.
Ƞ
= Fun
c
tion of
Gaussia
n
probability den
sity.
With the form
ula of function of Gaus
sian probability density as follows [8]:
η
X
,
μ
,
Σ
π
/
|
Σ
|
/
e
μ
Σ
μ
(2)
And the cova
rian
ce matrix
is assum
ed in
forms of [9]:
Σ
,
σ
I
(3)
For the next
image fram
e, every pixel is
com
b
in
ed with every K of of G
aussia
n
distrib
u
tion m
odel o
n
the
same
corre
s
p
o
nding
pixel,
starting fro
m
the di
stributio
n mod
e
l with
the
large
s
t to the
small
e
st p
r
o
bability. A pixel is dete
r
m
i
ned to b
e
suitable
with
one of
Gau
s
sian
distrib
u
tion m
odel
s wh
en it
belong
s to t
he ra
nge
2.
5
of deviation
stand
ard. O
n
the other
ha
nd
,
whe
n
a
pixel
has the val
u
e
inste
ad
of 2.
5 of d
e
vi
ation
stan
da
rd, it i
s
stated t
o
b
e
un
suitabl
e
wi
th
that Gaussia
n
distrib
u
tion
model [9, 10].
μ
k
−
2.5 *
σ
k
< X
t
<
μ
k
+ 2.5
*
σ
k
(4)
Whe
r
e,
X
t
=
Pixel c
o
lour vec
t
or (RGB) at time t.
μ
k
= Vecto
r
of mean value of
pixel (RGB
) from k-th G
a
u
s
sian.
σ
k
= Value of de
viation stand
ard fro
m
k-th
Gau
ssi
an.
If a pixel is suitable
with one of G
aussia
n
distribution mod
e
l
s, the para
m
eter of
Gaussi
an model will be updated. To
update the wei
g
ht value, the fo
llowing form
ula is used [9]:
ω
k,t
= (1 -
α
)
ω
k,t-1
+
α
(M
k,t
) (
5
)
Whe
r
e,
ω
k,t
=
Weight from k
-
th Gaussian at time t.
α
= Lea
rnin
g ra
te.
M
k,t
= The value i
s
1 for the sui
t
able model a
nd 0 for othe
r model.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Adaptive Ba
ckgroun
d Extraction for Vid
eo Ba
se
d Tra
ffic Counte
r
…
(Ra
y
m
ond Sutjiadi)
1009
After upd
atin
g the
wei
ght
value, do
the
norm
a
liz
ation
to ma
ke
the
total wei
ght from K of
Gau
ssi
an di
stribution m
ode
l = 1.
To update th
e mean value
,
the following
formula is u
s
ed [9]:
μ
t
= (1 –
ρ
)
μ
t-
1
+
ρ
X
t
(6)
Whe
r
e
ρ
=
αȠ
(X
t
|
μ
k
,
σ
k
)
The valu
e of
deviation
sta
ndard n
eed
s
perio
dic
up
da
te wh
en th
ere is pixel val
ue that i
s
suitabl
e with
distrib
u
tion. T
he followi
ng formul
a is u
s
e
d
for the upd
ate [9]:
σ
t
2
= (1 -
ρ
)
σ
2
t-1
+
ρ
(X
t
–
μ
t
)
T
(X
t
–
μ
t
) (
7
)
If a pixel is not suitable with
all Gaussian
distrib
u
tion m
odel
s on co
rresp
ondi
ng pi
xel, the
Gaussi
an model with
the smallest
probability will
be removed
and replaced with
Gaussi
an
model for th
e new
pixel colo
ur. The
new
Gau
ssi
a
n
model
will
be initialized
with mea
n
value
based on the
vector value,
high varia
n
t value, and lo
w weight value.
The
next ste
p
is to
determine th
e pix
e
l in th
e
ba
ckgroun
d a
n
d
foreg
r
ou
nd
o
b
ject
s a
nd
the sele
ction i
s
done. At the beginni
ng, the sele
ct
ion i
s
done by so
rting the existing model ba
sed
on the value of
ω
/
σ
2
(fi
t
ness value) where the most optima
l distribution as the background is still
placed
on th
e
top p
r
io
rity, while
the
distribution th
at d
oes not
refle
c
t the ba
ckgro
und i
s
pla
c
ed
on
the lowe
st priority. Some highe
st value
s
from
tho
s
e
distributio
n model
s are
selecte
d
until the
weig
ht value
fulfils th
e th
reshold
valu
e that
i
s
det
ermin
ed befo
r
e.
T
he sel
e
cted dist
ributi
on
model i
s
th
e
n
dete
r
min
e
d
as the
ca
nd
idate of
ba
ckgrou
nd. T
he
followin
g
formula i
s
u
s
e
d
to
cho
o
se B of backgroun
d di
stributio
n [9]:
B
a
r
g
∑
ω
(8)
Whe
r
e T is th
e smalle
st proportio
n
from
the
data and
it should b
e
counted a
s
the
backgroun
d.
Whe
n
the pixel colo
ur b
e
lo
ngs to the
ca
tegory of one
candi
date of
backg
roun
d
model,
the pixel will
be consi
dered as the background (pix
el
gets value 0/
black colour). Moreover, the
pixel that do
es not b
e
lon
g
to the cat
egory of ba
ckgroun
d mod
e
l will be
co
nsid
ere
d
a
s
the
foreg
r
ou
nd (p
ixel gets valu
e 1/white).
The
result in form
of bi
nary image
will be
processed
furtherm
ore. T
h
e advanced
pro
c
e
ss i
s
the forming of b
o
xes of dete
c
tion, countin
g
,
and obje
c
t classificatio
n
.
4. Sy
stem
Arch
itecture
The sy
stem consi
s
ts of sub
-
proc
esse
s a
s
explain
ed in
Figure 2.
Figure 2. System Block
Di
agra
m
a)
Input system
is in form of video file from
the CCTV recordin
g with certain du
ratio
n
and in a
format of Aud
i
o Video Interl
eave (AVI).
b)
In the image
optimizatio
n bloc
k, the video file will be
extracted int
o
its com
p
o
s
i
ng frame
s
,
whe
r
e ea
ch f
r
ame
will be pro
c
e
s
sed, a
s
su
ch
that the resulting im
age will be o
p
timized fo
r
the next process.
c)
In the detecti
on blo
ck, thre
e pro
c
e
ss a
r
e
runnin
g
as e
x
plained in Fi
gure 3.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 1006 – 10
13
1010
Figure 3. Det
a
il of Detectio
n Block
Diag
ram
(1)
Backgroun
d extraction
i
s
a
m
e
thod
for sep
a
ratin
g
bet
wee
n
backg
rou
nd and
foreg
r
ou
nd o
b
ject
s usi
ng
Gau
ssi
an Mixture Mod
e
ls (GMM) al
gorit
hm.
(2)
Thre
sh
olding
is a meth
od for chan
ging t
he imag
e into
binary ima
g
e
.
Object (a se
t of
pixel) d
e
tecte
d
a
s
b
a
ckg
r
o
und
will
be
gi
ven
bin
a
ry code as
0 (bl
a
ck
colou
r), while
the object (a set of pixel) detected as
foreground will be given binary code as 1
(white colo
ur).
(3)
Blob d
r
a
w
ing
is a
step
fo
r d
r
a
w
ing
of
boxes ci
rclin
g the
fore
gro
und
obje
c
ts that
rep
r
e
s
ent the
presen
ce of detecte
d obje
c
ts.
d)
In the
counting block, the
detected foregr
ound objects, whi
c
h
are r
epresented as
bl
ob, will
be co
unted.
e)
In the classification blo
c
k, each blob wil
l
be classifie
d
to obtain the data on the number of
vehicle
s
ba
se
d on group. T
h
is cl
assification pr
ocess i
s
ba
sed
on the dime
nsio
n
(width a
n
d
length) of the
vehicle
s
.
f)
Output from the syste
m
are:
(1)
Statistic data
of total number of vehicle
s
.
(2)
Statistic data
of total number of vehicle
s
base
d
on gro
up.
(3)
Data of traffic density per
minute.
(4)
Detail data of
total number
of vehicle
s
ba
sed o
n
gro
up
in certai
n inte
rval time.
(5)
Video display
which is alre
ady mappe
d with
blob, so use
r
ca
n mo
nitor the accu
racy
of
sy
st
em.
(6)
Export repo
rt data into text
and CSV file format.
5.
Sy
stem Desi
gn and Fea
t
ures
This system wa
s
develo
p
ed
u
s
ing Op
enCV so
ftwa
r
e. Open
CV i
s
a
C lang
ua
ge library
whi
c
h p
r
ovid
es
several fe
ature
s
for
co
mputer vi
sion
and digital i
m
age p
r
o
c
e
s
sing
appli
c
ati
ons
[11].
Here are the feature
s
avail
able
in this a
pplication (Fi
gure 4
)
:
(1)
Video playe
r
: for playing A
V
I video file.
(2)
Traffic de
nsit
y counter: for
c
ontrolling tra
ffic density function.
(3)
System settin
g
: for setting the paramete
r
s va
lue which
are re
quired
for run
n
ing th
e system.
(4)
Summary
re
p
o
rt: for displ
a
y
i
ng the sum
m
ary
rep
o
rt.
(5)
Detail report: for dis
p
laying t
he report in s
u
c
h
interval
time.
(6) Video
display
.
6. Experimenta
l
Result
Experiment f
o
r thi
s
rese
arch i
s
d
o
n
e
by u
s
ing
a sim
u
lation
and
sa
mpl
e
video.
Simulation vi
deo i
s
u
s
ed
to sim
u
late a
n
ideal
roa
d
co
ndition, which
is
suitable
wi
th the nee
ds
o
f
softwa
r
e, i.e. obje
c
t of vehicle
s
record
ed pe
rp
e
ndi
cularly from
a
bove, video
minimized from
other obj
ect
s
that may interfere det
e
c
tion
, and there is
no other n
o
ise.
Then
re
co
rde
d
sample
vid
eo
captu
r
e
s
t
he real
co
ndit
i
on of
high
wa
y is u
s
e
d
in
the te
st.
Due to the di
fficulty in recordin
g the video in an ide
a
l way, the existing re
cord
ing is take
n in a
con
d
ition that
is almo
st ide
a
l. The re
co
rded
video
of Wa
ru- Pe
ra
k
High
way, Surabaya is
use
d
for this
experiment.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Adaptive Ba
ckgroun
d Extraction for Vid
eo Ba
se
d Tra
ffic Counte
r
…
(Ra
y
m
ond Sutjiadi)
1011
Figure 4. Software Use
r
Interface
The re
adin
g
of softwa
r
e u
s
ing
simulatio
n
video is stat
ed as follo
ws, and the re
sul
t
is as Tabl
e 1.
The re
al num
ber of vehicl
e
= 274 vehi
cl
es
Total vehicle
cou
n
ted by system = 29
8 vehicle
s
Differen
c
e
= 298 – 27
4 =
+24 vehi
cle
s
The pe
rcenta
ge of readi
ng
error =
(24 / 2
74) * 10
0 % = + 8,76 %
Tabel 1. Te
sting Re
sult Usi
ng Simulation
Video
Perangkat
Lunak
Manual %
Erro
r
Grou
p 1
0
0
0
Grou
p 2
14
13
+7,69
Grou
p 3
39
49
-20,41
Grou
p 4
75
81
-7,41
Grou
p 5
69
61
+13,11
Grou
p 6
101
70
+44,29
To cou
n
t the number of
vehicle ba
se
d on gro
up usin
g simul
a
tion video is
done by
determi
ning t
he thre
sh
old
value of dim
ensi
on of
ea
ch vehi
cle g
r
oup. Group
1
is the
small
e
st
dimen
s
ion of
vehicle a
nd g
r
oup 6 i
s
the bigge
st.
The re
adin
g
of softwa
r
e u
s
ing
re
corde
d
sam
p
le vid
eo is
stated
as follo
ws, , and the
result is as T
able 2.
The re
al num
ber of vehicl
e
= 124 vehi
cl
es
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 1006 – 10
13
1012
Total vehicle
cou
n
ted by system = 12
2 vehicle
s
Differen
c
e
= 124 - 12
2 =
+ 2 vehicle
s
.
The pe
rcenta
ge of readi
ng
error =
(2 / 12
4) * 100 % =
+1.61%
Table 2. Te
sting Re
sult Usi
ng Sample Vi
deo
Software
Manual
Reading
% Erro
r
Grou
p 1
99
89
+11,24
Grou
p 2
23
35
-34,29
Grou
p 3
0
0
0
Grou
p 4
0
0
0
Grou
p 5
0
0
0
Grou
p 6
0
0
0
The rea
d
ing
on numb
e
r of
vehicle ba
se
d on its grou
p is determi
n
ed as follo
ws: group 1
is the
sm
all vehicl
e/com
m
on pa
ssen
ge
r an
d g
r
ou
p
2 is
big vehi
cle/bus/tru
c
k.
The g
r
ou
p of
big
vehicle
can
be cate
gori
z
ed into more
spe
c
ific gr
o
ups. Howeve
r, it is not possible
sin
c
e
th
e
con
d
ition of the re
co
rdin
g of sample vid
eo is not in id
eal co
ndition.
7. Conclu
sion
The alg
o
rithm
of Gau
ssi
an
Mixture Mod
e
l
s is
quite reli
able to u
s
e a
s
on
e of ba
ckgrou
nd
extraction
m
e
thod
s
with
several
con
d
i
t
ions
of vide
os. T
h
is algo
rithm i
s
a
b
le
to ad
apt to
the
existing chan
ges, so the a
c
cura
cy in de
tecting fore
ground o
b
je
ct is maintain
ed.
The ave
r
ag
e
results
of co
u
n
ting on total
v
ehicle
have
a quite hi
gh
accuracy, a
b
o
ve 80
percent. The
result is variou
s on eve
r
y con
d
ition
of video. Wh
en the ro
ad
become
s
mo
re
cro
w
d
ed, the accuracy is al
so re
du
ced.
The ave
r
age
result of co
unting on ve
hicle
b
a
sed
on its g
r
ou
p
has a
rel
a
tively low
accuracy
with variou
s re
sults
on the
con
d
it
ion of
video and
different vehi
cle g
r
ou
ps.
The
con
d
ition d
u
ri
ng the recording of
samp
le video t
hat
is not id
eal
by forming
certain el
evation
angle
al
so
causes the
error in
cla
s
sification
of ve
hicle
group.
Due
to m
a
n
y
vehicle
s
with
overlap
p
ing
l
ook in vide
o
frame. T
he
majority
of l
a
rge
s
t pe
rcent
age
of erro
r
happ
en
s on
t
h
e
smalle
st
and
the big
g
e
s
t vehicl
e g
r
ou
ps. The
erro
r
in
the small
e
st vehicle
group
(moto
r
cycle
)
i
s
cau
s
e
d
by it
positio
n that i
s
n
ear ea
ch
o
t
her,
so the
system read
s
many moto
rcycles a
s
on
e
unit
of vehicle
with big dime
nsion. Thu
s
, that con
d
it
ion
cau
s
e
s
the e
rro
r in
cla
ssif
y
ing the vehi
cle.
Mean
while, t
he e
rro
r in
group of bi
gge
st vehicl
e
is
cau
s
e
d
by th
e takin
g
tech
nique
of sa
m
p
le
video that is too nea
r with the su
rf
ace of the road, so the big vehi
cl
e is not fully
captu
r
ed in o
n
e
frame. Th
e takin
g
of sa
mple video
that is too
n
ear
with the
roa
d
surfa
c
e also cau
s
e
s
the
reflectio
n
of sunlight on th
e
body
su
rface
and the wi
nd
shiel
d
gla
ss
o
f
vehicle that
is ca
ptured by
came
ra th
at
make
s th
e system often
adapt its
Ga
ussian
pixel
value extrem
ely. These t
w
o
con
d
ition
s
make the bi
g o
b
ject
s unread
able an
d un
countabl
e.
Although th
e
existen
c
e
of sha
d
o
w
i
s
minimi
zed by
system, th
e
extreme
existen
c
e of
sha
d
o
w
can
be o
ne of
m
a
in fa
ctors to
ca
use t
he
e
rro
r in
re
adin
g
. The
existe
nce
of
shad
o
w
make
s t
w
o v
ehicl
es with
near po
sition
rea
d
a
s
one
unit of
obje
c
t sin
c
e it m
a
ke
s two bl
ob
s of
vehicle
con
n
e
c
ted to ea
ch
other.
This
system
i
s
mo
re
suita
b
l
e to u
s
e o
n
road
se
ction t
hat is
pa
ssed
by motor ve
h
i
cle
with
4 wheel
s o
r
more (toll ro
ad) an
d avoi
d the conditi
on of durin
g severe traffic jam (whe
n the
vehicle
can
n
o
t
move at all).
The sy
stem accuracy is
really determi
ned
by the con
d
ition of video re
co
rdi
ng, the
con
d
ition of road sectio
n, and the
cho
o
s
ing of
suitab
le input pa
ra
meters value
for the algo
rit
h
m
p
r
oc
es
s
.
Referen
ces
[1]
Dinas P
e
rhu
b
u
nga
n Kota Sur
aba
ya. Sur
a
b
a
y
a Inte
lli
ge
nt
T
r
ansp
o
rt S
y
ste
m
(SIT
S). 2015
.
[2]
Hon
g
ji
n Z
hu,
Hon
ghu
i F
an,
Shuq
ian
g
Gu
o. Movin
g
Ve
hicl
e Detecti
on
an
d T
r
acking i
n
T
r
affic Image
s
base
d
on
Hor
i
zonta
l
Edg
e
s.
T
E
LKOMNIKA Indon
esia
n
Journ
a
l of El
e
c
trical Eng
i
n
e
e
rin
g
. 201
3;
11(1
1
); 647
7-6
483.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Adaptive Ba
ckgroun
d Extraction for Vid
eo Ba
se
d Tra
ffic Counte
r
…
(Ra
y
m
ond Sutjiadi)
1013
[3]
Bada
n Pen
e
liti
an da
n Pen
g
e
m
ban
gan K
e
m
entria
n
Pekerj
a
an Umum. PL
AT
O: Penghitu
ng La
lu Li
ntas
Otomatis. Kementeri
an Pek
e
rj
aan Um
um. 20
10.
[4]
Lim R, T
h
iang, Rizal.
Pe
n
gukur
an K
e
p
a
datan
Arus
L
a
lu
Li
ntas Me
ngg
un
akan
Se
nsor K
a
mera
.
Semin
a
r Nasi
o
nal Ind
u
strial E
l
ectron
ics 200
3
(IES 2003). 20
03.
[5]
Shiso
ng Z
hu,
Min Gu, Jing
Liu. Movi
ng V
ehicl
e Det
e
ctio
n and T
r
ackin
g
Algor
ithm in
T
r
affic Video.
Te
lkom
n
i
ka
. 20
13; 11(6); 3
053
-305
9.
[6]
C Wren, A Aza
r
ba
ye
jan
i
, T
Darrell, AP Pe
ntla
nd. Pf
ind
e
r: re
al time track
i
ng
of the h
u
man
d
bod
y.
IEEE
Transactio
n
on
Pattern Analys
is and Mac
h
in
e
Intellig
enc
e.
1997; 19(
7): 780
-785.
[7]
Ai
yan L
u
, Lu
o
Z
hong, L
i
n L
i
, Qingb
o W
ang.
Movin
g
Veh
i
cl
e Reco
gn
ition
and F
e
ature E
x
tractio
n
from
T
unnel Monitor
i
ng Vi
deos.
T
e
l
k
omnik
a
. 201
3; 11(10); 60
60-
606
7
[8]
Rostia
nin
g
sih
S, Adipran
a
ta R, W
i
bisono F
S
.
Adaptive Ba
ckgrou
nd de
ng
an Metod
e
Gaussia
n
Mixtur
e
Mode
ls Untuk
Real-T
ime T
r
a
cking.
Jurnal Teknik
Inform
atika
. 2008; 9(
1); 68-7
7
.
[9]
Stauffer C, Grimson W
E
L. Adaptiv
e
Backgro
un
d Mixtur
e Mod
e
l
s for Rea
l
-T
ime T
r
acking.
Massach
usetts Institute of
T
e
chno
log
y
. 1
999
.
[10]
Kae
w
T
r
a Kul
Pong
P, Bo
w
d
en R.
A
n
I
m
pr
oved
Ada
p
tive
Backgro
un
d
Mixture Mo
de
l
for Rea
l
-T
i
m
e
T
r
acking w
i
th
Sha
dow
Det
e
ction.
Pr
oc.2
nd E
u
rop
e
a
n
W
o
rkshop
o
n
Adv
ance
d
Vide
o Bas
e
d
Surveil
l
a
n
ce S
y
stems. 20
01.
[11]
Kaeh
ler A, Bradski G. Lear
nin
g
Open
CV:
Computer V
i
s
i
on
w
i
t
h
the O
pen
CV Li
brar
y
.
2nd ed
itio
n.
Califor
ni
a, Unit
ed States of Am
erica: O’Rei
l
l
y
Med
i
a, Inc.
Evaluation Warning : The document was created with Spire.PDF for Python.