Indonesi
an
Journa
l
of
El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
13
,
No.
1
,
Jan
uar
y
201
9
,
pp.
2
2
~
2
6
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v1
3
.i
1
.pp
2
2
-
2
6
22
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Develop
ment of f
ra
m
ewo
rk for d
etectin
g s
mokin
g scene
in vide
o clips
Po
onam G
,
S
h
as
h
an
k
B.
N
,
At
hri
G R
ao
Depa
rtment
o
f
C
om
pute
r
Scie
n
ce a
nd
Engi
n
ee
rin
g,
Rashtr
eey
a
Vi
d
y
a
lay
a
Col
le
g
e of
Engi
n
ee
rin
g
(
RVCE),
Benga
luru
,
Ind
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
ug
21
, 201
8
Re
vised
Oct
22
, 2
01
8
Accepte
d
Nov
5
, 2
01
8
Acc
ording
to
Gl
obal
Adult
Tob
ac
co
Surve
y
20
16
-
17,
61.
9%
of
peopl
e
ar
e
quit
ti
ng
tob
ac
co
.
Th
e
r
ea
son
w
as
the
warni
ngs
display
e
d
on
t
he
produc
t
cove
rs,
vid
eo
cli
ps,
and
adve
r
ti
sm
ent
s.
The
foc
u
s
of
thi
s
pape
r
is
to
aut
om
ate
the
proc
ess
of
display
ing
warn
ing
m
e
ss
age
s
in
vide
o
clips.
Thi
s
paper
expl
a
ins
the
dev
el
opm
ent
of
a
s
y
stem
to
au
tomaticall
y
de
tect
t
he
sm
oking
sce
nes
using
image
re
cogni
t
ion
appr
oac
h
in
vi
deo
cl
ips
and
t
hen
add
the
warni
ng
m
essage
to
the
vie
w
er.
The
appr
o
ac
h
ai
m
s
to
de
te
c
t
t
he
c
iga
re
tte
obje
c
t
using
Tens
orflow’s
object
d
et
e
ct
ion
AP
I.
T
ensorflow
i
s
an
ope
n
source
software
li
bra
r
y
for
m
a
c
hine
l
ea
rning
pr
ovide
d
b
y
Goog
le
which
is
broa
dl
y
used
in
the
fie
ld
imag
e
rec
ogni
ti
on.
At
pre
sent,
Faster
R
-
CNN
(Regi
on
-
base
d
Convolut
ional
Neura
l
Network
s)
wi
th
Inc
ept
io
n
ResNet
is
the
T
ensorflow’s
slowest
but
m
ost
ac
cur
ate
m
odel
.
Faster
R
-
CNN
with
Inc
eption
Resne
t
v2
m
odel
is
us
ed
to
det
ect
sm
o
king
sce
nes
b
y
t
rai
ning
the
m
odel
with ci
g
ar
et
t
e
as
an
ob
ject
.
Ke
yw
or
d
s
:
Deep l
ear
ning
Faste
r
R
-
C
N
N
Object
detect
io
n
S
m
ok
ing
scene
d
et
ect
io
n
Tens
orflo
w
Copyright
©
201
9
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
:
Poon
am
G
,
Dep
a
rtm
ent o
f C
om
pu
te
r
Scie
nce a
nd E
ng
i
ne
erin
g,
Ra
sh
treey
a
Vidyal
ay
a
Coll
eg
e of E
ng
i
neer
i
ng (
R
VCE)
, B
eng
al
uru, I
nd
ia
.
Em
a
il
:
po
onam
ghuli@r
vce.
ed
u.
in
1.
INTROD
U
CTION
In
t
his
a
ge
of
onli
ne
a
nd
s
ocial
m
edia,
ci
nem
as
re
m
ai
n
a
prom
inent
form
of
ente
rtai
nm
ent
and
influ
e
nce
on
y
ou
t
h.
India
pro
du
ce
s
ap
pro
xi
m
at
ely
800
to
1000
m
ov
ie
s
a
ye
ar,
le
a
ding
to
the
require
m
ent
of
disp
la
yi
ng
the
warnin
g
m
es
sage
w
he
n
th
e
s
m
ok
in
g
sc
ene
is
showca
sed.
T
he
wa
r
nings
as
of
now
a
re
disp
la
ye
d m
anu
al
ly
.
This
pro
po
se
d
work
e
xp
la
ins
the
de
vel
op
m
ent
of
a
f
ram
ewo
r
k
t
o
a
uto
m
atical
ly
detect
the
sm
ok
ing
scenes
us
i
ng
ne
ur
al
netw
ork
m
od
el
and
the
n
disp
la
y
t
he
r
equ
i
red
wa
r
ning
m
essage.
Th
e
chall
en
ge
to
dete
ct
the
sm
ok
ing
sc
enes
in
vi
deo
c
li
ps
is
that
onl
y
the
sm
all
po
r
ti
on
of
t
he
sm
ok
in
g
e
ven
t
m
ay
be
s
howcase
d
a
nd
it
m
ay
be
disp
l
ay
ed
f
or
f
racti
on
of
a
sec
ond.
To
ove
rco
m
e
this
chall
e
ng
e
,
obj
ect
detect
ion
m
et
ho
ds
a
r
e
us
e
d
to
detect
diff
e
ren
t
ki
nd
of
ci
gar
et
te
s
.
T
hes
e
ci
gar
et
te
s
m
ay
hav
e
va
ryi
ng
s
ha
pes,
c
olo
rs
a
nd
siz
e.
The
n
a
warnin
g
m
essage s
uc
h
as “
Sm
ok
i
ng K
il
ls”
or “Sm
ok
ing i
s i
njurio
us
to
h
ea
lt
h”
is dis
play
ed
in
the
vide
o
cl
ip.
Goo
gle’s
O
bje
ct
detect
ion
AP
I
is
bu
il
t
on
t
op
of
Te
ns
orfl
ow.
Th
e
re
are
di
ff
e
re
nt
pre
-
trai
ne
d
Tens
orflo
w
m
od
el
s
avail
a
ble
fo
r
ob
j
ect
detect
ion
suc
h
as
Sing
le
Shot
Mult
ibo
x
Detect
or
(
SS
D)
with
Mob
il
eNet,
S
S
D
with
In
ce
ption
V
2,
Re
gi
on
-
Ba
sed
F
ully
Conv
olu
ti
onal
N
et
work
s
(R
-
FC
N)
with
Re
sn
et
101,
Faste
r
R
-
C
NN
with
Re
s
net
101
a
nd
Faste
r
R
-
CN
N
with
In
c
eptio
n
Re
s
net
v2.
I
n
our
pro
posed
ap
proa
c
h
Faste
r
R
-
CN
N
with
I
ncep
ti
on
Re
sn
et
v2
i
s
chosen
.
Fast
er
R
-
CN
N
has
achieve
d
m
uch
bette
r
s
pee
d
an
d
accuracy.
F
uture
m
od
el
s
f
ollow
e
d
var
i
ous
appr
oach
es
but
co
uld
outpe
rfor
m
Faste
r
R
-
CNN
by
a
sig
nificant
m
ar
gin
.
Faste
r
R
-
CNN
m
ay
n
ot
be
t
he
sim
pl
est
or
fastest
m
et
hod
f
or
ob
j
ec
t
detect
ion
,
but
it
is
sti
l
l
on
e
of
the
best
perform
in
g.
At
pr
ese
nt,
Faste
r
R
-
C
N
N
with
I
ncep
ti
on
Re
s
Net
m
od
el
of
Ten
sorfl
ow
is
the
slo
w
est
but
m
os
t acc
ur
at
e
m
od
el
.
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
Develo
pm
e
nt
of
fram
ew
or
k f
or detec
ti
ng s
m
okin
g
sce
ne
i
n vi
deo
cli
ps
(
Po
onam
G
)
23
2.
RELATE
D
W
ORKS
Re
search
w
ork
carrie
d
out
in
[1
]
e
xp
l
oits
the
Re
gion
P
rop
osa
l
Netw
ork
(
RPN)
of
the
F
ast
er
R
-
CN
N
m
od
el
to
detect
ped
est
ria
ns
.
Even
t
hough
R
-
CNN
’s
Re
gion
Pro
posal
Netw
ork
(R
PN
)
perf
or
m
s
well
,
the
res
ults
ca
n
be
de
gr
a
de
d
by
dow
ns
tream
cl
assifi
ers.
The
two
m
ai
n
reas
on
s
that
m
ay
l
ead
to
this
sit
ua
ti
on
are:
handlin
g
of
sm
al
l
instances
due
to
t
he
insuffici
ent
reso
l
ution
of
t
he
featu
res
a
nd
the
m
ining
of
hard
neg
at
ive
cases
is dif
ficult
due
near
ly
no prese
nece if
bo
otstr
apin
g
m
et
ho
do
log
ie
s t
o
ac
hieve
the
sam
e.
Anothe
r
rese
ar
ch
w
ork
c
ar
ried
out
in
[
2]
co
ncen
t
rates
on
t
he
dataset
s'
s
i
m
pact
on
dee
p
le
arn
in
g
a
nd
the
ap
plica
ti
on
an
d
t
he
im
po
rt
ance o
f
dee
p
le
arn
i
ng
thr
ough
Faste
r
R
-
CN
N
s.
T
he
wor
k
tri
es
to
s
umm
ariz
e
the
deep
le
ar
ning
al
gorithm
s
and
com
m
on
data
set
s
us
ed
in
t
he
fie
ld
of
c
ompu
te
r
visio
n.
A
dd
it
io
nally
,
the
stud
y
bu
il
ds
a
ne
wer
dataset
in
acc
orda
nce
to
the
pr
e
viously
ava
il
able
and
com
m
on
ly
us
ed
da
ta
set
s.
Faste
r
R
-
CN
N
is t
hen ap
plied
ov
e
r
t
his n
e
wl
y bu
il
t datase
t.
Applic
at
ion
of
the
faster
R
-
C
NN
is
e
xp
l
or
e
d
on
va
rio
us
be
nch
m
ark
s
on
wh
ic
h
the
al
go
rithm
has
a
pro
ven
im
pr
oved
resu
lt
s
ra
nging
from
obj
e
ct
detect
ion
to
face
rec
ogniti
on
in
[
3].
T
he
w
orke
d
ca
rr
ie
d
ou
t
pro
vid
es
t
he
re
su
lt
s
of
trai
ning
a
Faste
r
R
-
CNN
m
od
el
on
t
he
la
r
ge
scal
e
face
dataset
WI
DER
[
4].
T
he
w
ork
al
so
trie
s
to
exp
la
in
t
he
res
ults
on
WIDE
R
dataset
al
ong
with
tw
o
m
or
e
sta
te
of
t
he
art
and
wi
de
y
us
ed
dataset
s F
DD
B
and
IJB
-
A.
On
li
ne
ha
ndw
ritt
en
gra
ph
ic
s
m
a
y
con
ta
in
m
at
he
m
at
ic
a
l
expressi
on
s
an
d
diag
ram
s.
D
et
ect
ion
of
these
sym
bo
ls
consi
st
of
m
et
h
od
s
de
sig
ne
d
f
or
a
sin
gle g
ra
ph
ic
ty
pe
[
5].
I
n
this w
ork,
eva
luati
on
of
the
Faste
r
R
-
CNN
obj
ect
detect
io
n
al
go
rithm
as
a
ge
ne
ral
m
et
ho
d
for
detect
ion
of
sy
m
bo
ls
in
ha
ndwr
it
te
n
gra
phic
s
is
carried
ou
t.
Di
ff
e
ren
t
c
onfi
gurati
on
s
of
t
he
Faste
r
R
-
C
N
N
m
e
thod
are
e
valuated
,
a
nd
i
ssu
es
relat
ive
t
o
the
hand
wr
it
te
n
na
ture
of
the
data
are
pointe
d
ou
t.
Co
ns
i
der
i
ng
the
onli
ne
recog
niti
on
c
onte
xt,
eval
uation
of
eff
ic
ie
ncy
an
d
accuracy
trad
e
-
offs
of
usi
ng
Dee
p
Ne
ur
a
l
Network
s
of
diff
e
ren
t
com
plexiti
es
as
featur
e
extracto
rs
is c
a
rr
ie
d o
ut.
3.
DEEP L
EA
R
NING
MO
DE
LS FO
R
OB
JE
CT D
ET
EC
TION
Deep
Lear
ning
is
a
pa
rt
of
m
achine
le
a
rn
i
ng
wh
ic
h
giv
es
ou
tst
a
nd
i
ng
pe
rfor
m
ance
in
t
he
im
age
an
d
vid
e
o
cl
assifi
cat
ion
ta
sk
s
.
In
d
eep
le
ar
ning
t
her
e
a
re
var
i
ou
s
arch
it
ect
ur
e
s
includi
ng
r
ecu
r
ren
t
ne
ural
netw
ork
s
and
dee
p
ne
ur
a
l
netw
orks
w
hich
hav
e
m
ajo
r
app
li
cat
io
ns
in
the
fiel
d
of
c
om
pu
te
r
visio
n,
m
achine
tra
ns
l
at
io
n
and
natu
ral
la
ngua
ge
proce
s
sing.
Dee
p
ne
ur
al
net
wor
k
arch
it
ect
ure
has
m
ulti
ple
hid
de
n
la
ye
rs
betw
een
it
s
input
and
ou
t
put
la
ye
rs
wh
ic
h
are
fee
d
f
orward.
Data
fro
m
the
inp
ut
la
ye
r
flows
to
th
e
ou
tp
ut
la
ye
r
without
loopin
g
bac
k.
The
m
ai
n
app
li
cat
ion
of
com
pu
te
r
visio
n
in
vo
l
ves
ob
j
ect
detect
ion
wh
ic
h
has
a
m
ai
n
fo
cu
s
on
researc
h. The
progr
e
ss in
ob
j
e
ct
d
et
ect
ion i
s
m
ai
nly becaus
e of Co
nvol
ution
al
Ne
ur
al
Ne
tworks
(CN
N).
Object
detect
ion
is im
pr
ov
e
d f
ro
m
sing
le
ob
je
ct
to
m
ult
iple object
detect
ion
in rece
nt years. T
he first
can
detect
a
sing
le
ob
j
ect
in
an
i
m
age
wh
ic
h
can
be
us
e
d
for
cl
assifi
cat
ion
ta
s
ks
.
I
n
th
e
la
te
r
app
r
oac
h
not
on
ly
m
ulti
ple
obj
ect
s
in
a
n
i
m
age
are
dete
ct
ed
but
al
so
t
heir
e
xact
loc
at
ion
in
th
e
im
age
is
ind
ic
a
te
d
by
rectan
gu
la
r
bo
xes
or
m
asks.
Mov
i
ng
obj
ect
d
et
ect
ion
h
as be
en
ex
plore
d
e
xtensi
vely
by
va
rio
us
aut
hors
by
the
app
ly
in
g
dif
fe
ren
t
m
et
ho
ds
[6
-
7]
an
d
so
on.
But
m
os
t
of
the
pro
p
os
e
d
w
orks
ha
s
con
si
der
e
d
sta
ti
on
a
ry
ca
m
eras
or
a
sta
ti
on
ary
bac
kgr
ound.
T
he
CNN
a
rc
hitec
ture
is
co
ns
ta
nt
ly
i
m
pr
ov
in
g
from
ALexN
et
[8
]
,
the ZF Net
[9
]
,
the
VGG Net
[10], the
ResN
et
[
11]
starti
ng
from
the year 2012. A
n o
bj
ect
d
et
ect
ion al
go
rithm
was
f
or
m
ed
ba
sed
on
R
-
C
N
N
in
dee
p
le
a
rn
i
ng
a
nd
a
num
ber
of
well
-
kn
own
datase
ts
are
consi
de
red
to
i
m
pr
ove thes
e
al
gorithm
s w
it
h
im
pr
ovem
ent
in
the
acc
ur
ac
y of detect
io
n.
Var
i
ou
s
cha
ng
es
to
t
he
netw
ork
str
ucture
ha
s
im
pr
ov
e
d
t
he
dee
p
le
ar
ni
ng
that
the
netw
or
k
us
es.
T
h
e
m
os
t well
-
known
ser
ie
s
of al
gorithm
s f
or obj
ect
detect
io
n are
based o
n
R
-
CN
N wh
ic
h
i
nclu
de
the
foll
ow
i
ng
:
3.1
R
-
C
N
N (
Reg
i
on
-
Base
d
Convoluti
onal
N
eur
al
Net
w
orks)
Im
ages
in
the
dataset
are
la
be
ll
ed
or
t
he
re
gi
on
o
f
our
inter
est
is
cro
ppe
d
ou
t
f
r
om
the
i
m
age
and
t
his
crop
ped
reg
i
on
is
giv
e
n
as
input
to
the
c
onvoluti
on
neur
al
network.
Wh
en
t
hey
are
gi
ven
as
in
put
to
the
netw
ork,
a
pp
li
cat
ion
of
recta
ngular
boundi
ng
box
r
eg
ress
or
is
pr
ovi
ded.
Fo
r
th
e
cl
assifi
cat
ion
pur
pose
SV
M
is
us
e
d.
In
te
r
m
s
of
both
s
pa
c
e
an
d
ti
m
e,
trai
nin
g
bec
om
es
ve
ry
ex
pe
ns
i
ve.
The
ob
j
ect
detect
ion
in
im
ages
is
of
te
n
slo
w
a
nd
it
takes aroun
d on
e
m
inu
te
p
er
i
m
age.
3.2
F
as
t
R
-
C
NN
Fast
R
-
CN
N
ta
kes
the
c
ha
ra
ct
erist
ic
s
of
both
R
-
C
N
N
an
d
SP
P
-
Net.
Fa
st
R
-
CNN
ta
ke
s
the
entire
i
m
age
and
fe
at
u
re
m
ap
is
create
d
by
f
orwardin
g
it
to
the
conv
olu
ti
onal
la
ye
r.
The
n
re
gi
on
of
i
nterest
is
f
ou
nd
by
RoI
po
olin
g
la
ye
r.
RoI
la
ye
r
is
a
sing
le
SPP
la
ye
r
that
is
app
li
ed
on
top
of
co
nvol
ut
ion
al
la
ye
r
wh
ic
h
is
then
at
ta
ched
to
fu
ll
y
conne
ct
ed
la
ye
r.
Bou
ndin
g
bo
x
re
gr
ess
ors
an
d
so
ftm
ax
cl
assifier
s
are
ap
plied
for
cl
assifi
cat
ion
.
Ba
sed
on
both
boun
ding
box
re
gr
ess
ors
a
nd
s
of
tm
ax
cl
assifi
ers,
m
ulti
task
l
os
s
is
c
ompu
t
e
d.
By
this,
the
la
ye
r
belo
w
si
ngle
SPP
la
ye
r
is
m
ade
trai
na
ble
an
d
the
pr
ob
le
m
associat
ed
with
SPP
-
Net
is
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,
Vo
l.
13
, N
o.
1
,
Ja
nu
a
ry 20
19
:
2
2
–
2
6
24
so
lve
d.
It
high
er
detect
ion
qual
it
y
in
the
m
ai
n
i
m
pr
ove
m
ent
done
ove
r
R
-
C
NN
a
nd
SPP
-
Net.
He
re
al
l
the
la
ye
rs
can
be
updated
durin
g
trai
ning
pro
cess
an
d
it
do
es
not
require
the
featu
res
t
o
be
st
or
e
d
i
n
a
disk
.
The
Fa
st
R
-
C
NN
trai
ning
ti
m
e
is
9
tim
e
s
faster
w
he
n
c
om
par
ed
to
R
-
C
NN
w
hich
is
3
tim
es
faster
th
an
S
PP
-
Net
an
d
te
sti
ng
ti
m
e
req
uire
d
is
213
tim
es
faster
t
han
R
-
CNN
a
nd
whe
n
c
om
par
ed
t
o
SPP
-
Net
it
s
10
tim
es
faster.
A
l
ong wit
h
the
d
ec
rea
se in trai
ning ti
m
e, ther
e is i
nc
rease in
the le
ve
l of acc
uracy
.
3.3
F
as
ter
R
-
CNN
Re
gional
Pro
posal
Net
work
ta
kes
im
age
of
any
siz
e
as
in
pu
t
a
nd
outp
ut
a
set
of
obj
ec
t
proposals
each
with
ob
je
ct
ness
sc
or
e
[12
-
13]
.
SPP
-
Net
an
d
Fast
R
-
CNN
has
re
du
ce
d
the
exe
cution
ti
m
e
of
obj
ect
detect
ion
but
m
or
e
tim
e
is
req
ui
red
f
or
re
gi
on
al
pro
posal
.
Faste
r
R
-
CN
N
s
olv
es
this
pro
blem
by
us
i
ng
dee
p
netw
orks
for
tr
aditi
on
al
pr
act
i
ces
to
c
om
pu
te
a
pro
posal
box.
Faste
r
R
-
CN
N
c
on
sist
s
of
two
m
od
ules.
F
irst
is
the full
y co
nn
e
ct
ed
co
nvol
ution
al
netw
ork
a
nd the
seco
nd i
s Fast R
-
C
NN
de
te
ct
or.
4.
PROP
OSE
D MET
HO
DOL
OGY
The follo
wing
sect
ion
giv
es
t
he deta
il
s o
f
th
e m
e
tho
dol
og
y
u
se
d
i
n
the
pr
opos
e
d wor
k.
4.1
Ex
peri
ment
al Setup
The
detai
ls
of
hard
war
e
c
ho
sen
f
or
our
e
xp
e
rim
ent
are
show
n
in
Ta
ble
1.
Ex
pe
rim
ental
set
up
requires
f
or
w
hich
a
syst
e
m
with
m
ini
m
u
m
2
GB
of
N
VIDIA
G
PU
card,
CUD
A
a
nd
c
uDN
N
in
sta
ll
ed.
The
m
od
el
us
e
s
the
A
nacon
da
Pyt
hon
an
d
the
pyth
on
pac
kag
e
s
inclu
ding
Ten
sorfl
ow,
Op
e
nC
v,
m
at
p
lotl
ib
and p
a
ndas.
Table
1.
Hard
war
e
Re
quirem
ents
Hardwar
e co
m
p
o
n
en
ts
Co
n
f
i
g
u
ration
Proces
so
r
Proces
so
r
Sp
eed
RAM size
OS
GPU
VRAM
Intel i5
72
0
0
2
.6 GHz
8
GB
W
in
d
o
ws 1
0
NVID
IA
Ge
Fo
rce GTX 94
0
M
2
GB
4.2
D
atase
t
Hun
dr
e
ds
of
im
ages
are
require
d
to
trai
n
the
cl
assifi
er
fo
r
good
det
ect
ion
.
Vi
deos
con
ta
inin
g
sm
ok
ing
sce
ne
s ar
e c
ollec
te
d. These
vi
deo
s
con
ta
in
ciga
rett
es w
it
h diff
e
re
nt sh
a
pe, si
ze a
nd co
l
or
.
From
these
vid
e
os
5
f
ram
e
s
are
extracte
d
per
seco
nd.
A
ll
these
fr
am
es
are
conve
rted
to
200
X
200
JPEG
im
ages
wh
ic
h
form
s
ou
r
trai
ning
dataset
.
The
trai
ni
ng
da
ta
set
con
ta
ins
660
im
ages.
The
ci
ga
rett
es
in
the
i
m
ages
hav
e
var
ie
ty
of
li
ghti
ng
co
nd
it
io
ns
and
bac
kgr
ound
s
.
Als
o
ther
e
are
i
m
ages
i
n
w
hich
ci
gar
e
tt
e
is
par
ti
al
ly
seen
.
Using
Labeli
ng
too
l,
the
loc
at
ion
of
ci
ga
re
tt
e
in
an
i
m
age
is
m
ark
ed
by
dr
a
wing
rectan
gles.
The
locat
ion
of
ci
gar
et
te
is
stored
in
an
XML
file
wh
ic
h
co
nt
ai
ns
inform
at
i
on
a
bout
i
m
age
heigh
t,
width
and
the
co
ordi
nates
of
the
boundi
ng
rectan
gle
drawn
.
Each
im
age
siz
e
is
le
s
s
than
10
0K
B
since
the
tim
e
required
for
trai
ning
beco
m
es large
if the im
age size is hi
gh.
4.3
Tr
ainin
g
Trainin
g
the
m
od
el
f
or
detect
ing
the
ci
ga
rett
es
can
be
do
ne
on
G
oogle
cl
oud
se
rv
ic
es
,
CPU
or
GPU
.
Durin
g
trai
ning,
f
or
a
par
t
ic
ul
ar
tim
e
interval
Tenso
r
flo
w
stores
the
che
ck
po
i
nts.
Loss
is
repor
te
d
in
each
ste
p
of
trai
ni
ng.
T
he
lo
ss
repo
rted
by
the
m
od
el
is
a
com
bin
at
ion
of
cl
as
sific
at
ion
l
os
s
an
d
regressi
on
lo
ss.
Trainin
g
t
he
cl
assifi
er
is
sto
pped
w
hen
l
os
s
is
droppe
d
to
0.0
4.
T
he
la
te
st
checkp
oin
t
cre
at
ed
by
Te
ns
or
flo
w
at
a loss o
f 0.
4 i
s u
se
d for
detect
ion
the
ciga
r
et
te
s.
5.
RESU
LT
S
The
m
od
el
ca
n
input
in
3
different
form
s:
i
m
age,
vi
de
o
or
a
li
ve
we
bca
m
feed
.
Th
e
s
yst
e
m
resu
lt
s
are
e
valuate
d
usi
ng
im
ages
an
d
vid
e
os
.
First
the
s
yst
em
was
eval
uated
us
i
ng
differe
nt
dataset
s
with
di
fferent
i
m
ages.
Datas
et
1
co
ntain
10
im
ages
a
m
on
g
w
hich
5
im
ages
hav
e
ci
ga
rett
es.
Dataset
2
c
on
ta
in
20
i
m
ages
a
m
on
g
w
hich
10
im
ages
ha
ve
ci
ga
rett
es.
Dataset
3
co
ntain
30
im
ages
am
on
g
wh
i
ch
15
im
ages
ha
ve
ci
gar
et
te
s.
Acc
ur
acy
,
se
ns
it
iv
it
y
and
sp
eci
fici
ty
are
the
pe
rfor
m
ance
m
easur
e
s
co
ns
id
ered
t
o
eval
ua
te
the
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
Develo
pm
e
nt
of
fram
ew
or
k f
or detec
ti
ng s
m
okin
g
sce
ne
i
n vi
deo
cli
ps
(
Po
onam
G
)
25
m
od
el
with
th
ese
dataset
s.
T
he
r
esults
a
re
sh
ow
n
in
Ta
ble
2.
Since
data
set
s
2
a
nd
3
c
on
ta
in
im
ages
wh
e
re
ci
gar
et
te
is
part
ia
lly visible a
nd also
due t
o
i
ll
u
m
inati
on
c
ha
ng
e
s in
the im
age,
t
he
acc
ur
a
cy
is reduce
d.
Table
2.
E
val
ua
ti
on
Re
s
ults
Datasets
Pe
rf
o
r
m
an
ce
m
eas
u
res
Accurac
y
Sen
sitiv
ity
Sp
e
cif
icity
Dataset 1
90%
80%
100%
Dataset
2
75%
70%
80%
Dataset 3
76%
73%
80%
Nex
t t
he
m
od
el
is
evaluated
usi
ng
a
v
i
deo
da
ta
set
that con
ta
in 10 vide
os
. T
he
re
su
lt
s ar
e
di
sp
la
ye
d
i
n
Table
3.
T
he
prop
os
ed
app
ro
a
ch gives
an ave
rag
e
accu
racy
94.08 f
or
t
he v
ideo data
set
c
onside
red.
Table
3.
Res
ult
A
naly
sis
of
Videos
Sl.
No
.
Details o
f
con
f
u
sion
m
atrix
Fra
m
e
size
Tr
u
e po
sttiv
es
Tr
u
e neg
ativ
es
False
p
o
sitiv
es
False
n
eg
ativ
es
Accurac
y
(%)
Vid
eo
1
244
210
24
6
4
9
5
.90
Vid
eo
2
92
83
6
2
1
9
6
.74
Vid
eo
3
303
276
16
3
8
9
6
.37
Vid
eo
4
483
417
36
7
23
9
3
.79
Vid
eo
5
351
278
39
12
22
9
0
.31
Vid
eo
6
132
119
9
0
4
9
6
.97
Vid
eo
7
567
401
77
19
10
8
4
.30
Vid
eo
8
821
733
81
0
7
9
9
.14
Vid
eo
9
573
423
99
8
43
9
1
.10
Vid
eo
1
0
885
763
88
9
25
9
6
.16
Our
re
su
lt
s
a
re
com
par
e
d
with
the
sm
ok
in
g
eve
nt
de
te
ct
ion
rati
o
histo
gr
am
m
et
hod
[
13]
.
The
c
om
par
isi
on is s
how
n
in
Table
4.
Table
4.
C
om
par
isi
on of
Re
s
ul
ts
Dataset
Details o
f
con
f
u
sion
m
atrix
Fra
m
e
size
Tr
u
e po
sttiv
es
False
p
o
sitiv
es
False
n
eg
ativ
es
Accurac
y
(%)
Vid
eo
s co
n
sid
ered
in
his
to
g
ra
m
m
e
th
o
d
2196
1824
30
120
9
3
.2
Ou
r
v
id
eo
datas
et
4451
4003
66
207
9
3
.87
The
resu
lt
s
pro
ve
that
obj
ect
detect
ion
t
hro
ugh
faster
R
-
C
NN
can
be
us
e
d
f
or
detect
io
n
of
sm
ok
ing
scenes
by c
ons
iderin
g
ci
gar
et
t
e as a
n object.
6.
CONCL
US
I
O
N
The
res
ults
show
that
propo
sed
m
et
ho
d
can
be
a
dopted
for
dis
play
ing
warnin
g
m
ess
ages
duri
ng
sm
o
kin
g
s
cene
s.
Our
pro
pos
ed
wor
k
disp
l
ay
s
warni
ng
m
essage
by
de
te
ct
ing
ci
ga
rett
es.
T
his
wor
k
can
be
exten
ded
to
detect
s
m
ok
in
g
sc
enes
wh
ic
h
do
no
t
c
on
ta
i
n
a
c
igarett
e
but
ex
haling
the
sm
ok
e.
In
I
nd
ia
n
m
ov
ie
s
and
te
le
visi
on
sh
ows
,
a
sim
i
l
ar
ki
nd
of
m
essages
a
re
dis
play
ed
duri
ng
t
he
even
t
of
al
c
ohol
c
onsu
m
ption
f
or
this p
rop
os
e
d work ca
n be e
xt
end
e
d.
REFERE
NCE
S
[1]
Zha
ng,
Li
l
ia
ng
,
et
a
l.
“
Is
faste
r
R
-
CNN
doing
w
el
l
for
p
ede
stri
a
n
det
e
ct
ion
?
”
.
E
uropean
Confe
renc
e
on
Comput
er
Vi
sion.
Springer
,
Cham,
2016
;
44
3
-
457
.
[2]
Zhou,
Xin
y
i,
W
ei
Gong,
W
enL
ong
Fu,
Fengtong
Du.
“
Applic
at
io
n
of
dee
p
le
arn
i
ng
in
obje
ct
dete
ct
ion
”.
Compute
r
and
Information Sci
en
ce (
ICIS)
.
IEE
E/
ACIS
16
th Inte
rna
ti
ona
l
Co
nfe
ren
c
e
on
,
I
EEE.
2017
;
631
-
63
4.
[3]
H.
Jiang
and
E.
L
ea
rn
ed
-
Mill
er.
“
Face
De
tecti
on
wi
th
the
Faster
R
-
CNN
”.
2017
12th
IE
EE
Int
ernati
ona
l
Confe
renc
e
on
A
utomati
c
Face
&
Gesture
R
ec
ogn
it
ion
(
FG
2017)
.
W
ashingt
on,
DC,
2017;
650
-
65
7.
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,
Vo
l.
13
, N
o.
1
,
Ja
nu
a
ry 20
19
:
2
2
–
2
6
26
[4]
S.
Yang, P. Luo, C.
C
.
Lo
y
,
and
X.
Ta
ng
.
“
W
IDER
FA
CE:
A f
ace
de
tecti
on
ben
c
hm
ark
”,
C
VP
R
,
2016
.
[5]
F.
D.
Jul
ca
-
Agui
l
ar
and
N.
S.
T.
Hira
ta,
“
S
y
m
bol
Detect
ion
in
Online
Handwrit
t
en
Graphi
cs
Us
ing
Faster
R
-
CNN
”,
2018
13th
I
APR
Inte
rnational
W
orkshop on
Doc
ument
Ana
ly
sis
Syste
ms
(
DAS)
,
Vienna
,
Aus
tri
a
,
2018,
151
-
156.
[6]
Xu,
P.
“
Study
o
n
Moving
Objec
ts
b
y
Video
Monitori
ng
S
y
st
em
of
Rec
ognition
a
nd
Tra
ci
ng
Sch
e
m
e”
.
Indon
esian
Journal
of
Elec
t
rical
Engi
ne
erin
g
and
Computer
Sci
en
ce
(
IJE
ECS
)
.
2013;
11(9)
,
4
847
-
4854.
[7]
Mengxin
Li
,
Jin
gji
ng
Fan,
Ying
Zha
ng,
Rui
Zh
ang,
W
ei
ji
ng
Xu
,
Dingding
Hou1.
“
Moving
Obj
ec
t
Det
ec
t
ion
an
d
Tra
ck
ing
Algo
rithm
”.
Indone
sian
Journal
of
El
e
ctr
ic
al
Engi
n
ee
rin
g
and
Computer
Sci
en
ce
(
IJE
ECS
)
.
2013;
11(10),
5539
–
5544
.
[8]
A.
Krizh
evsk
y
,
I.
Suts
keve
r,
an
d
G.
Hinton.
“
I
m
age
Net
cl
assifi
ca
t
ion
with
deep
convol
uti
ona
l
neur
al
net
works
”.
NIPS
,
2012.
[9]
M.
D. Ze
i
le
r
and
R.
Fe
r
gus.
“
Visualizing and
und
ersta
nding
conv
olut
ional
n
eur
al
net
works
”. I
n
E
CCV
,
[10]
K.
Sim
ony
an
a
nd
A.
Zi
ss
erman.
“
Ver
y
deep
convol
uti
on
al
net
works
for
la
rge
-
sc
ale
ima
ge
rec
ogn
it
ion
”
.
ICLR
.
2015.
[11]
K.
He, X. Zha
ng
,
S.
Ren, a
nd
J.
Sun.
“
Dee
p
residu
al
le
arn
ing
for
i
m
age
re
cogni
t
io
n”.
C
VPR
,
2016.
[12]
Ren,
Shaoq
ing,
Kaiming
He,
R
oss
Girshick,
Ji
an
Sun.
“
Faster
R
-
CNN
:
towar
ds
rea
l
-
ti
m
e
obj
ec
t
detec
t
ion
wi
th
reg
ion
proposa
l
net
works
”
.
IE
EE
transacti
ons
on
pat
te
rn
an
aly
sis
and
mac
hine
int
e
ll
ig
ence
.
2017
;
39
(
6),
1137
-
1149.
[13]
W
u,
Pin,
Jun
-
We
i
Hs
ie
h,
Jiun
-
C
heng
Cheng,
Shy
i
-
Ch
y
i
Ch
eng,
Shau
-
Yin
Tseng.
“
Hu
m
an
s
m
oking
eve
nt
detec
tion
using
visual
intera
c
ti
on
c
lue
s”
.
20th
Inte
rnati
onal
Confe
renc
e
on
Pat
t
ern
R
ec
ogni
ti
on
(
ICPR
)
IEE
E
,
2010
;
4344
-
4347.
Evaluation Warning : The document was created with Spire.PDF for Python.