Indonesi
an
Journa
l
of
El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
10
,
No.
1
,
A
pr
il
201
8
, p
p.
184
~
1
9
0
IS
S
N:
25
02
-
4752
, DO
I: 10
.11
591/
ijeecs
.
v
10
.i
1
.pp
1
8
4
-
1
9
0
184
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Measuri
ng the R
oa
d
Traf
fic Inten
sity
U
sin
g Neural
Networ
k
with C
om
pu
ter
Visi
on
Muhamm
ad
Ha
m
dan
1
,
O
t
hma
n
Omr
an
Kha
li
f
ah
2
,
Te
ddy Sur
ya G
u
na
w
an
3
1
,3
Depa
rtment
of
Elec
tr
ical and
C
om
pute
r
Engi
n
e
eri
ng,
Kulliyy
a
h
of
Engi
n
ee
rin
g
2
Depa
rtment of
Inform
at
ion
S
y
s
t
ems
,
Kulliyy
ah
of
ICT
Inte
rna
ti
ona
l
Isl
a
m
ic
Univer
sit
y
Malay
s
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ja
n
12
, 2
01
8
Re
vised
Ma
r
16
,
201
8
Accepte
d
Ma
r
30
, 201
8
Tra
ffi
c
conge
st
i
on
pla
gues
al
l
dr
ive
r
aro
und
the
world.
To
solve
thi
s
proble
m
computer
vision
ca
n
b
e
used
a
s
a
tool
to
dev
el
op
a
lterna
t
ive
route
s
an
d
el
iminate
tr
aff
i
c
conge
stions.
I
n
the
cur
r
ent
gene
ra
ti
on
with
inc
re
asing
num
b
er
of
ca
m
era
s
on
the
street
s
and
lower
cost
for
Inte
rne
t
of
Thi
ngs
(IoT
)
thi
s
soluti
on
will
have
a
gr
ea
t
er
i
m
pac
t
on
cur
re
n
t
s
y
stems
.
In
this
pape
r,
t
h
e
Mac
roscopic
Ur
ban
Tra
ff
ic
m
odel
is
used
using
computer
vision
as
it
s
source
and
tr
aff
ic
inten
sit
y
m
onit
oring
s
y
stem
is
imple
m
ent
ed.
The
in
put
of
th
is
progra
m
is
ext
ra
ct
ed
from
a
tra
ff
ic
survei
ll
an
ce
c
amera
and
anot
h
er
progra
m
running
a
neur
a
l net
work c
l
assificat
ion
which ca
n
cl
assif
y
and
dist
i
nguish
the
vehi
c
le
t
y
pe
is
on
the
roa
d
.
T
he
neur
al
net
w
ork
tool
box
is
tra
ine
d
with
positi
ve
and
ne
gat
iv
e
input
to
inc
re
ase
ac
cur
acy
.
Th
e
accur
acy
of
th
e
progra
m
is
compare
d
to
o
the
r
r
elate
d
works
done
and
the
tre
nds
o
f
the
tr
aff
i
c
int
ensity
from
a
roa
d
is
al
so
calc
ula
t
ed.
Ke
yw
or
d
s
:
N
eu
ral
N
et
w
ork
O
bject
D
et
ect
ion
T
raffic
I
ntensit
y
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Muh
am
m
ad
H
a
m
dan
,
Dep
a
rtm
ent o
f El
ect
rical
an
d
Com
pu
te
r
E
ng
i
neer
i
ng, Kulli
yy
ah
of E
nginee
rin
g
,
In
te
r
natio
nal Is
lam
ic
U
niv
er
sit
y M
al
ay
sia
,
Jal
an Go
m
bak
,
5310
0 Ku
al
a
Lum
pu
r, (+
603)
6196
4521
.
Em
a
il
:
ha
m
dan
hg@yah
oo.c
om
1.
INTROD
U
CTION
Urba
n
de
vel
opm
ent
in
ci
ti
es
and
lo
wer
e
d
c
os
t
of
owni
ng
a
ve
hicle
ha
s
le
d
to
rise
i
n
num
ber
of
veh
ic
ular
in
t
he
street
cau
sing
tr
af
fic
co
ngest
io
n
[
1]
.
N
ow
a
days,
it
is
norm
al
to
be
stuck
in
t
raff
ic
fo
r
prolo
nged
pe
riod
an
d
ti
m
e
is
sp
e
nt
f
or
unne
cessary
pa
rt
of
the
day
to
go
th
rou
gh
traf
f
ic
.
Fig
ur
e
1
s
hows
a
sam
ple
of
traffi
c
j
am
in
Asian
ci
ti
es.
H
owever,
with
t
he
current
im
pr
ov
e
m
ent
to
the
a
vaila
bili
ty
of
inter
net
and
vast
c
ov
e
r
age
of
the
t
raffic
surveil
la
nc
e
cam
era
has
le
d
m
any
research
e
rs
to
opt
into
us
in
g
c
ompu
te
r
visio
n
to
so
l
ve t
he
co
ngest
io
n pro
blem
.
In
t
he
rece
nt
ye
ars,
a
dv
a
nce
m
ents
in
In
te
r
net
of
Thi
ngs
(IoT
)
a
s
well
a
s
lowe
red
c
os
t
of
owning
a
sm
artph
on
e
wi
th
fast
inte
rn
et
co
nn
ect
ivit
y
can
be
use
d
t
o
co
nv
ey
neces
sary
inf
orm
ati
on
to
the
r
oad
us
e
r.
Applic
at
ion
s
s
uch
a
s
W
aze
a
nd
G
oogle
m
a
p
can
pro
vid
e
com
m
uters
with
im
po
rtant
de
ta
il
s
reg
ar
ding
thei
r
com
m
ute
ahe
a
d
of
tim
e
and
this
has
hel
ped
reducin
g
the
traf
fic
j
am
s
in
ci
ti
es.
Ho
we
ve
r,
these
a
pp
li
c
at
ions
sti
ll
dep
en
d
on
the
us
er
to
se
nd
in
form
at
ion
to
the
ser
ve
rs
to
c
om
pu
te
th
e
sta
te
of
the
ro
a
d
t
hey
are
in
to
convey
the
inf
or
m
at
ion
to
oth
er
.
I
f
no
us
e
r
s
are
us
i
ng
th
e
ap
plica
ti
on
,
t
he
n
the
pro
ba
bili
ty
of
the
c
onge
sti
on
occurri
ng is
not detec
te
d.
Estim
at
ion
of
the
am
ou
nt
of
traff
ic
on
the
ro
a
ds
at
any
giv
e
n
point
of
tim
e
is
the
fi
rst
ste
p
in
m
itigati
ng
tra
ffi
c
congesti
on
[
2]
.
A
c
omm
on
m
e
tho
d
is
to
pl
ace
sens
or
s
on
the
ro
a
d
a
nd
c
ount
the
num
ber
of
tim
es they are
act
uated
by t
he
p
assin
g wh
eel
s o
f
a
ve
hicle
. Th
is ap
proac
h suffers fr
om
f
our
m
ai
n
prob
le
m
s
: a)
it
is
exp
ensi
ve
to
de
plo
y,
as
the
sens
ors
ne
ed
to
be
pa
rtia
ll
y
e
m
bed
ded
i
n
the
ta
rm
ac,
b)
the
sens
ors
on
t
he
ro
a
d
are
pr
on
e
to
theft,
c)
se
nsors
need
to
be
placed
at
m
ult
iple
entry
and
exit
po
ints
on
t
he
r
oad,
to
m
a
i
ntain
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Meas
ur
in
g
t
he Ro
ad Tr
affi
c I
ntensity
Using
Ne
ur
al Net
wo
r
k wi
th Co
mput
er Visi
on (Mu
ham
m
ad Ha
mda
n)
185
accurate
c
ount
s,
a
nd
d)
e
ve
n
on
a
si
ng
le
str
et
ch
of
r
oad,
t
he
se
nsors
nee
d
to
be
placed
at
regular
inter
vals
to
est
i
m
at
e the d
ensity
on diffe
re
nt se
gm
ents o
f t
he
r
oa
d.
Figure
1
.
Tr
a
ffi
c congesti
on i
n Ku
al
a
Lum
pu
r
, Mal
ay
sia
Ba
sed
on
the
li
te
ratur
e
re
view
done
,
t
he
w
orks
t
hat
s
hows
si
m
il
arities
with
c
urre
nt
resea
rch
co
pe
is
discusse
d
in d
e
ta
il
in
this
sect
i
on.
I
n
[
3]
,
the au
th
or
s u
ses b
l
ob
a
naly
sis
to
track
a
nd
d
et
ect
m
ov
in
g
cars on
the
ro
a
d.
Bl
ob
ana
ly
sis
us
es
backgro
und
subtrac
ti
on
in
each
fr
a
m
e
to
gen
erate
a
blo
b
of
m
oving
pi
xels
across
the
fr
am
e.
Bl
ob
analy
sis
is
the
si
m
plest
way
to
fin
d
m
ov
in
g
obj
ect
in
subse
quent
f
ram
es.
T
he
auth
or
us
e
d
traff
ic
ca
m
eras
places
in
2
diff
ere
nt
areas
wh
ic
h
a
re
ca
m
eras
facing
3
la
nes
du
bb
e
d
as
NIPA
and
TOL
L
PLAZ
A
with
2
la
nes
of
r
oad.
For
trai
ning
the
im
ages,
Haa
r
Ca
sca
de
was
us
e
d
to
cl
assify
the
im
ag
es
an
d
X
ML
file
was
ge
ner
at
e
d
us
ing
580
po
sit
ive
i
m
ages
and
1500
ne
ga
ti
ve
sa
m
ples
t
o
ge
ner
at
e
the
dataset
.
Finall
y,
the
perform
ance
of
the
syst
e
m
is
evaluate
d
us
in
g
posit
ive
dete
ct
ion
com
par
e
d
to
act
ual
nu
m
ber
of
veh
ic
l
es.
This
pap
e
r
pr
ov
i
ded
a
good
platf
o
r
m
fo
r
this
res
e
arch
to
pic,
howev
e
r
the
casc
ade
us
e
d
f
or
th
is
researc
h
was
base
d
on
al
gorit
hm
s
us
e
d
for
face
tracki
ng
an
d
re
su
lt
will
be
m
or
e
accu
rate
if
the
featur
es
a
re
m
or
e
disti
nc
t
in
dataset
im
ages.
The
ne
xt
pa
pe
r
fo
c
us
e
d
on
usi
ng
f
oreg
rou
nd
-
base
d
m
e
thod
to
detect
m
ov
in
g
obj
ect
ac
ro
ss
fr
am
es
in
[4]
.
Im
age
w
as
acq
uire
d
usi
ng
tra
ff
ic
ca
m
eras
and
re
siz
ed
into
sm
al
ler
fr
am
es.
The
n
the
po
sit
ive
i
m
ages
of
cars
flo
wing
thr
ough
the
traf
fic
is
captur
e
d
into
po
sit
ive
and
neg
at
ive
f
ram
es.
The
posit
ive
i
m
ages
a
re
then
run
ning
thr
ough
the
Haar
Ca
scade
li
ke
the
work
of
[
3
]
but
the
key
dif
f
eren
ce
is
t
he
s
a
m
ple
was
not
done
us
in
g
t
he
fi
rst
fr
am
e,
bu
t
t
he
first
nu
m
ber
of
fram
es
was
us
e
d
to
ide
ntify
the
fore
gro
und
of
the
im
ages
t
o
increase the d
e
te
ct
ion
accur
ac
y and
r
em
ov
in
g
unwa
nted noi
se in each
f
ra
m
e b
ei
ng
p
r
oc
essed.
T
he
i
m
a
ge
wit
h
foregr
ound
det
ect
ion
bec
om
e
s
m
or
e
dilat
ed
and
easi
e
r
to
detect
i
m
ages
that
are
m
ov
in
g.
T
he
ad
va
nta
ge
of
this
is
the
am
ount
of
processi
ng
re
quired
to
detect
the
ca
rs
are
reduce
d
due
to
sm
al
le
r
ROI
but
aut
hor
do
es
no
t
e
xp
la
in
on
the
key
pe
rform
ance
detai
ls
su
c
h
as
f
ram
e
rates
or
eval
ua
ti
on
on
the
c
ha
ng
e
s
in
f
ram
es
siz
es
on the acc
urac
y.
Anothe
r
pa
per
discuss
i
ng
on
t
he
sam
e
issue
is
w
ritt
en
by
[5]
wh
e
re
the
im
age
ca
pturin
g
m
et
ho
d
wa
s
us
e
d
to ide
ntif
y t
he
nu
m
ber
of cars o
n
the ro
ad.
T
he pr
ocedur
e
us
e
d
by the
w
rite
r was
bac
kgr
ou
nd subtra
ct
ion
to
ide
ntify
the
obj
ect
s
on
the
ro
a
d
a
nd
m
eas
ur
i
ng
t
hem
.
In
s
te
ad
of
m
easuri
ng
t
he
num
ber
of
bl
ob
s
,
the
a
uthor
counts
t
he
num
ber
of
pi
xels
in
wh
ic
h
t
he
i
m
age
is
dif
fere
nt
f
or
m
the
f
irst
i
m
age
wit
h
no
co
ngest
io
n.
B
y
est
i
m
ation
the
range
of
val
u
e
s
in
these
dif
fe
ren
t
co
ngest
io
n
rates
m
app
ed
to
ra
ng
e
of
pi
xel
in
each
sta
te
,
the
con
cl
us
io
n
is
reache
d.
Key
idea
of
this
pa
per
is
the
sim
pl
ic
it
y
of
the
al
go
rithm
wh
ere
the
im
ages
are
processe
d
a
s
is
an
d
no
post
pr
ocessin
g
is
re
quire
d
to
ob
ta
in
data
howe
v
e
r,
the
data
will
be
not
m
eaningf
ul
if
the
patte
rn
of
cars
or
ve
hicle
ty
pes
are
so
m
et
hin
g
im
po
rta
nt
that
need
to
be
identifie
d.
The
aut
hor
al
so
doe
s
no
t
pro
vid
e
any
insig
ht on
t
he
m
easur
e
m
ent for
t
he
c
onges
ti
on
detect
ion s
yst
e
m
s.
Ed
ge
de
te
ct
ion
is
al
so
a
k
ey
m
et
ho
d
i
n
im
a
ge
cl
assifi
cat
io
n
to
ol
as
resea
rch
e
d
by
[
5]
w
hich
us
es
I
P
ca
m
eras
places
in
strat
e
gic
lo
cat
ion
s
on
the
ro
a
d
t
o
pre
dict
traf
fic
co
ngest
ion
.
T
he
obj
ect
detect
ion
pr
oc
ess
is
done
us
i
ng
e
dge
detect
ion
a
nd
the
pa
ram
e
te
rs
are
passe
d
on
to
a
fuzzy
log
ic
syst
e
m
f
or
traf
fic
est
im
at
ion
.
Ed
ge
detect
io
n
in
this
pa
per
us
es
a
S
obel
f
il
te
r
m
ixed
with
Kalm
an
filt
er
to
trac
k
the
m
ov
ing
ob
j
ect
.
The
nu
m
ber
of
m
ov
in
g
ob
j
ect
is
then
passe
d
to
a
fu
zzy
log
ic
analy
zer
that
consi
ders
thre
e
par
am
et
ers
su
ch
as
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vol
.
10
, N
o.
1
,
A
pr
il
201
8
:
1
8
4
–
1
9
0
186
veh
ic
le
de
ns
it
y,
distance
bet
w
een
ve
hicle
and
la
stl
y
veh
ic
le
siz
e.
The
logi
c
analy
zer
then
pr
e
dicts
the
traf
fic
into
dif
fer
e
nt
cat
egories
base
d
on
the
log
ic
ta
ble
of
the
f
uz
zer.
This
m
eth
od
can
acc
ur
a
te
ly
pr
edict
the
traff
ic
intensit
y
of
th
e
m
ov
ing
t
raff
ic
howe
ver
the
reg
i
on
of
tra
c
king
in
these
i
m
ages
is
not
s
et
therefo
re
th
e
ROI
trackin
g of m
ov
in
g veh
ic
le
s
m
us
t be don
e
on th
e
wh
ole fra
m
es w
hic
h wil
l i
ncr
ease
the
pr
ocessin
g
ti
m
e.
To
detect
the
traf
fic
inten
sit
y,
the
re
a
re
3
c
r
it
ic
al
app
r
oac
h
in
t
he
m
od
el
li
ng
of
t
h
e
tra
ff
i
c
flo
ws
in
ro
a
ds
nam
el
y
the
spa
ti
al
m
od
el
li
ng
[6]
–
[
9]
,
ti
m
e
pr
edict
ion
m
od
e
ll
ing
[
10
]
–
[
12]
a
nd
non
-
pa
r
a
m
et
ric
m
od
el
li
ng
.
S
pa
ti
al
m
od
el
li
ng
re
fers
to
us
i
ng
t
he
ph
ysi
ca
l
sp
ace
on
real
tim
e
to
est
i
mate
the
m
otion
of
t
he
veh
ic
le
s
base
d
on
the
ph
ysi
c
al
par
am
et
ers
su
c
h
as
r
oad
di
sta
nce,
ca
r
le
ng
t
h,
r
oad
cu
r
vatu
re
a
nd
ot
he
r
key
el
e
m
ents.
Ti
m
e
pr
e
dicti
on
m
od
el
s
us
e
sta
ti
sti
cal
data
colle
ct
ed
ove
r
tim
e
to
pre
dict
th
e
curre
nt
tren
d
on
th
e
ro
a
d
an
d
updat
e
the
data
on
r
eal
tim
e
if
necessary.
N
on
-
pa
ram
et
ric
m
od
el
li
ng
ref
e
rs
to
us
in
g
com
pu
t
at
ion
al
intensive
al
gor
it
h
m
s
su
c
h
as
arti
fici
al
neura
l
netw
ork
a
nd
w
ork
i
nd
e
pe
ndent
f
ro
m
any
real
-
ti
m
e
inp
ut
but
reli
es
on
colle
ct
ion
of
data
to
synthesiz
e
an
outp
ut.
H
oweve
r,
in
this
i
m
ple
m
entat
i
on
the
s
pace
base
d
m
od
el
li
ng
tech
nique is
us
e
d
t
o
sim
plify t
he
r
esult.
2.
RESEA
R
CH MET
HO
D
Figure
2
sho
w
s
the
overall
i
m
ple
m
entat
ion
of
t
he
pro
pos
ed
s
olu
ti
on.
T
he
pr
ocess
sta
r
ts
wh
e
n
the
traff
ic
cam
era
is
trai
ned
on
th
e
foregr
ound
a
nd
bac
kgr
ound
i
m
ages
on
a
n
area.
T
he
n
the
pro
gr
am
disp
la
ys
the
i
m
ages
to
the
us
er
to
i
den
ti
f
y
the
Re
gio
n
of
In
te
rest(RO
I)
fo
r
th
e
desire
d
locat
io
n
to
sta
rt
c
ountin
g
th
e
cars
.
The
us
er
will
c
hoos
e
t
he
lo
we
r
an
d
uppe
r
re
gion
of
t
he
RO
I
to
determ
ine
traff
ic
flo
w
int
o
a
nd
out
of
th
e
li
nk
as sho
wn in f
ollow
in
g
F
ig
ur
e
3.
Figure
2
T
he o
ver
al
l
process t
o
ide
ntify t
ra
ff
i
c intensit
y
The
fi
gure
s
hows
the
ov
e
rall
i
m
ple
m
entat
io
n
of
the
propo
sed
s
olu
ti
on.
T
he
proce
ss
sta
r
ts
wh
e
n
th
e
traff
ic
cam
era
is
trai
ned
on
th
e
foregr
ound
a
nd
bac
kgr
ound
i
m
ages
on
a
n
area.
T
he
n
the
pro
gr
am
disp
la
ys
the
i
m
ages
to
t
he
us
er
to
ide
ntif
y
the
Re
gion
of
I
ntere
st
(R
OI)
f
or
th
e
desire
d
l
ocati
on
to
sta
rt
co
unti
ng
th
e
cars
.
The
us
er
will
c
hoos
e
t
he
lo
we
r
an
d
uppe
r
re
gion
of
t
he
RO
I
to
determ
ine
traff
ic
flo
w
int
o
a
nd
out
of
th
e
li
nk
as
well
as
cam
era
place
m
e
nt
is
s
hown
i
n
f
ollow
i
ng
F
i
gure
3.
This
m
et
ho
d
is
not
i
m
ple
m
en
te
d
in
oth
e
r
researc
h d
on
e
in
the
li
te
ratur
e
rev
ie
w
as m
os
t w
orks runs
tra
ckin
g on entire
v
ide
o fr
am
e.
Figure
3
.
Tr
ac
king locati
on of
veh
ic
le
s
for
tr
aff
ic
inte
ns
it
y and cam
era p
la
ce
m
ent
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Meas
ur
in
g
t
he Ro
ad Tr
affi
c I
ntensity
Using
Ne
ur
al Net
wo
r
k wi
th Co
mput
er Visi
on (Mu
ham
m
ad Ha
mda
n)
187
Nex
t
the
fram
e
captur
e
d
fro
m
ca
m
era
su
bt
racted
from
th
e
cur
re
nt
fr
am
es
and
the
dif
f
eren
ces
ar
e
m
ark
ed
us
i
ng
blob
a
naly
sis
and
it
s
outp
ut
is
filt
ered
for
reg
i
on
s
that
ar
e
la
rg
e
r
tha
n
5000
pix
el
s
w
hich
i
s
dep
e
ndent
on
t
he
ori
gina
l
vide
o
siz
e.
Ne
xt
the
cent
ro
i
d
of
the
m
ov
ing
obje
ct
s
is
cal
culated
f
ro
m
the
blo
b
a
nd
it
is p
asse
d
to
the classi
ficat
io
n
to
ol.
Figure
4
.
GUI
of the ca
r usin
g
Ma
tl
ab
The
cl
assifi
cat
ion
to
ol
co
ntains
the
pr
et
rai
ne
d
ob
j
ect
outp
ut
of
im
ages
f
ro
m
veh
ic
le
s
gi
ven
before
the
pro
gr
am
is
init
ia
li
zed.
The
trai
nin
g
proce
ss
involves
us
i
ng
posit
ive
i
m
ages
of
cars
a
nd
oth
e
r
ve
hicle
s
and
neg
at
ive
im
age
con
ta
ini
ng
ba
ckgr
ound
an
d
unrelat
ed
ra
ndom
i
m
ages
to
al
low
the
co
m
pu
te
r
to
dif
f
eren
ti
at
e
the
ty
pe
of
ob
je
ct
giv
e
n
to
it
. Th
e obj
ect
f
il
e
is
ge
ner
at
ed
us
ing
n
e
ur
al
n
et
work
to
olbo
x
i
n
Ma
tl
ab
w
hich
giv
es
a
high
acc
ur
ac
y
of
data
f
or
obj
ect
cl
assifi
ca
ti
on
.
When
t
he
i
m
age
is
passe
d
to
t
his
pro
gra
m
,
it
will
dete
rm
ine
if the ce
ntr
oid
giv
e
n
to
it
cont
ai
ns
wh
ic
h vehi
cl
e fr
om
the tr
ai
ned
obj
e
ct
.
3.
RESU
LT
S
A
ND D
I
SCUS
S
ION
3.1.
Pro
gra
m
A
cc
urac
y
The
pro
gr
am
is
r
un
on
a
r
oa
d
an
d
the
data
c
ollec
te
d
is
disc
us
se
d
in
this
se
ct
ion
.
The
sam
ple
locat
io
n
sel
ect
ed
wa
s
i
n
Pe
na
ng,
Ma
l
ay
sia
wh
e
re
a
ca
m
era
is
plac
ed
on
the
ce
nter
of
the
r
oad
facin
g
the
li
nk
.
T
he
accuracy
of
th
e
cl
assifi
cat
ion
pro
gr
am
is
as
sho
w
n
in
T
ab
le
2.
This
res
ul
t
sh
ows
t
he
optim
u
m
conditi
on
i
s
us
e
d
for
data
c
ollec
ti
on
an
d
im
age
processi
ng.
The
ass
umpti
ons
us
e
d
for
veh
ic
le
tracki
ng
is
the
car
is
m
ov
ing
in
a
c
onsta
nt
velocit
y
an
d
s
peed
bel
ow
60km
/h
as
fa
ste
r
ve
hicle
s
trac
king
is
no
t
pr
act
ic
al
to
track
us
i
ng
backg
rou
nd subtracti
on.
Table
1
. O
ver
a
ll
A
ccu
racy
Ty
p
e of
veh
icle
d
etected
Nu
m
b
e
r
o
f
veh
icle
d
etected
Nu
m
b
e
r
o
f
v
eh
icle
u
n
d
etected
Accurac
y
(
%)
Av
erage Det
ectio
n
ti
m
e(
m
s)
Car
85
15
85
150
Moto
rcy
cl
e
30
7
80
150
The
pro
gram
c
an
te
ll
with
alm
os
t
85
pe
rce
nt
prob
a
bili
ty
if
the
detect
ed
obj
ect
is
a
ca
r
and
80
%
accuracy
of
t
he
detect
ed
obj
e
ct
is
a
m
oto
rcy
cl
e.
Ba
sed
on
t
hese
data,
it
pr
ov
i
des
a
good
ref
e
ren
ce
t
o
optim
iz
e
the
syst
e
m
in
t
he
f
uture
th
ough
is
sli
gh
tl
y
lo
wer
t
han
[3
]
,
[6]
.
Nex
t
the
pro
gr
am
is
ru
n
du
rin
g
the
night
tim
e
to
get
the
accu
rac
y
wh
e
n
t
he
c
on
diti
on
is
da
r
k
a
nd
T
a
ble
3
s
hows
t
he
acc
ur
ac
y
of
the
data
c
ollec
te
d
on
the
sam
e
locat
ion
.
Table
2
. Acc
uracy
o
f
the
pro
gram
d
ur
in
g nig
ht
Ty
p
e of
veh
icle
d
etected
Nu
m
b
e
r
o
f
veh
icle
d
etected
Nu
m
b
e
r
o
f
veh
icle
n
o
t detected
Accurac
y
(
%)
Av
erage Det
ectio
n
ti
m
e
(
m
s
)
Car
8
24
25
220
Moto
rcy
cl
e
2
15
11
230
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vol
.
10
, N
o.
1
,
A
pr
il
201
8
:
1
8
4
–
1
9
0
188
The
data
in
T
able
3
s
hows
that
the
accu
r
acy
has
dro
pp
ed
sig
nifica
ntly
com
par
ed
to
the
vid
e
o
captu
red
du
rin
g
the d
ay
. Th
es
e resu
lt
s ar
e n
ot
co
m
par
ed
to
oth
e
r
researc
he
s as this test
was
no
t d
on
e b
y
them
.
Ther
e
a
re
a
fe
w
fact
or
s
th
at
con
t
rib
uted
to
this
dro
p
in
th
e
accur
acy
s
uc
h
as
di
ff
e
ren
t
l
igh
ti
ng
on
the
roa
d
causin
g
the
ch
ang
e
s
in
t
he
s
had
e
of
car
be
lowe
r
tha
n
it
sh
oul
d
be
w
he
n
the
blob
a
na
ly
sis
cod
e
is
r
unning.
Nex
t,
the
ca
r
at
nig
ht
us
es
li
ght
to
il
lu
m
inate
the
ro
a
d
thus
changin
g
a
ke
y
featur
e
extra
ct
ed
durin
g
tra
i
ning
process
.
The
li
gh
t
proj
e
ct
ed
by
the
car
al
s
o
incre
ased
bri
gh
tne
ss
that
m
akes
the
cam
era
un
a
ble
to
see
the
act
ual
car
sh
a
pe
but
on
ly
se
e
ci
rcle
du
ri
ng
blo
b
a
naly
sis.
Last
ly
,
the
color
s
of
the
li
ght
colo
r
al
so
change
wh
e
n
t
he
ca
r
is
m
ov
ing
f
orwa
rd an
d b
ac
kw
a
rd as the
h
ea
dli
gh
t i
s
whit
e an
d
the
brea
k
li
ght i
s r
e
d.
3.1.
Tr
affic In
ten
sity
The
tra
ff
ic
i
ntensity
durin
g
l
ow
traf
fic
an
d
high
tra
ff
ic
m
ov
em
ent
is
cal
culat
ed
a
nd
th
e
res
ult
is
as
sh
ow
n
in
Ta
ble 3
a
nd
4.
Table
3
.
T
ra
ff
i
c intensit
y d
uri
ng lo
w
c
ongest
ion
Ti
m
e
Star
t
Ti
m
e
end
Nu
m
b
e
r
o
f
car
s
into
lin
k
(per
m
in
u
te)
Nu
m
b
e
r
o
f
car o
u
t of
lin
k
(per
m
in
u
te)
Tr
af
f
ic inten
sity
at
ti
m
e
k
(
n
o
C
o
n
g
estion)
0
1
26
25
0
.01
8
3
6
7
3
4
7
1
2
9
10
-
0
.01
7
3
0
7
6
9
2
2
3
7
6
0
.01
7
0
4
5
4
5
5
3
4
3
4
-
0
.01
6
9
1
7
2
9
3
4
5
22
21
0
.01
8
0
7
2
2
8
9
5
6
12
12
0
6
7
4
5
-
0
.01
6
9
8
1
1
3
2
7
8
11
10
0
.01
7
3
0
7
6
9
2
8
9
13
13
0
Total
107
106
0
.01
9
5
8
6
6
6
5
Table
4
. T
ra
ff
i
c intensit
y d
uri
ng m
ediu
m
to hig
h
c
ongestio
n
Ti
m
e
Star
t
Ti
m
e
end
Nu
m
b
e
r
o
f
car
s
in
to
lin
k
(per
m
in
u
te)
Nu
m
b
e
r
o
f
car o
u
t
o
f
link
(per
m
in
u
te)
Tr
af
f
ic inten
sity
at
ti
m
e
k
(Co
n
g
estio
n
prese
n
t)
0
1
20
17
0
.05
3
3
5
9
6
8
4
1
2
17
14
0
.05
2
7
3
4
3
7
5
2
3
22
20
0
.03
6
3
4
19
17
0
.03
5
5
7
3
1
2
3
4
5
20
21
-
0
.01
8
0
7
2
2
8
9
5
6
22
20
0
.03
6
6
7
13
13
0
7
8
23
21
0
.03
6
1
4
4
5
7
8
8
9
20
23
-
0
.05
4
6
5
5
8
7
to
tal
176
166
0
.17
7
0
8
3
6
Wh
e
n
the
re
is
no
c
onge
sti
on,
the
total
traff
i
c
intensit
y
in
each
tim
e
is
low
an
d
the
t
raff
ic
intensit
y
durin
g
pea
k
hours
are
high
there
fore
the
pr
ogram
will
be
able
to
te
ll
the
diff
e
ren
ce
in
the
traff
ic
intensit
y
cl
early
.
The
use
r
can
set
the
t
hr
es
hold
wh
e
r
e
the
traff
ic
in
te
ns
it
y
is
hig
h,
m
edium
or
low
de
pe
nd
i
ng
on
the
ro
a
d
c
onditi
on
as ex
plaine
d
in
the
researc
h p
aper
[
13
]
.
Kno
wn
li
m
it
ation
of
the
i
m
plem
entat
ion
are
sh
a
dows,
c
ar
colo
r
an
d
perform
ance.
Durin
g
li
gh
t
conditi
on
cha
nges,
t
he
accu
ra
cy
of
the
pro
gra
m
var
ie
s
due
to
sh
a
pe
of
s
ha
dow
inc
rease
s
an
d
dec
rea
se
s
ove
r
tim
e.
An
ot
her
lim
it
at
ion
is
th
e
car
c
olors
li
ke
black
an
d
gr
ey
ca
us
es
th
e
syst
e
m
to
not
detect
t
he
m
ov
ing
obj
ect
betwee
n
f
ram
es
as
i
t
is
the
sam
e
colo
r
of
the
ro
a
d
hen
c
e
r
edu
ci
ng
the
a
ccur
acy
.
Final
ly
,
the
perform
ance
dr
ops
sig
nifican
tl
y
if
m
or
e
t
han
4
ve
hicle
s
ar
e
tracked
at
th
e
sa
m
e
t
i
m
e
howev
e
r
the
f
ram
es
siz
e
us
e
d
ty
pical
ly
con
ta
in
s
ab
out
3
cars
m
axi
m
u
m
.
To
bette
r
ev
al
uate
the
pe
rfor
m
ance
of
our
pro
po
se
d
m
eth
od,
a
traff
ic
c
row
d
si
m
ula
ti
on
base
d o
n
intel
li
ge
nt
agen
t c
ould
be
dev
el
op
e
d.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Meas
ur
in
g
t
he Ro
ad Tr
affi
c I
ntensity
Using
Ne
ur
al Net
wo
r
k wi
th Co
mput
er Visi
on (Mu
ham
m
ad Ha
mda
n)
189
Figure
5
.
Sam
ple G
U
I
C
reated
to Display
Ou
t
pu
t
4.
CONCL
US
I
O
N
AND
F
UT
U
RE W
ORKS
In
t
his
pa
pe
r,
a
m
e
tho
d
of
ide
ntifyi
ng
t
he
tra
ff
ic
inte
ns
it
y
of
the
ro
a
d
is
im
plem
ented
us
in
g
c
om
pu
te
r
visio
n.
Ba
sed
on
the
data
acqu
i
red,
the
program
can
te
l
l
accuratel
y
if
the
ro
a
d
is
co
ng
e
ste
d
base
d
on
the
equ
at
io
n
of
t
he
Ma
c
ro
sc
opic
Urba
n
Tra
ff
ic
m
od
el
s.
How
ever,
us
i
ng
co
m
pu
te
r
visio
n
sti
ll
has
lim
it
ation
s
as
discusse
d
in
t
he
pr
e
vious
sec
ti
on
.
I
n
the
f
utu
re
,
a
bette
r
pro
gr
am
wh
ic
h
colle
ct
m
ulti
pl
e
ro
a
d
sta
ti
sti
cs
and
gen
e
rate
a
pre
dicti
on
al
go
rithm
based
on
the
r
oad
t
rends
is
need
e
d
to
be
able
to
te
ll
ro
ad
us
e
rs
inf
or
m
at
ion
about
their
j
ou
rn
ey
on
the
road.
Mo
re
ov
e
r,
a
traff
ic
cr
owd
sim
ulatio
n
c
ou
l
d
be
dev
el
op
e
d
to
e
valua
te
our
m
et
ho
d.
ACKN
OWLE
DGE
MENT
The
a
uthors
w
ou
l
d
li
ke
to
e
xpress
t
heir
gr
at
it
ud
e
to
the
Ma
la
ysi
an
Min
ist
ry
of
Hi
gher
Ed
ucati
on
(MO
HE),
w
hich
has
pro
vid
e
d
f
unding
f
or
the
researc
h
th
rou
gh
the
F
un
dam
ental
Re
se
arch
Gr
a
nt
Sc
hem
e,
FRGS
15
-
194
-
0435.
REFERE
NCE
S
[1]
A.
S.
Abdelf
ata
h,
et
al
.
,
“
Ev
al
u
at
ing
th
e
Sus
ta
i
nabi
lit
y
of
Traff
ic
Grow
th
in
Malay
si
a,
”
J.
Tr
af
fi
c
Logis
t.
Eng.
,
vol
/i
ss
ue:
3
(
1
)
,
p
p.
6
–
11
,
2015
.
[2]
A.
As
aha
ra
,
et
al.
,
“
Mac
roscop
ic
Stru
ct
ura
l
Su
m
m
ari
za
ti
on
of
Road
Networks
for
Mobil
e
Tr
a
ffic
Inform
a
ti
on
Servic
es,
” in
7th
Inte
rnat
ional
C
onfe
renc
e
on
Mo
bil
e
Data
Mana
geme
nt
(
MDM’06)
,
pp.
42
–
42
,
2
006
.
[3]
Y.
Hasan,
et
al.
,
“
Com
par
at
ive
ana
l
y
sis
of
vehicle
detec
ti
on
in
urba
n
tra
ffic
e
nvironment
using
Haa
r
ca
sca
de
d
cl
assifi
ers
and
bl
ob
statistic
s,
”
FTC 2016
-
Proc. Futur.
Techno
l. C
onf.
,
pp.
547
–
55
2,
2017
.
[4]
Q
.
W
u,
et
al.
,
“
Rea
l
-
ti
m
e
veh
i
cl
e
de
tecti
on
w
it
h
fore
ground
-
base
d
ca
sc
ade
cl
assifi
er,”
IET
Image
Proce
ss
.
,
vol
/i
ss
ue:
10
(
4
)
,
pp.
289
–
296
,
20
16.
[5]
B.
Ea
m
th
ana
kul
,
et
a
l.
,
“
The
Traffic
Congest
ion
Inve
stigating
S
ystem
b
y
Im
age
Proce
ss
ing
from
CCTV
Camera
,
”
in
2017
In
te
rnati
onal
Conf
ere
nce on
Digi
tal Arts,
Me
dia
and
Tech
nolog
y
(
ICDAMT)
,
pp.
240
–
245
,
2017
.
[6]
F.
S
.
Ferna
nd
e
z,
et
a
l.
,
“
Background
subtraction
for
ae
r
ia
l
surveil
l
anc
e
co
ndit
ions,
”
14th
Int.
Conf
.
Inno
v.
Comm
unit
y
Serv
.
“T
ec
hnolog
ie
s
Ev
ery
on
e”,
I4C
S
2014
-
Conf
.
Proc.
,
pp.
28
–
33
,
2014
.
[7]
K
.
Phalke
and
R.
Hega
di
,
“
Objec
t
tra
ck
ing
usin
g
Mutipl
e
H
y
po
the
sis,”
in
2016
3rd
Inte
rnationa
l
Confe
ren
ce
on
Signal
Proce
ss
in
g
and
Int
egrate
d
Net
works (
SPIN
)
,
pp.
707
–
710
,
2016
.
[8]
F.
Yan,
et
al
.
,
“
An
improved
tra
ffic
net
work
m
odel
for
m
odel
-
ba
sed
pre
dic
t
ive
c
o
ntrol
,
”
in
Proc
eedings
of
the
33rd
Chine
se
Con
trol
Confe
ren
ce
,
pp
.
3405
–
3410
,
201
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vol
.
10
, N
o.
1
,
A
pr
il
201
8
:
1
8
4
–
1
9
0
190
[9]
A.
Boulm
akoul
,
et
al.
,
“
Towa
rds
Scal
abl
e
Distribut
ed
Fram
ework
for
Urb
an
Congesti
on
Tra
ffi
c
Patt
ern
s
W
are
housing,
”
Appl
.
Comput.
I
nte
ll.
So
ft Comput.
,
vol
.
2015
,
20
15.
[10]
L.
Li
,
“
A
cont
in
uum
m
odel
of
tr
aff
ic
flow
on
roa
d
net
works
,
”
in
2015
Inte
rnation
al
Confe
ren
ce
o
n
Tr
anspor
tat
ion
Information
and
Safe
t
y
(
ICTIS)
,
v
o
l
.
2
,
pp
.
595
–
59
8
,
2015
.
[11]
L.
Sindhia,
et
al.
,
“
Eff
i
ci
en
t
det
e
ct
ion
tr
ac
ki
ng
of
m
ult
ipl
e
m
oving
obje
ct
s
in
te
m
pora
l
dom
ai
n,
”
in
201
6
Inte
rnational
Co
nfe
renc
e
on
Rec
ent
Tr
ends
in
Inf
orm
ati
on
Techn
ology
(
ICRTIT)
,
pp.
1
–
6
,
2016
.
[12]
A.
Ouess
ai
and
M.
Kec
h
e,
“
Deve
lopment
o
f
r
oad
traffic
flow
m
odel
for
nigh
tt
ime
cond
it
ions
,
”
3rd
Int
.
Conf
.
Control.
Eng. In
f.
Te
chnol.
CEIT
2015
,
2015.
[13]
H.
G.
M
.
H
amdan,
et
a
l.
,
“
Tra
f
fic
in
te
nsi
t
y
m
onit
oring
using
m
ult
ipl
e
ob
ject
det
e
ct
ion
with
t
ra
ffic
surveillan
ce
ca
m
era
s,
”
IOP
C
onf.
S
er.
Ma
te
r.
Sci
.
Eng
.
,
vol
/
issue:
260
(
1
)
,
pp
.
2
8
–
30,
2017
.
Evaluation Warning : The document was created with Spire.PDF for Python.