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.
1
70
~
1
78
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v1
3
.i
1
.pp
1
70
-
1
78
170
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Gazing
as actual
p
arameter for d
rowsin
ess assessm
ent
in d
riving s
imu
lato
rs
Art
h
ur
M
ou
ri
ts Rum
ag
i
t,
Iz
z
at
A
uli
a Akb
ar,
Mit
aku
Utsunomi
ya, T
a
ka
m
asa
Mo
ri
e
,
Tomohi
ko Ig
asaki
Facul
t
y
of
Adva
nce
d
Sc
ie
n
ce
an
d
Technol
og
y
,
K
um
amoto
Univer
sit
y
2
-
39
-
1
Kurokam
i,
Chuo
-
war
d,
Kum
amoto
860
-
8555,
Japa
n
,
telp/
fax
: +81
-
96
-
342
-
3613
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
ug
28
, 201
8
Re
vised
N
ov
2
1
, 2
01
8
Accepte
d
Nov
3
0
, 201
8
Man
y
tr
aff
i
c
a
ccide
nts
ar
e
due
t
o
drows
y
dr
ivi
n
g.
How
eve
r,
to
dat
e
,
onl
y
a
few
studie
s
ha
ve
bee
n
condu
ct
ed
on
th
e
ga
zi
ng
prope
rt
ie
s
rel
ated
to
drows
ine
ss
.
Thi
s
study
was
conduc
te
d
with
the
obje
ct
iv
e
of
esti
m
at
ing
th
e
rel
a
ti
onship
be
t
wee
n
ga
zi
ng
p
rope
rties
and
d
rows
ine
ss
in
thre
e
fa
cial
expr
ession
ev
aluati
on
(FE
E)
c
at
egor
ie
s:
al
er
t
(FEE
=
0)
,
li
g
htly
drows
y
(FEE
=
1−2)
,
h
ea
vi
l
y
drows
y
(
FEE
=
3−4)
.
D
rows
ine
ss
was
inve
stiga
t
ed
base
d
on
the
se
e
y
e
-
g
azing
prope
rti
es
b
y
anal
y
zi
n
g
the
gazing
signal
utilizin
g
an
e
y
e
ga
ze
trac
ker
and
FEE
in
a
drivi
n
g
sim
ulator
environm
ent
.
The
result
s
obta
in
ed
indicate
tha
t
g
azing
prope
rties
have
si
gnifi
c
ant
diff
erence
s
among
the
three
drows
ine
ss
condi
ti
ons,
with
p
<
0.
00
1
in
a
Krus
kal
–
W
al
li
s
te
st
.
Furthermore,
th
e
over
al
l
class
ifi
ca
ti
on
ac
cu
racy
of
the
thr
ee
drows
ine
ss
condi
ti
ons
b
ase
d
on
gaz
ing
pro
per
ties
using
a
support
vec
tor
m
ac
hine
was
76.
3%.
Thi
s
ind
ic
a
te
s
th
at
our
proposed
gazing
prope
rties
c
an
be
used
to
quant
itati
v
ely
as
sess
drows
ine
ss
.
Ke
yw
or
d
s
:
Dr
i
ving sim
ula
tor
Drowsi
nes
s
Ey
e g
aze t
racke
r
Gazin
g
Suppor
t
v
ect
or m
achine
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
:
Tom
oh
ik
o
I
gas
aki,
Faculty
of
Adv
anced Sci
en
ce
and Tec
hnolog
y,
Ku
m
a
m
oto
Un
iversity
,
2
-
39
-
1
Kur
ok
a
m
i, Chuo
-
ward
, Kum
a
m
oto
860
-
8555, J
a
pa
n, te
lp/fa
x: +81
-
96
-
342
-
3613.
Em
a
il
:
iga@cs.kum
a
m
oto
-
u.
a
c.jp
1.
INTROD
U
CTION
Drowsi
ness
durin
g
dri
ving
is
a
sever
e
prob
l
e
m
and
is
belie
ved
t
o
be
a
di
rect
con
tri
bu
ti
ng
ca
us
e
of
traff
ic
acci
de
nts
[1
,
2].
It
plac
es
the
li
ves
of
dr
i
ver
s
an
d
pa
s
sen
ger
s
at
risk
and
ca
n
cause
serio
us
acci
de
nts
on
m
ajo
r
r
oa
ds
.
Accor
ding
to
a
U.S.
Nati
on
al
Highway
T
raf
fic
Sa
fety
Ad
m
inist
rati
on
(
NH
T
SA)
re
port
in
201
7
[
3],
dro
w
siness
an
d
fall
ing
asl
ee
p
w
hile
dr
ivi
ng
w
as
r
esp
on
s
i
ble
for
at
le
ast
10
0,000
autom
ob
il
e
crash
e
s
and
84
6
deat
hs
within
a
y
ear.
T
he
Nati
on
al
Po
li
ce
A
gen
cy
of
Ja
pa
n
al
s
o
release
d
data
s
howing
that
appr
ox
im
at
ely
434
,
000
tra
ff
ic
acci
den
ts
occ
urred
in
201
7
[4
]
.
P
rev
i
ou
s
s
tud
ie
s
the
or
iz
e
d
that
the
caus
es
of
acci
d
ents
m
ig
ht
be
relat
ed
to
facto
rs
su
c
h
as
la
ck
of
co
ncen
t
rati
on
durin
g
dr
i
ving
and
poor
dr
i
ving
sk
il
ls.
Howe
ver, th
ose
shor
tc
om
ing
s can be
recti
fie
d by im
pr
ov
i
ng
dr
i
ver
a
wa
re
ness
a
nd
dr
i
ving s
kill
s.
Var
i
ou
s
m
et
ho
ds
for
detect
ing
dr
ow
si
nes
s
hav
e
been
pro
po
se
d.
Among
the
m
os
t
popu
la
r
is
i
m
ple
m
enting
a
traject
or
y
se
ns
or
i
ns
ide
t
he
ta
rg
et
veh
ic
le
[5,
6].
This
se
ns
or
m
easur
es
the
m
agn
it
ud
e
of
t
he
ste
ering
wh
eel
ang
le
an
d
it
s
velocit
y,
as
well
as
the
frequ
e
ncy
with
wh
ic
h
the
dro
wsy
dr
i
ver
c
orrectl
y
po
sit
io
ns
the
s
te
ering
wh
eel
ang
le
.
Plac
i
ng
the
sensor
ins
ide
the
ve
hicle
is
m
or
e
conv
enient
f
or
the
dr
i
ver
instea
d
of
at
ta
chin
g
it
to
th
e
dri
ve
r
direct
ly
.
Howe
ver,
the
r
oa
d
s
urfa
ce
an
d
c
onditi
on
m
ay
red
uc
e
th
e
detect
ion acc
ur
acy
o
f
the se
nsor
.
We
the
or
iz
e
t
hat
dro
wsin
es
s
it
sel
f
is
strongly
relat
ed
to
the
physi
cal
conditi
on
of
a
pe
rson,
and
the
refor
e
,
drowsi
ness
det
ect
ion
ca
n
be
im
pr
ov
e
d
by
directl
y
inv
est
ig
at
ing
t
his
c
ondi
ti
on
.
Fun
dam
e
ntall
y,
the
act
ual
sta
t
e
of
the
hum
a
n
body
is
us
ua
ll
y
deter
m
ined
by
placi
ng
el
ect
rodes
or
bio
-
se
nsors
on
t
he
body
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
Ga
zi
ng as
actu
al par
amet
er for dr
owsiness
as
sess
me
nt in drivi
ng sim
ula
t
or
s
(
Arth
ur Mo
ur
it
s R
uma
git
)
171
it
sel
f.
Howe
ve
r,
seve
ral
pre
vious
stu
dies
hav
e
repo
rted
ano
t
her
ap
pr
oach
for
dete
ct
ing
dro
wsi
ne
ss
that
involves
us
in
g
hu
m
an
bio
lo
gical
sign
al
s,
su
c
h
as
ey
e
m
ov
e
m
ents
and
ey
e
blink
in
g
obta
ined
usi
ng
a
n
el
ect
ro
oc
ulog
r
a
m
(EOG)
[7,
8],
hea
rtbeats
us
in
g
a
n
el
ect
r
ocardio
gr
am
(
ECG)
[
9
-
11
]
,
br
ai
n
act
ivit
y
us
in
g
a
n
el
ect
ro
ence
pha
logram
(EEG
)
[12
-
14
]
,
m
on
it
or
i
ng
m
us
cl
e
a
ct
ivit
y
us
ing
an
el
ect
ro
m
yogra
m
(EMG)
[15
,
16
]
,
and
al
s
o
pu
lse
rate
act
ivit
y
[1
7].
H
oweve
r,
sk
in
co
ntact
by
el
ect
ro
des
or
a
bio
-
se
nso
r
cou
l
d
cause
dr
i
ver
discom
fo
rt
dur
ing
dri
vi
ng.
A
no
t
her
a
pprop
r
ia
te
m
et
ho
d
f
or
detect
ing
a
nd
est
i
m
a
ti
ng
the
dr
owsi
ness
sta
te
of
a
dr
i
ver
with m
inim
a
l or
no s
kin
c
on
ta
ct
is t
he
refor
e
n
ee
de
d
.
An
al
te
r
native
m
et
ho
d
of
es
tim
a
ti
ng
dro
w
siness
is
by
u
sing
a
cam
era
to
record
ey
e
beh
a
vior.
Pr
e
vious stu
dies r
ep
or
te
d
tha
t i
t
is p
os
sible t
o
detect
d
r
owsiness u
si
ng
num
ero
us
less
-
int
ru
si
ve
te
chn
i
ques an
d
m
ini
m
iz
ing
sk
i
n
co
ntact
by
placi
ng
a
cam
era
in
fr
ont
of
the
dr
ive
r
to
capt
ur
e
the
face
an
d
ey
e.
For
inst
an
ce,
ey
el
id
m
ov
em
ent
[
18
]
, g
aze and
h
ea
d
[
19,
20
]
,
ey
e
trac
ki
ng
a
nd p
upil
posit
ion
[
21]
, f
a
ce
expressi
on
d
et
ect
ion
[22],
face
e
xpr
ession
m
on
it
ori
ng
[
23
]
,
blin
k
detect
ion
[
24
]
,
ey
e
sta
te
analy
sis
[25,
26]
,
po
rtion
of
ey
e
cl
os
ure
[27],
a
nd
ey
el
id
cl
osu
re
[28
]
hav
e
been
inv
e
sti
gated.
T
hese
m
et
ho
ds
ha
d
t
he
sam
e
goal
of
pr
ovidin
g
inf
or
m
at
ion
rel
at
ed
to
t
he
s
ubje
ct
/driv
e
r
co
nd
it
ion
w
hile
pe
r
form
ing
var
i
ous
ta
sk
s o
r
un
de
r
va
rio
us
c
ondi
ti
on
s
(e.
g.
,
rest,
fati
gu
e
,
an
d
dro
w
siness)
.
H
ow
e
ver,
al
though
pr
e
vious
ap
pr
oach
e
s
co
uld
est
i
m
at
e
the
dr
owsy
conditi
on,
ther
e
wer
e
draw
ba
cks
su
c
h
as
th
e
necessit
y
to
pro
vid
e
a
cl
ear
view
an
d
sta
bl
e
po
sit
ion
i
ng
of
th
e
ca
m
era dur
i
ng
the r
ec
ordin
g p
ro
ces
s.
Our
pro
po
se
d
syst
e
m
e
m
plo
ys
an
ey
e
t
racke
r
se
ns
or
m
ounted
on
the
he
ad
to
obta
in
e
ye
prop
e
rtie
s
durin
g
dr
i
ving.
W
e
c
onfirm
ed
that
this
kind
of
a
rr
a
ngem
ent
has
rar
el
y
be
e
n
us
ed
to
date,
even
th
ough
a
head
-
m
ou
nted
ey
e
tracker
ca
n
ove
r
com
e
view
and
po
sit
io
n
lim
it
a
ti
on
s
w
hile
eva
luati
ng
the
dr
i
ve
r’
s
ey
e
pr
op
e
r
ti
es.
As
previ
ously
descr
i
bed,
m
o
st
stud
ie
s
us
e
d
the
subj
ect
’s
e
ye
and
facial
m
ov
e
m
ent
i
m
a
ges
to
e
valuate
their
conditi
on,
esp
eci
al
ly
dr
ow
si
ness.
Howe
ver,
we
co
uld
not
find
a
ny
cl
ear
inform
ation
on
how
to
util
i
ze
the
gazin
g
of
the
dr
i
ver
to
deter
m
ine
his/her
c
onditi
on.
I
n
th
is
stud
y,
we
f
oc
us
e
d
on
ey
e
-
gazin
g
beca
use
of
the
lim
it
ed
extent
to
w
hich
it
has
been
util
iz
ed.
To
e
valuate
th
e
dro
wsin
ess
c
onditi
on,
we
e
valuated
the
s
ubj
ect
’s
conditi
on
us
i
ng
facial
ex
pres
sion
e
valuati
on
(
FEE)
[
29
]
i
n
acco
r
dan
ce
with
the
e
xper
i
m
ent’s
locat
io
n
an
d
env
i
ronm
ent.
We
util
iz
ed
this
evaluati
on
m
et
hod
with
the
obj
ect
ive
of
obser
ving
t
he
a
ct
ual
co
nd
it
io
n
of
t
he
su
bject
by c
onsiderin
g seve
ra
l po
i
nts
of
vie
w wit
h
the
sam
e sour
ce
in
for
m
at
ion
.
Th
us
,
we
est
i
m
at
ed
the
relat
ion
s
hip
betwee
n
dro
wsin
e
ss
a
nd
gazin
g
pa
ra
m
e
te
rs
in
three
cat
ego
ries
of
dro
wsin
es
s.
W
e
hy
pothesi
zed
that
these
three
cat
eg
ori
es
ha
ve
a
str
ong
relat
io
nship
with
the
ey
e
-
ga
zi
ng
pro
per
ti
es,
es
pecial
ly
fo
r
est
i
m
ating
the
co
nd
it
io
n
be
fore
act
ual
drowsi
ness
to
preve
nt
acci
den
ts
.
We
in
vestigat
e
d
wh
et
her
e
ac
h
feat
ur
e
of
th
e
gazin
g
pro
pe
rtie
s
has
a
sig
ni
ficant
dif
fer
e
nc
e
an
d
e
xam
in
ed
th
e
perform
ance
of
relat
ed
featu
r
es
us
in
g
a
s
up
port
vect
or
m
achine
(SVM).
Finall
y,
we
co
nf
irm
ed
w
heth
er
the
gazin
g
si
gn
al
c
ou
l
d be
us
e
d
as
an
act
ual
par
a
m
et
er to
assess
drowsine
s
s
w
hile d
riving.
2.
METHO
D
S
2
.
1.
S
ubj
e
cts
Ele
ven
healt
hy
m
al
es
of
ages
in
the
range
of
21−
35
ye
a
rs
pa
rtic
ipate
d
in
this
stud
y.
Be
fo
re
th
e
exp
e
rim
ent,
wr
it
te
n
inform
e
d
co
ns
ent
f
or
this
stud
y
was
ob
ta
ine
d
fro
m
each
par
ti
cip
ant.
T
he
pa
rtic
ipants
wer
e a
sk
e
d
t
o get su
ff
ic
ie
nt s
le
ep
duri
ng the
n
ig
ht a
nd
ha
ve
their lu
nc
h bef
or
e
pa
rtic
ipati
ng in
t
he
e
xperi
m
ent.
They
wer
e als
o aske
d n
ot to
c
on
s
um
e alc
ohol
o
r
caf
feine
be
fore the
expe
rim
ent.
2
.
2
.
T
as
ks
We
us
e
d
a
dr
i
ving
sim
ulator
(D
A
-
1110,
H
onda
Mot
or
,
J
apan),
ey
e
gaz
e
tracker
(
Tal
kEye
Lit
e,
Takei
Scie
ntifi
c
In
str
um
ents,
Japan),
web
c
a
m
era
(HD
Pro
W
e
bcam
C92
0,
Lo
gico
ol,
China)
,
com
pute
r
to
record
t
he
dri
ver’s
facial
expressi
on,
an
d
dri
ving
sim
ulator
syst
em
con
tr
ol,
as
sh
ow
n
in
Fig
ur
e
1.
Each
sub
j
ect
was
aske
d
to
dr
i
ve
on
the
oval
track
wit
hout
obsta
cl
es
durin
g
the
day
tim
e
in
an
aut
om
atic
transm
issi
on
c
ar
w
hile
m
ai
nt
ai
nin
g
a
s
peed
of
100
km
/h
for
50
m
inu
te
s.
The
e
xperim
ent
was
sc
he
du
le
d
twic
e
per
day
from
8
:00
am
to
10:0
0
am
an
d from
1
:0
0 pm
to
3:00
pm
, r
especti
ve
l
y. Th
us
, eac
h su
bj
e
ct
p
a
rtic
ipate
d
in eigh
t t
rial
s
duri
ng
t
he
ex
pe
r
i
m
ent o
n
di
ff
e
r
ent d
ay
s. All
pro
ce
dures use
d i
n
this stu
dy were ap
pro
ved
by
the
Ethic
al
Com
m
i
tt
ee of
t
he
Fac
ulty
o
f
Ad
van
c
ed
Scie
nce a
nd
Tech
nolo
gy, Kum
a
m
oto
Un
iversity
.
2
.
3
.
Reco
r
dings
2
.
3
.
1
P
hy
si
ological
Measure
ment
We
m
ou
nted
t
he
ey
e
gaze
tr
acker
,
as
show
n
in
Fig
ur
e
2,
on
th
e
hea
d
to
ob
ta
in
a
nd
re
cord
the
ey
e
gaze
sig
nal
at
a
sa
m
pling
rate
of
3
0
Hz.
I
n
a
dd
it
io
n,
the
s
ubj
ect
’s
face w
a
s
recorde
d
us
in
g
the
we
b
cam
era
f
or
ps
yc
holo
gical
m
ea
su
rem
ent.
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
:
1
7
0
–
1
7
8
172
Figure
1. Re
se
arch en
vir
on
m
ent
Figure
2
.
Eye
gaze trac
ke
r
2
.
3
.
2
Ps
ycholo
gical
Measure
ment
FEE
is
perfor
m
ed
with
evaluati
on
from
di
ff
e
ren
t
pe
rs
pec
ti
ves
by
evalu
at
or
s.
T
he
eva
luators
m
us
t
evaluate
the
s
ubj
ect
s
’
sta
te
wi
th
the
sam
e
facial
exp
res
sio
n
recordi
ng
sou
rce
duri
ng
dri
vi
ng
.
Eval
uato
rs
al
so
need
to
m
at
ch
their
judgm
e
nt
w
hile
pro
vi
ding
the
eval
ua
ti
on
a
nd
stri
ve
f
or
the
pe
r
cepti
on
of
dif
fer
e
nt
evaluato
rs
to
be
relat
ively
si
m
i
la
r.
T
he
m
at
ching
of
pe
r
cepti
ons
is
ca
r
ried
out
jointl
y
at
the
e
nd
of
the
evaluati
on.
Theref
or
e
, we e
xpect
ed FEE t
o pro
vid
e a
reli
a
ble m
easur
e of
the con
diti
on of the
sub
j
ect
.
FEE
was
us
e
d
to
e
valuate
e
ach
s
ubj
ect
’s
drowsi
ness
co
nd
it
io
n.
It
c
on
sist
s
of
a
five
-
le
vel
(0
−
4)
drowsi
ness
qu
est
ionnaire,
i
n
wh
ic
h
e
ver
y
nu
m
ber
r
ep
res
ents
a
dr
ow
si
ne
ss
de
gr
ee
,
f
rom
ver
y
al
ert
to
ve
ry
sle
epy
.
O
n
c
om
ple
ti
on
of
th
e
exp
e
rim
ent,
four
e
xam
iners
evaluate
d
ea
ch
s
ubj
ect
’s
dro
wsin
es
s
co
ndit
ion
ever
y
e
poch
(
1
ep
och
= 30
s
),
acco
rd
i
ng
to
the
F
EE questi
onnai
re
s
how
n
i
n
Ta
ble 1
,
by w
at
chin
g
the vi
deo
of
the
subj
ect
’s
f
aci
al
exp
ressi
on
rec
orde
d
by
the
web
cam
era
w
hile
the
su
bject
wa
s
dr
iving.
The
fin
al
FEE
evaluati
on sc
or
e w
as
d
eci
ded
by the m
ajorit
y v
ote of t
he
f
ou
r
e
xam
iners.
Be
fore
an
d
a
fter
each
t
rial
,
e
ach
s
ubj
ect
w
as
instr
ucted
t
o
m
ai
ntain
a
resti
ng
sta
te
in
the
dri
vi
ng
si
m
ulator’
s
se
at
for
5
m
in.
The
n,
the
sub
je
ct
s
wer
e
a
ske
d
to
dri
ve
for
50
m
in
in
the
dr
ivi
ng
sim
ula
tor
,
and a
vid
e
o of
each s
ubj
ect
’s face rec
orde
d.
Table
1.
Faci
al
Ex
pr
e
ssio
n
E
va
luati
on
an
d
it
s
Crit
eria
Grade
Drows
in
ess
stag
e
Actio
n
cr
iteri
a
0
No
t dro
wsy
Qu
ick
and
f
requ
en
t ey
e
sh
if
t,
activ
e b
o
d
y
m
o
v
e
m
en
t
1
So
m
ewh
at dro
ws
y
Op
en
lip, slo
w e
y
e
m
o
v
e
m
en
t
2
Drows
y
Slo
w and
f
requ
en
t
ey
eb
lin
k
,
m
o
u
th
m
o
v
e
m
en
t
3
Qu
ite dro
wsy
Co
n
scio
u
s ey
eb
lin
k
,
h
ead s
h
ak
e,
f
requ
en
t y
awn
4
Ver
y
dro
wsy
Clo
se ey
elid
,
h
ead
tilt f
o
rwar
d
or f
all
b
eh
in
d
2
.
4
.
Analyses
2
.
4
.
1
Fe
ature
Extr
act
i
on
Be
fore
e
xtracti
ng
ga
zi
ng
pro
pe
rtie
s,
the
th
re
sh
ol
d
t
o
judge
the
gazin
g
had
to
be
determ
ined
.
Ga
zi
ng
was
co
ns
i
der
e
d
as
a
feat
ur
e
wh
e
n
the
m
ov
i
ng
s
pee
d
was
m
ai
ntained
bel
ow
t
he
consi
der
e
d
th
r
esh
old
.
To
c
onfirm
the
op
ti
m
u
m
thresh
ol
d,
we
cal
cul
at
ed
the
num
ber
of
f
ram
es
in
wh
ic
h
gazi
ng
occurre
d
(1
fr
a
m
e
=
1/30
s)
pe
r
ep
oc
h
by
us
in
g
th
e
m
ini
m
u
m
and
m
axi
m
u
m
th
reshold
.
By
co
ns
ide
rin
g
the
m
axi
m
u
m
su
m
of
t
he
diff
e
re
nces
(SOD
)
value
as
t
he
op
ti
m
u
m
threshold
ca
nd
i
da
te
,
we
cal
cula
te
d
the
S
OD
of
the
f
ram
es
in
w
hich
gazin
g
occ
urre
d,
as
s
how
n
i
n
Fig
ure
3.
I
n
this
ex
per
im
e
nt,
a
th
reshold
of
2−3
de
g/s
was
f
ound
to
be
the
m
axi
m
u
m
SO
D
val
ue.
Acc
ordin
g
to
t
hat
conditi
on,
we
therefo
re
ch
ose
2
or
3
deg/
s
as
our
final
SOD
t
hr
es
ho
l
d
ca
nd
idate
.
D
uri
ng
t
his
e
xp
e
rim
ent
,
ba
sed
on
t
his
data,
the
le
ft
side
of
t
he
t
hresh
old
ca
ndida
te
was
the
th
res
ho
l
d
1−2
de
g/s
a
nd
the
rig
ht
sid
e
was
the
thre
s
ho
l
d
3
=
4
deg
/
s.
I
f
t
he
neig
hbori
ng
ga
p
fro
m
th
e
m
axi
m
u
m
SO
D
value was cl
os
er to
t
he
le
ft,
t
hen
2
deg
/s
be
ca
m
e the o
ptim
u
m
thr
esh
old;
o
ther
wise, if
the g
a
p
from
the
m
axim
u
m
SO
D
of
t
he
ca
nd
i
date
of
the
final
thre
s
ho
l
d
value
was
cl
os
er
to
t
he
righ
t,
the
n
a
thre
sh
ol
d
of
3
deg
/s
bec
a
m
e
the
op
ti
m
um
threshold.
In
t
his
case,
from
the
op
ti
m
u
m
SO
D
can
dida
te
,
a
thres
hold
of
2
deg
/s
was
cons
idere
d
as t
he o
pti
m
u
m
thr
esh
old
.
T
her
e
fore,
w
e
us
e
d
a t
hr
e
sh
ol
d value
of
2 deg
/s a
s
our gazi
ng
occurre
nce
t
hresh
old
.
Eve
ry
su
bject
ha
d
a
diff
e
re
nt
opti
m
um
threshold
i
n
eac
h
tria
l.
In
a
total
of
88
tria
ls,
there
wer
e
31
tr
ia
ls
with
an
op
ti
m
u
m
thresh
ol
d
of
2
de
g/
s,
30
tria
ls
wit
h
an
optim
u
m
thres
ho
l
d
of
3
deg
/s
,
and 27 t
rial
s wi
th an
opti
m
u
m
thr
es
hold
of
4 deg
/s
.
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
Ga
zi
ng as
actu
al par
amet
er for dr
owsiness
as
sess
me
nt in drivi
ng sim
ula
t
or
s
(
Arth
ur Mo
ur
it
s R
uma
git
)
173
The
gazin
g
s
ign
al
was
ge
ner
at
e
d
by
m
ov
i
ng
sp
ee
d.
On
obta
inin
g
the
op
ti
m
um
threshold
,
we
co
ns
tr
ucted
and
extract
e
d
the
gazin
g
sig
na
l,
as
sh
own
in
Figu
re
4.
Wh
e
n
the
m
ov
ing
s
peed
was
le
ss
than
the
thres
ho
l
d
and
m
or
e
than
z
ero,
gazin
g
oc
curred
.
In
c
on
t
rast,
w
hen
the
m
ov
ing
sp
ee
d
was
great
er
th
an
the
thres
ho
l
d,
the
r
e
was
no
gazi
ng.
Wh
e
n
the
m
ov
in
g
sp
ee
d
was
ze
ro,
bl
ink
in
g
occ
urr
ed.
I
n
this
e
xam
ple,
a
thres
ho
l
d
of
2
de
g/s
was
us
e
d.
Th
e
nu
m
ber
of
gazin
g
occ
urre
nces,
blink
i
ng
occurre
nces,
or
non
-
gazi
ng
occurre
nces
co
uld
be
cal
c
ulate
d
on
a
f
ram
e
by
f
ram
e
basis.
The
co
ntinuo
us
occurre
nces
durin
g
a
certai
n
per
i
od
wer
e
c
ounte
d
as
a
cl
ust
er.
Nine
featu
res
of
the
gazing
si
gn
al
,
li
ste
d
in
Ta
ble
2,
c
ou
l
d
be
e
xtract
ed
an
d
com
pu
te
d
e
very
ep
oc
h. T
he p
ro
ces
s was r
ep
eat
ed
f
or all
th
e featu
res
i
n
ea
ch
tria
l.
Table
2
. Fea
tu
r
es
Ext
racted
from
Gazing
Sig
nal
Para
m
eter
Ab
b
rev
Featu
re
Gazing
f
ra
m
e
GF
Nu
m
b
e
r
o
f
f
ra
m
es
in
which
gazin
g
occu
rr
ed
per
ep
o
ch
Gazing
clus
ter
GC
Nu
m
b
e
r
o
f
clus
ters
in wh
ich
gazin
g
occu
rr
ed
per
epo
ch
No
n
-
g
azin
g
f
ra
m
e
NF
Nu
m
b
e
r
o
f
f
ra
m
es
in
which
no
n
-
g
azin
g
occu
rr
ed
per ep
o
ch
No
n
-
g
azin
g
clus
ter
NC
Nu
m
b
e
r
o
f
clu
sters
in wh
ich
no
n
-
g
azin
g
occu
rr
ed
per ep
o
ch
Blin
k
f
ra
m
e
BF
Nu
m
b
e
r
o
f
f
ra
m
es
in
which
blin
k
s o
c
cu
rr
ed
per
ep
o
ch
Blin
k
clus
ter
BC
Nu
m
b
e
r
o
f
f
ra
m
es
in
which
blin
k
s o
c
cu
rr
ed
per
ep
o
ch
Ratio
of
gazin
g
f
ra
m
e
s v
s. c
lu
sters
RG
GF/GC
Ratio
of
no
n
-
g
azin
g
f
ra
m
es v
s. c
lu
sters
RN
NF/NC
Ratio
of
blin
k
f
ram
e
s v
s. c
lu
sters
RB
BF/B
C
2
.
4
.
2.
Statis
tics
Be
fore
co
nduc
ti
ng
sta
ti
sti
cal
analy
sis,
we
inv
est
igate
d
w
het
he
r
each
fe
at
ur
e
co
rr
el
at
e
d
with
the
conditi
on
of
th
e
subj
ect
by
usi
ng
F
EE.
T
he
n,
a
K
olm
og
oro
v
–
Sm
irno
v
te
st
was
us
ed
t
o
e
xam
ine
wh
et
he
r
the
gazin
g
sig
nal
sh
owe
d
a
norm
al
distribu
ti
on.
On
e
-
way
A
N
OVA
analy
sis
was
us
e
d
if
the
distribu
ti
on
da
ta
ha
d
a
no
rm
al
di
stribu
ti
on;
oth
e
r
w
ise
,
W
il
coxo
n
-
rank
su
m
analy
sis
was
us
ed
.
The
res
ults
of
the
featur
e
e
xtr
act
ion
process
we
re
di
vid
ed
into
th
r
ee
cat
egories:
al
ert
(F
EE
=
0),
li
ghtl
y
dro
w
sy
(F
EE
=
1−
2),
a
nd
hea
vily
drowsy
(F
EE
=
3−
4)
.
Af
te
r
div
i
ding
the
gazi
ng
si
gnal
into
t
hese
three
cat
e
gories
,
sta
ti
sti
cal
analy
sis
was
perf
orm
ed
to
in
vestigat
e
t
he
sig
nifica
nt
diff
e
re
nces
within
t
he
th
ree
c
at
egories.
A
va
lue
of
p
<
0.0
5
was
c
onside
re
d
to
be
sta
ti
sti
cally
sign
ific
ant.
2.4.
3
.
Classi
fic
at
i
on
In
our
stu
dy,
a
n
S
VM
was
use
d
as
a
cl
assifi
er
to
co
nduct
perform
ance
evaluati
on
of
th
e
featu
res
i
n
the
three
cat
e
gories
—
al
e
rt
(FEE
=
0),
li
ghtl
y
drow
sy
(F
E
E
=
1−
2)
,
an
d
heav
il
y
dro
ws
y
(F
EE
=
3−4)
—
by
us
in
g
t
he
L
IB
SV
M
li
brary
[
30
]
,
w
hich
has
al
so
bee
n
util
iz
ed
by
A
kbar
et
al
.
[31].
T
he
feat
ures
we
re
first
com
bin
ed
int
o
one
dataset
;
t
hen,
on
e
half
was
us
ed
to
m
ake
the
t
rainin
g
da
ta
an
d
the
oth
e
r
half
the
te
sti
ng
data
(trainin
g
set
:
50
%,
te
st
set
:
50
%)
.
Th
e
aver
a
ge
per
c
entage
of
total
true
detect
ion
fr
om
4
-
f
old
c
ro
ss
-
validat
io
n
was
us
e
d
as
a
m
ea
su
re
of
cl
assifi
cat
ion
acc
ur
ac
y.
A
ra
dial
bas
is
functi
on
(R
BF)
wa
s
us
e
d
as
th
e
SV
M
kernel
f
unct
ion.
T
he
be
st
value
of
cos
t
and
gam
m
a
par
am
et
er
of
the
RB
F
ke
rn
el
w
as
set
autom
at
i
cal
ly
by u
si
ng L
IBS
VM.
To
op
ti
m
iz
e
t
he
cl
assifi
cat
ion
proc
ess,
w
e
us
e
d
the
S
VM
rec
ursive
featu
re
el
im
i
nation
(RF
E
)
m
et
ho
d
f
or
each
sub
j
ect
,
w
hi
ch
is
a
wr
ap
pe
r
-
base
d
m
e
tho
d.
The
S
VM
RFE
was
de
ve
lop
e
d
by
G
uyon
et
al
.
[32]
an
d
has
be
en
use
d
in
ge
ne
sel
ect
ion
for
cance
r
cl
assif
ic
at
i
on
,
an
d
by
Ebr
a
him
i
e
t
al.
[33]
f
or
a
utom
at
ic
sle
ep
sta
gi
ng.
The
ste
ps
in
th
e
SV
M
RF
E
fe
at
ur
e
sel
ect
io
n
al
gorithm
us
ed
in
this
st
ud
y
wer
e
a
s
f
ollo
w
s:
first,
on
e
featu
re
w
as
rem
ov
ed,
a
nd
t
he
accu
ra
cy
com
pu
te
d.
Subseque
ntly
,
the
featu
re
tha
t
con
trib
uted
t
o
t
he
highest
accu
ra
cy
was
el
im
inate
d.
T
he
fe
at
ure
el
i
m
inate
d
in
the
pre
vious
s
te
p
co
uld
be
use
d
in
the
nex
t
ste
p.
This
operati
on
was
rep
eat
e
d
for
eve
ry
featu
re
rem
ov
ed.
T
he
featu
re
el
im
inate
d
first
was
co
ns
ide
re
d
as
the
worst
c
on
tri
bu
ti
ng
featur
e
an
d
the
fea
t
ur
e
el
i
m
inate
d
la
st
was
c
onside
r
ed
as
t
he
best
-
co
ntributi
ng
f
eat
ur
e.
The
fi
nal
ste
p
was
to
s
ort
the
featur
e
s
f
ro
m
best
to
worst,
t
hen
c
om
pu
te
the
accu
racy
f
r
om
the
best
fe
at
ur
es
com
bin
at
ion
of each
sub
j
ect
.
3.
RESU
LT
S
Figure
5
s
how
s
the
relat
ion
s
hip
bet
wee
n
f
eat
ur
es
acc
ord
ing
to
the
c
onditi
on
re
presen
te
d
by
FEE
.
Gazin
g
f
ram
e
(G
F
),
gazin
g
c
luster
(
GC),
non
-
gazin
g
cl
ust
er
(N
C
),
a
nd
the
rati
o
of
G
F/GC
(RG
)
s
how
a
decr
easi
ng
te
ndency
with
i
nc
reasin
g
FEE
.
In
co
ntrast,
bl
ink
fr
am
e
(BF)
,
blink
cl
us
t
er
(BC),
the
r
at
io
of
BF/
BC
(RB),
a
nd
the
rati
o
of
non
-
gazi
ng
f
ra
m
e
(N
F
)/NC
(
RN)
sho
w
a
n
i
ncr
easi
ng
te
nd
ency
with
inc
r
easi
ng
FEE. NF
does
no
t
show
a
ny tend
e
ncy a
nd is
inconsist
ent
w
it
h
FEE.
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
:
1
7
0
–
1
7
8
174
Table
3
s
umm
arizes
the
ty
pical
sta
ti
st
ical
resu
lt
s
of
al
l
the
featur
e
s
in
t
he
three
cat
ego
ries
:
al
ert, li
gh
tl
y drow
sy,
hea
vily
d
r
owsy
. I
t ca
n be see
n
that t
he
G
F, GC, a
nd
NC exhibit
a st
at
ist
ic
ally si
gn
ific
ant
diff
e
re
nce
(
p
<
0.001;
Krus
kal
–
Wall
is
te
st)
acco
r
ding
to
the
dif
fer
e
nce
s
in
each
cat
egory
f
ollow
e
d
by
the
decr
easi
ng
t
re
nd
as
well
.
BF,
the
rati
o
of
BF/
BC
(RB),
an
d
the
rati
o
of
NF
/
NC
(RN)
al
s
o
ex
hi
bit
a
sta
ti
sti
cally
si
gn
i
ficant
dif
fe
ren
ce
(
p
<
0.001;
Krus
kal
–
Wall
is
te
st)
a
ccordin
g
to
th
e
diff
er
ences
i
n
each
cat
egory
f
ollo
wed
by
the
i
nc
reasin
g
tre
nd
a
s
well
.
I
n
a
no
t
her
cas
e,
eve
n
though
BC
an
d
th
e
r
at
io
of
GF
/GC
(RG)
te
nded
t
o
FEE
durin
g
th
e
dro
wsin
ess
s
ta
te
,
these
pa
ra
m
et
ers
wer
e
not
co
ns
ide
re
d
to
ha
ve
a
sig
nif
ic
ant
diff
e
re
nce
in
a
ny
of
the
dr
owsiness
cat
e
gories.
Mo
reove
r,
t
he
NF
al
s
o
has
no
sig
nifica
nt
diff
e
re
nce
ow
i
ng
to
it
s
inco
ns
ist
ent
te
ndency
t
o
F
EE
durin
g
t
he
drowsi
ness
sta
te
.
Re
ga
rd
i
ng
the
sta
ti
sti
cal
r
esults
in
these
three
cl
asses,
we
ob
ta
ined
t
he
res
ults
f
or
bo
t
h
i
ncr
ease
d
an
d
decr
ease
d
pro
per
ti
es
t
o
repr
esent
the
dro
wsin
ess
conditi
on, es
pe
ci
al
ly
f
or
cate
gory
2, w
hich
desc
ribes
the sta
te
b
ef
or
e
b
e
co
m
ing
dr
ow
sy.
Figure
3
.
S
OD
of fram
es in which gazi
ng
oc
curred
Figure
4
.
Pro
duct
ion o
f gazi
ng
from
m
ov
ing
sp
ee
d for
featur
e
ex
tract
i
on
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
Ga
zi
ng as
actu
al par
amet
er for dr
owsiness
as
sess
me
nt in drivi
ng sim
ula
t
or
s
(
Arth
ur Mo
ur
it
s R
uma
git
)
175
Figure
5
.
Featu
res
te
nde
ncy a
ccordin
g
t
o
FE
E
Table
3
. St
at
ist
ic
al
Re
su
lt
s
of
al
l
Feat
ur
es
,
T
hr
ee
Cat
egorie
s: Alert,
Lig
htly
D
r
owsy
, Hea
vily
D
r
owsy
Para
m
eter
Alert
Ligh
tly
d
rows
y
Heavil
y
dro
ws
y
GF
2
1
1
.39
±
5
0
.8
6
1
6
5
.88
±
5
9
.9
1
*
*
*
1
1
2
.34
±
0
5
8
.34
*
*
*
,
#
#
#
GC
1
2
8
.33
±
2
1
.7
8
0
9
8
.63
±
3
0
.2
8
*
*
*
0
6
9
.08
±
0
3
0
.61
*
*
*
,
#
#
#
NF
6
5
8
.36
±
5
0
.1
0
6
6
7
.54
±
7
6
.4
5
*
*
*
6
2
7
.28
±
1
0
9
.76
*
*
*
,
###
NC
1
3
3
.67
±
2
0
.6
8
1
0
6
.06
±
2
8
.8
4
*
*
*
0
7
8
.45
±
0
2
8
.13
*
*
*
,
#
#
#
BF
0
2
2
.06
±
2
6
.7
4
0
3
6
.15
±
3
2
.8
6
*
*
*
1
2
8
.31
±
1
1
4
.43
*
*
*
,
#
#
#
BC
00
5
.16
±
0
3
.3
9
00
7
.54
±
0
4
.7
8
*
*
*
0
1
0
.07
±
00
5
.93
*
*
*
,
#
#
#
RG
00
1
.63
±
0
0
.1
8
00
1
.58
±
0
0
.1
6
*
*
*
00
1
.48
±
00
0
.19
*
*
*
,
#
#
#
RN
00
5
.10
±
0
1
.2
0
00
6
.86
±
0
2
.3
9
*
*
*
00
9
.01
±
00
3
.69
*
*
*
,
#
#
#
RB
00
4
.82
±
0
1
.4
1
00
9
.95
±
0
4
.6
6
*
*
*
0
2
1
.18
±
0
1
7
.02
*
*
*
,
#
#
#
*
*
*
p < 0.0
0
1
vs
.
alert,
##
#
p < 0.0
0
1
vs
.
lig
h
tly
dro
wsy;
all
v
alu
es ar
e exp
r
ess
ed
as
m
e
an
±
S
D
The
res
ults
of
the
t
hr
ee
cat
egories
s
how
that
the
gazin
g
par
am
et
ers
cou
l
d
be
us
e
d
to
est
im
at
e
drowsi
ness
.
H
ow
e
ve
r,
sev
er
al
su
bject
s
di
d
not
show
a
ny
sign
i
ficant
diff
e
ren
ce
s
corres
pondin
g
to
the
ps
yc
holo
gical
m
easur
em
ents u
sin
g
FE
E in
a
ll
cate
go
ries.
We ass
um
ed
that i
t was cau
se
d
by the
dif
fere
nces in
per
ce
ptio
n
duri
ng
the
e
xam
iners’
eval
uation
w
hile
exam
ining
the
sub
j
ect
s’
ph
ysi
cal
sta
te
duri
ng
dri
ving
an
d
wh
e
n watc
hi
ng the
vid
e
o reco
rd
i
ng of the
s
ubj
ect
s
drivi
ng
as w
el
l.
We
use
d
al
l
nin
e
pa
ram
et
er
featur
es
durin
g
the
cl
assifi
cat
ion
a
naly
sis.
Ta
ble
4
s
how
s
th
at
the
SVM
was
able
to
de
te
ct
the
dr
owsi
ness
with
a
n
overall
accu
rac
y
of
76.
3%
in
the
three
cat
eg
or
ie
s
of
sta
te
:
al
ert,
li
gh
tl
y dro
wsy,
and
heav
il
y
drow
sy.
Table
4.
Cl
assi
ficat
ion
Re
s
ults
of
al
l Feat
ure
s
f
or
t
he
T
hree
Ca
te
go
ries:
Al
ert, Lig
htly
Dr
ow
sy,
Heav
il
y D
r
ow
s
y
Su
b
ject
Accurac
y
[
%]
Bes
t co
m
b
in
atio
n
0
1
8
5
.4
NC, BF,
BC, RG
0
2
6
6
.2
GF,
GC
,
NF
,
BC
,
RG, RN
0
3
6
9
.2
GF,
GC
,
NF
,
NC
,
BC
0
4
6
8
.0
NC, BF,
RG,
R
N
0
5
6
3
.8
NF,
BF
,
BC
,
R
G,
RN
0
6
7
3
.8
GC, NF
,
NC
,
BF,
BC
,
RG, R
N
0
7
7
8
.3
GF,
GC
,
NF
,
BF,
BC
,
RG, R
N
0
8
8
5
.8
GC, NF
,
BF,
BC,
RG, RN
0
9
7
7
.6
GF,
GC
,
NF
,
NC
,
BC
,
RG, R
N,
RB
10
8
7
.6
RB
11
8
3
.4
GF,
GC
,
NF
,
BF,
BC
,
RG, RB
Ov
erall
7
6
.3
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
:
1
7
0
–
1
7
8
176
4.
DISCU
SSI
ON
We
in
vestigat
e
d
the
relat
io
nship
betwee
n
ga
zi
ng
prop
e
rtie
s
and
dro
wsi
ness
usi
ng
fea
tures
of
the
gazin
g
par
am
et
er
an
d
t
he
dro
wsin
ess
co
ndit
ion
us
i
ng
FEE
scor
e
s
du
rin
g
dr
i
ving.
Dro
ws
iness
is
c
onsid
ered
to
be
relat
ed
t
o
t
he
hum
an
co
nd
i
ti
on
.
The
sim
plest
way
t
o
det
erm
ine
the
sta
te
of
t
he
hu
m
an
body
is
by
di
r
ect
ly
askin
g
t
he
s
ubj
ect
t
heir
pr
esent
c
onditi
on
or
ha
ving
t
hem
and
t
he
exam
iner
co
m
ple
te
a
dro
wsin
ess
assessm
ent.
Sever
al
resea
rch
e
rs
hav
e
use
d
fa
ci
al
expressio
n
evaluati
on
(FEE
)
i
n
qu
e
sti
onnai
res
to
ob
ta
in
th
e
ph
ysi
cal
co
nd
i
ti
on
,
es
pecial
ly
the
drow
si
ne
ss
conditi
on,
of
a
subj
ect
.
M
or
e
over,
pr
e
vi
ou
s
st
ud
ie
s
ha
ve
al
so
evaluate
d
the
pe
rfor
m
ance
of
the
FE
E
(as
a
drowsi
ness
eva
luati
on
to
ol).
Con
s
eq
ue
ntly
,
it
has
bee
n
co
nc
lude
d
that
the
fluct
ua
ti
on
of
F
EE
values
represe
nts
the
c
onditi
on
of
the
s
ubje
ct
beco
m
ing
drowsy
.
I
n
thi
s
stud
y,
we
obser
ve
d
the
drowsi
nes
s
conditi
on
of
subj
ect
s
du
r
ing
dri
vi
ng
usi
ng
an
act
ua
l
dr
ivin
g
sim
ulator
.
We
obta
ine
d
t
he
ga
zi
ng
sig
na
l
by
us
in
g
a
he
ad
-
m
ounted
e
ye
-
trackin
g
de
vice,
the
n
e
xt
r
act
ed
an
d
a
naly
zed
the
featur
e
s
of
t
he
gazin
g
par
a
m
et
er
us
ing
t
he
FEE
pr
operti
es
as
our
evaluati
on
dro
wsin
ess
m
et
ho
d
t
o
evaluate
dro
ws
iness.
Seve
ral
stud
ie
s
hav
e
in
vestig
at
ed
dro
wsin
es
s
based
on
ey
e
pr
ope
rtie
s.
For
instance,
Jac
ks
on
et
al
.
[34]
in
vestiga
t
ed
sl
ow
ey
el
id
cl
os
ure
a
s
a
m
easur
e
of
dri
ver
drow
si
ne
ss
by
m
easur
ing
slo
w
ey
e
c
losure
(P
ERC
L
OS
)
wh
il
e
dr
i
ver
s
perform
ed
a
sim
ula
te
d
dri
ving
ta
s
k.
Howe
ve
r,
t
heir
st
ud
y
is
sti
ll
lim
it
e
d
in
it
s
discuss
i
on
of
t
he
par
am
et
ers
associat
ed
wi
th
the
ph
ysi
ca
l
condi
ti
on,
es
pecial
ly
the
drow
si
ness
c
ondi
ti
on
.
More
ov
e
r,
the
y
placed
the
c
a
m
era
in
fron
t
of
the
dr
i
ver,
wh
ic
h
does
no
t
pr
ovide
a
cl
e
ar
view
for
the
dr
ive
r
and is
un
sta
ble in term
s o
f
po
sit
ion
.
Ma
’touq
et
al
.
[
35
]
us
e
d
ey
e
blin
king
t
o
detect
dri
ve
r
drowsi
ness
.
T
hey
propose
d
a
de
vice
f
or
m
on
it
or
ing
a
dri
ver’s
drowsi
ne
ss
by
detect
in
g
an
d
cl
assify
i
ng
t
he
ey
e
blin
king
into
norm
al
blink
in
g
(NB
)
or
prolo
nged
blin
king (
PB)
. H
oweve
r,
they
d
id n
ot d
isc
us
s
t
he
relat
ion
s
hip
of
the
pa
ram
eter
s
ass
ociat
ed
w
it
h
th
e
ph
ysi
cal
c
o
ndit
ion
,
es
pecial
ly
the dr
ow
si
ness
cond
it
io
n.
Wang
a
nd
Xu
[36]
inv
est
igat
ed
dro
wsi
ness
base
d
on
ey
e
pro
per
ti
es.
T
he
y
detect
ed
the
drowsi
nes
s
by
us
i
ng
ey
e
f
eat
ur
es:
per
ce
ntage
of
ey
e
c
losure
(
PERC
LOS),
a
ver
a
ge
pupil
diam
et
e
r,
a
nd
blin
k
durati
on
com
bin
ed
wit
h
dri
ving
be
ha
vi
or
pa
ram
et
ers.
They
f
ur
t
her
us
e
d
m
ulti
le
vel
orde
red
lo
git
(MO
L),
order
lo
git
(O
L
),
an
d
a
rtif
ic
ia
l
network
(
ANN)
to
deter
m
ine
dro
wsin
e
ss
in
t
hr
ee
dr
owsiness
cat
eg
ori
es.
T
he
res
ults
of
their
stu
dy
sho
wed
t
hat
the
overall
accu
racy
us
in
g
MOL wa
s
64.15
%, OL
was
52.
70
%
,
an
d
A
NN
w
a
s
56.04
%
(MO
L
had
the
highest
detect
ion
acc
ur
acy
).
T
heir
stu
dy
al
so
c
onfirm
ed
that
ey
e
feat
ures
perform
ed
bette
r
than
dr
i
ving
be
hav
i
or
i
n
dro
wsin
ess
detect
ion
.
It
was
c
onfirm
ed
by
re
m
ov
ing
the
e
ye
featur
e
s
th
at
the
acc
ur
acy
was
reduce
d.
H
ow
ever,
their
stu
dy
has
a
lowe
r
accuracy
in
the
detect
ion
of
dro
wsin
e
s
s
than
our
st
ud
y.
Drowsi
ness
ha
s
al
so
bee
n
inv
est
i
gated
base
d
on
ey
e
pro
per
ti
es
usi
ng
m
achine
le
arn
i
ng
or
cl
assifi
cat
ion
m
et
ho
ds.
H
u
and
Z
he
ng
[
37]
us
ed
a
n
S
VM
to
cl
assify
the
dr
owsin
ess
conditi
on
in
thre
e
cat
egories
with
an
ov
e
rall
accuracy
of
80.
74%
in
a
dr
ivin
g
si
m
ulator
env
i
r
on
m
ent.
They
detect
ed
dro
ws
iness
via
ey
el
id
relat
ed
pa
ram
et
ers
us
in
g
E
OG.
Al
though
they
obta
ined
a
hi
gh
e
r
accuracy
tha
n
that
ob
ta
ine
d
i
n
our
stud
y,
thei
r
stu
dy
has
a
dra
wback
in
that
el
ec
tro
des
we
re
at
ta
ched
t
o
the
dr
iver,
w
hic
h
co
uld
ca
us
e
disc
om
fo
rt
durin
g dr
i
ving.
In
this
st
ud
y,
we
us
e
d
a
hea
d
-
m
ou
nted
ey
e
tracke
r
to
ove
rc
om
e
view
and
po
sit
io
n
li
m
it
ation
s
,
an
d
to
el
i
m
inate
intrusi
on
w
hile
extracti
ng
ey
e
prop
e
rtie
s.
This
kind
of
in
vesti
gation
has
rare
ly
been
cond
ucted.
To
assess
the
drow
si
ness
c
onditi
on,
we
cond
ucted
a
s
ubj
ect
ive
e
valuati
on
of
eac
h
sub
j
ect
’s
physi
cal
conditi
on
us
in
g
F
EE.
W
e
f
oc
us
e
d
on
th
ree
c
at
egories
f
or
es
tim
a
ti
ng
the
co
nd
it
io
n
befo
re
act
ual
dro
wsi
ne
ss
to
pr
e
ve
nt
acci
den
ts.
O
nly
a
few
stud
ie
s
ha
ve
been
co
nduc
te
d
on
ga
zi
ng
pro
per
ti
es
r
el
at
ed
to
dr
ows
iness
.
A
novel
pa
ra
m
et
er
was
pr
e
sented
in
this
stud
y.
W
e
fou
nd
that
the
fe
at
ur
es
of
the
gazin
g
ha
d
sig
nificant
sta
ti
sti
cal
diff
eren
ces
in
t
hr
ee
dro
wsin
es
s
ca
te
gories:
al
ert
(F
EE=
0),
li
ghtl
y
dr
owsy
(F
E
E=1−
2)
,
an
d
he
avily
drowsy
(F
EE
=3−4).
Se
ve
r
al
featur
es
of
the
gazi
ng
—
gazin
g
occ
urren
ce
fr
am
es
per
e
po
c
h
(G
F
),
gazin
g
occ
urre
nce
cl
us
te
rs
pe
r
epo
c
h
(
GC),
blink
occurre
nc
e
fr
am
es
per
epo
c
h
(BF
),
non
-
gazin
g
occ
urren
c
e
cl
us
te
rs
per
e
poch
(N
C
),
rati
o
of
blin
king
f
ram
es
ver
su
s
c
lusters
(RB
=
BF/
BC
),
an
d
the
rati
o
o
f
non
-
gazi
ng
fr
am
es
ver
sus
cl
us
te
rs
(R
N
=
NF
/NC)
—
ha
d
sign
ific
a
nt
sta
ti
sti
cal
diff
ere
nc
es
with
p
<
0.
001;
K
ru
s
kal
-
Wall
is
te
st. Overall
, se
ven feat
ur
es
wer
e
suff
ic
ie
nt
to dete
ct
the
dro
wsin
es
s c
ondi
ti
on
in
t
hr
ee
c
at
egories.
Ba
sed
on
th
os
e
resu
lt
s,
a
n
SVM
was
us
e
d
to
exam
ine
the
per
f
or
m
ance
of
the
featu
res
to
cl
assify
the
drowsi
ness
co
nd
it
io
n.
I
n
the
three
cat
e
gorie
s,
al
ert
(F
EE=
0),
li
ghtl
y
dro
wsy
(
FEE=
1−
2),
a
nd
hea
vily
dro
wsy
(F
EE=
3−
4)
,
th
e
SV
M
was
a
ble
to
detect
the
drowsi
ness
with
a
n
over
al
l
accuracy
of
76.
3%
.
I
n
a
dd
it
io
n,
in
two
cat
e
gories,
al
ert
(F
EE=
0)
a
nd
dro
wsy
(F
EE=
1−
4)
,
t
he
SV
M
was
a
bl
e
to
detect
the
drowsi
ness
with
a
n
ov
e
rall
accu
rac
y
of
89.0%.
N
ote
that
the
res
ult
of
t
he
cl
ass
ific
at
ion
is
s
ubj
ect
-
dep
e
ndent
as
it
was
cal
c
ulate
d
us
in
g
eac
h
s
ubj
ect
’s
data.
T
he
cl
assifi
cat
ion
res
ults
show
that
the
featu
re
s
of
gazi
ng
ca
n
be
use
d
t
o
de
te
ct
three
dro
wsin
e
ss
cat
eg
or
ie
s
from
the
best
fe
at
ur
es
c
om
bin
at
ion
of
eac
h
subj
ect
;
the
se
re
s
ults
we
re
c
onfirm
ed
via the
FEE
qu
est
ionnaire.
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
Ga
zi
ng as
actu
al par
amet
er for dr
owsiness
as
sess
me
nt in drivi
ng sim
ula
t
or
s
(
Arth
ur Mo
ur
it
s R
uma
git
)
177
The
feat
ur
es
of
the
gazin
g
pa
r
a
m
et
er
cou
ld
be
us
ed
as
a
ne
w
pa
ram
et
er
or
var
ia
ble
in
drow
si
ness
t
o
ob
ta
in
t
he
c
harac
te
risti
cs
and
sta
te
of
the
ey
es.
W
e
belie
ve
that
these
c
om
bin
at
ion
s
effe
ct
ively
rep
res
ent
th
e
aspects
of
the
ey
e
pro
per
ti
es
and
co
uld
be
use
d
t
o
determ
i
ne
the
dr
ow
sy
condit
ion
ef
fec
ti
vely
.
The
ey
e
s
ar
e
com
m
on
ly
kn
own
to
be
a
pa
r
t
of
the
hum
an
body
that
can
cl
early
rep
res
ent
the
hum
an
conditi
on
of
bein
g
asl
eep
or
awa
ke.
Usi
ng
the
gazin
g
pro
per
t
ie
s,
we
can
ge
ner
al
ly
say
t
hat
a
hum
an’
s
ey
es
easi
ly
be
com
e
unf
ocu
se
d
w
hile
sta
rtin
g
to
f
al
l
asl
eep
or
be
com
ing
dro
w
sy
w
hen
he/s
he
sta
rts
getti
ng
sle
e
py
or
be
com
e
drowsy
.
I
n
c
on
trast
,
a
hu
m
an’s
gaze
is
f
ocu
s
ed
wh
e
n
c
once
ntrati
ng
on
a
s
pecific
obj
ect
.
By
con
si
der
i
ng
tho
se
sp
eci
fic
phe
no
m
ena
and
us
i
ng
the
gazin
g
pro
per
ti
es,
we
obta
ined
us
e
f
ul
par
am
et
ers
an
d
s
howe
d
that
us
in
g
the
rati
o
of
diff
e
ren
t
pa
ram
e
te
rs
prov
i
de
d
sign
ific
a
nt
dif
fer
e
nces
an
d
cou
l
d
induce
cl
assifi
cat
ion
r
esults
betwee
n
s
ubj
e
ct
s
as
well
as
the
gazi
ng
fe
at
ur
es
t
hem
selv
es.
Gazin
g
pa
ram
et
ers
are
com
po
sed
of
s
ever
al
featur
e
s t
hat c
ould im
pr
ove t
he
estim
at
ion
of
drowsines
s
by co
m
bin
ing s
pe
ci
fic p
ar
am
et
er
s.
Howe
ver,
our
current
stu
dy
has
the
f
ollowi
ng
lim
it
ation
.
In
or
der
to
in
duce
dro
wsin
e
s
s,
we
aske
d
each
sub
j
ect
to
dri
ve
in
un
r
eal
ist
ic
con
diti
on
s
,
su
c
h
as
on
an
oval
trac
k
with
no
ob
s
ta
cl
es
and
no
sp
ee
d
changes
.
I
n
re
al
it
y,
peo
ple
dri
ve
on
var
io
us
ro
a
ds
w
hile
con
t
ro
ll
in
g
the
ir
veh
ic
le
’s
s
pe
ed
an
d
disce
r
ning
signp
os
ts
a
nd
oth
e
r
veh
ic
le
s.
I
n
s
uc
h
a
sc
enar
i
o
with
obsta
cl
es,
t
he
s
ubj
ect
w
ou
l
d
hav
e
to
lo
ok
at
the
ob
sta
cl
es
in
orde
r
t
o
dr
i
ve
safely
.
C
ons
equ
e
ntly
,
we
hypothesiz
e
t
ha
t
non
-
gazi
ng
w
ou
l
d
occ
ur
m
or
e
fr
e
qu
e
ntly
.
W
e
will
exa
m
ine
wh
et
her
our
pr
opos
e
d
feat
ur
e
s
sh
ow
the
sa
m
e
resu
lt
s
regardless
of
the
s
cenari
o
in futu
re
wor
k.
5.
CONCL
US
I
O
N
In
this
stu
dy,
a
novel
par
am
eter
and
it
s
f
eat
ur
es
wer
e
pro
pose
d
to
detect
drowsi
ness
a
nd
sta
ti
sti
cal
and
cl
assi
ficat
ion
te
c
hn
i
qu
e
s
wer
e
us
e
d
to
quantify
the
perform
ance
of
ga
zi
ng
pro
per
ti
es
represe
nting
s
ever
al
drowsi
ness
co
nd
it
io
n
le
vels.
Our
res
ults
in
dicat
e
that
the
pro
posed
gazing
pa
ram
et
er
ca
n
e
ff
ect
ively
asses
s
the dr
ow
si
ness
level o
f
a
driv
er.
ACKN
OWLE
DGE
MENTS
This
wor
k was
par
ti
al
ly
suppo
rted by JS
PS K
AK
E
N
HI grant
num
ber
JP
16K00
375.
REFERE
NCE
S
[1]
Dingus
TA,
Jahns
SK
,
Horow
it
z
AD
,
Knipli
ng
R.
Hum
an
fac
to
rs
design
issues
for
cra
sh
avoi
d
a
nce
s
y
st
ems
.
In:
Barf
ield
W
,
Dingus
TA.
Edi
tors.
Hum
an
Fact
ors
in
Inte
ll
ige
n
t
Tra
nsporta
ti
on
S
y
stems
.
New
York:
Ps
y
cho
lo
g
y
Press
.
1998:
55
–
93.
[2]
El
zoh
ai
r
y
Y
.
Fata
l
and
inj
ur
y
fa
ti
gue
-
r
el
a
te
d
cr
a
shes
on
Ontar
io
’
s
roa
ds:
A
5
-
y
e
ar
rev
ie
w.
Th
e
Highwa
y
Safe
t
y
Roundta
ble &
Fati
gu
e
Im
pai
rm
e
nt,
Driv
er
Fa
ti
gu
e
S
y
m
posium
.
T
oronto.
2007
.
[3]
Nati
ona
l
Highwa
y
Tra
f
fic
Saf
ety
Adm
ini
stra
ti
o
n.
As
le
ep
At
Th
e
W
hee
l:
A
Nat
iona
l
Com
pendium
of
Eff
orts
t
o
El
iminate
Drow
sy
Dr
ivi
ng.
U.S.
D
epa
rtment
o
f
T
ran
sportation.
20
17.
[4]
Tra
ffi
c
Bur
ea
u
.
The
num
ber
of tr
aff
ic acci
d
ent occurre
nc
es
in
201
7.
Nat
ional
Poli
c
e
Agenc
y
of
Jap
an.
2017
.
[5]
Li
u
CC,
Hos
kin
g
SG
,
Le
nné
M
G.
Predic
t
ing
dri
ver
drows
ine
ss
using
vehi
c
le
m
eas
ure
s:
Rec
en
t
in
sights
and
futur
e
cha
l
le
nges.
Jour
nal
of
Safe
t
y
R
e
sea
rch
.
2009;
40
(4):
239
–
245.
[6]
Friedri
chs
F,
Ya
ng
B.
Drow
sines
s
m
onit
oring
b
y
stee
ring
and
la
n
e
da
ta
b
ase
f
ea
tu
res
under
r
eal
dr
ivi
ng
cond
it
ions
.
Proce
edi
ngs of
t
he
18th
Europ
ean Signa
l
Proce
ss
ing
Confer
ence
(
EUSIP
CO).
Aalbor
g.
2
010:
209
–
213.
[7]
Ingre
M,
Åkerste
dt
T,
Pet
ers
B,
Anund
A,
Kec
klund
G.
Subjec
tive
slee
pin
ess,
sim
ula
te
d
drivi
ng
per
form
anc
e
and
bli
nk
dur
at
ion
: E
xamining
ind
ividual
d
iffe
r
enc
es
.
Journal
of
Sle
ep
Resea
r
ch. 2006; 15(1):
47
–
53
.
[8]
Sum
an
D,
Mali
ni
M,
Anchuri
S.
EOG
base
d
vigi
l
anc
e
m
onit
or
ing
sy
st
em.
Proce
e
dings
of
the
Annual
IEEE
Indi
a
Confer
ence
(IN
DICO
N).
New Del
hi. 2015: 6
p
age
s.
[9]
Góm
ez
-
Cla
per
s
J,
Casanella
R.
A
fast
and
ea
s
y
-
to
-
use
ECG
a
cq
uisit
ion
and
he
a
rt
rate
m
onit
or
in
g
s
y
stem
using
a
wire
le
ss
ste
eri
ng
whee
l
.
I
EEE
Se
nsors
Journal.
20
12;
12(3):
610
–
6
16.
[10]
Iga
saki
T
,
Naga
sawa
K,
Mura
yama
N,
Hu
Z.
Drow
siness
esti
m
at
ion
under
dr
ivi
ng
env
ironment
b
y
hea
r
t
ra
t
e
var
ia
b
il
i
t
y
and/
o
r
bre
at
h
ing
rate
var
ia
b
il
i
t
y
wi
th
logi
stic
r
egr
essi
on
ana
l
y
sis.
Pro
ce
ed
ings
of
the
8t
h
Inte
rn
at
ion
al
Confer
ence
on
B
iomedic
a
l
Eng
in
ee
ring
and
Infor
m
at
ic
s (BMEI)
.
Shen
y
ang
.
2015
:
189
–
193.
[11]
Vice
nt
e
J,
La
gu
na
P,
Bart
ra
A,
Bai
lón
R.
Drow
siness
det
ec
ti
on
using
hea
rt
rate
var
ia
b
il
i
t
y
.
Medi
ca
l
&
Biol
og
ic
a
l
Engi
ne
eri
ng
& Com
puti
ng.
201
6;
54(6):
927
–
93
7.
[12]
Akerste
dt
T
,
Gill
ber
g
M.
Subje
ct
iv
e
and
object
ive
slee
p
ine
ss
in
the
active
ind
ivi
dual.
Int
ern
ational
Journal
o
f
Neurosci
ence. 1
990;
52(1
–
2):
29
–
37.
[13]
Kaida
K,
T
aka
h
ashi
M,
Akerste
dt
T,
Nak
ata
A,
Otsuka
Y,
Hara
t
ani
T
,
Fukasawa
K.
Vali
da
ti
on
o
f
the
Karol
inska
slee
pin
ess sca
l
e aga
inst
per
form
a
nce
and EE
G v
ar
ia
bl
es.
Cl
inical
Neuroph
y
siolog
y
.
2006;
117(7):
1574
–
1581.
[14]
Akbar
IA,
Iga
sa
ki
T,
Mura
y
ama
N,
Hu
Z.
Drow
s
ine
ss
assessment
using
el
ec
troe
n
ce
pha
logra
m
in
drivi
ng
sim
ula
to
r
envi
ronm
ent
.
Proce
ed
ings
of
th
e
8t
h
In
te
rna
ti
on
a
l
Confer
ence
on
Biom
edi
cal
Eng
i
nee
ring
and
Info
rm
at
ic
s
(BMEI)
.
Shen
y
ang
.
2015
:
184
-
188.
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
:
1
7
0
–
1
7
8
178
[15]
Akin
M,
Kurt
MB,
Sezgi
n
N,
Ba
y
r
am
M.
Est
i
m
at
ing
vigilan
c
e
l
eve
l
b
y
usin
g
EE
G
and
EM
G
signal
s.
Neur
al
Com
puti
ng
and Applic
a
ti
ons.
20
08;
17(3):
227
–
3
6.
[16]
Saha
y
adha
s
A,
Sundara
j
K,
Murugappa
n
M.
Dr
ows
ine
ss
det
ecti
on
during
di
ffe
r
ent
t
imes
of
da
y
using
m
ult
iple
fea
tur
es.
Aus
tral
asia
n
Ph
y
sic
al &
Engi
n
ee
ring
Sci
enc
es
in
Med
ic
i
ne.
2013
;
36(2)
: 43
–
50.
[17]
Ha
y
ashi
K,
Ishihar
a
K,
Hashim
oto
H,
Oguri
K.
Indivi
dualiz
ed
drows
i
ness
det
ec
t
ion
during
driving
b
y
pulse
wave
ana
l
y
sis wi
th
n
e
ura
l
n
et
work.
Proce
ed
ings o
f the
IEE
E
Intelli
g
ent
Tra
nsporta
ti
on
S
y
stems
.
Vi
enna. 2005:
901
–
906.
[18]
Yang
F,
Yu
X,
Huang
J,
Yang
P,
Meta
x
as
D.
R
obust
e
y
e
li
d
trac
king
for
f
at
igu
e
det
e
ct
ion
.
Proc
e
edi
ngs
of
the
19
t
h
IEE
E
Internat
ion
al
Conf
ere
n
ce o
n
Im
age
Proc
essing
(ICIP).
Orla
nd.
2012:
1829
–
1832.
[19]
Kawat
o
S,
Te
ts
uta
ni
N
.
Det
ec
t
i
on
and
tr
ac
k
ing
of
e
y
es
for
ga
z
e
-
ca
m
er
a
con
tro
l.
Im
age
and
Vi
sion
Com
puti
ng.
2004;
22(12):
10
31
–
1038.
[20]
Sm
it
h
P,
Shah
M,
d
a
Vitor
ia
Lobo
N
.
Mon
it
oring
hea
d
/ey
e
m
oti
on
for
dri
ver
a
le
r
tne
ss
with
on
e
c
amera
.
Proce
edi
ngs of
t
he
15th
Int
ern
ati
onal
Conf
ere
n
ce on
Pat
te
rn
Rec
o
gnit
ion (ICP
R).
Barc
e
lona.
2000
:
636
–
642.
[21]
Ghos
h
S,
Nand
y
T,
M
anna
N.
R
ea
l
ti
m
e
e
y
e
d
etec
t
ion
and
tracki
ng
m
et
hod
for
d
r
ive
r
assistanc
e
s
y
stem.
In
:
Gupt
a
S,
Sandip
B
,
Ka
rab
i
G,
Indra
n
ath
S,
Papun
B
.
E
dit
ors.
Advan
cem
ent
s
of
Medical
E
lectr
oni
cs:
P
roc
ee
d
ings
of
th
e
1st I
CAM
E
Inter
nat
ion
al
Conf
erence
.
New De
lhi
:
Springer
Indi
a; 2
015:
13
–
25.
[22]
Hac
hisuka
S,
Is
hida
K,
En
y
a
T,
Kam
ij
o
M.
Faci
a
l
expr
essio
n
m
ea
surem
ent
for
det
e
cting
drive
r
drows
ines
s.
Engi
ne
eri
ng
Ps
ychol
og
y
and
Co
gnit
ive E
rgonom
ic
s.
2011
;
135
–
4
4.
[23]
Sigari
MH
,
Fath
y
M,
Sor
y
ani
M.
A
drive
r
fac
e
m
onit
or
ing
sy
st
em
for
fatigue
and
distr
a
ct
ion
de
tecti
on
.
Inte
rna
ti
ona
l
Jou
rna
l
o
f
Vehi
cul
a
r
Technol
og
y
.
2
013;
(263983):
1
1
page
s.
[24]
Danism
an
T,
Bi
la
sco
IM,
Dj
eraba
C,
Ih
adda
d
e
ne
N.
Drow
s
y
drive
r
de
tecti
on
s
y
stem
using
e
y
e
bli
nk
p
atter
n
s.
Proce
edi
ngs of
t
he
Int
ern
a
ti
ona
l Confere
nc
e
on
Mac
hine a
nd
W
eb
Int
el
l
ige
n
ce
(
ICMW
I).
Algie
r
s.
2010:
230
–
23
3.
[25]
Fazl
i
S,
Esfe
ha
ni
P.
Tracki
ng
e
y
e
sta
te
for
f
a
ti
gue
d
et
e
ct
ion
.
Proce
edi
ngs
of
the
Int
ern
ation
al
Confer
ence
o
n
Advanc
es
in
Co
m
pute
r
and
Elec
tri
c
al
Engi
n
ee
rin
g
(ICACEE)
.
M
ani
l
a. 2012:
17
–
20.
[26]
Jo
J,
Le
e
SJ
,
Park
KR,
Kim
I
J,
Kim
J.
Dete
ct
in
g
drive
r
dro
ws
ine
ss
using
fea
ture
-
le
v
el
fusion
a
nd
user
-
spec
ific
cl
assifi
ca
t
ion. Expert
S
y
stems
with
Appli
cations. 2014;
41(4)
:
11
39
–
1152.
[27]
Garc
ia
I
,
Bronte
S,
Berga
sa
LM,
Alm
az
án
J,
Yebe
s
J.
Vision
-
base
d
drows
ine
ss
det
ec
tor
for
re
al
dri
ving
condi
ti
ons
.
Proce
edi
ngs of
t
he
IE
EE Int
e
ll
ig
ent
Veh
ic
l
es
S
ym
posium
(IV).
Alca
l
a
d
e
Hen
ar
es.
2012:
618
–
6
23.
[28]
Alvar
o
PK
,
Jac
kson
ML,
Berl
o
witz
DJ
,
Sw
ann
P,
How
ard
ME
.
Prolonged
e
y
e
l
id
cl
osure
epi
so
des
during
sle
ep
depr
ivation in
pr
ofe
ss
iona
l
dr
ive
r
s.
Journal
of
Cl
i
nic
a
l
Sle
ep
Med
i
ci
ne
.
2016
;
12(8
):
1099
–
1103.
[29]
Tsuchi
da
A,
B
hui
y
an
MS
,
Oguri
K.
Esti
m
ati
on
of
drows
ine
ss
le
vel
base
d
on
e
y
el
id
cl
osu
re
and
he
art
ra
te
var
ia
b
il
i
t
y
.
Proc
ee
dings
of
the
31st
Annual
Inte
rna
t
iona
l
Conf
ere
nc
e
of
the
I
EE
E
Eng
ine
er
in
g
in
Medic
ine
a
nd
Biol
og
y
Soci
ety
(EM
BC).
Minne
apol
is.
2009:
25
43
–
2546.
[30]
Chang
CC,
L
in
CJ.
LIBSVM
:
A
li
bra
r
y
for
supp
ort
vector
m
ac
hi
nes.
ACM
Tra
nsac
t
ions
on
Inte
l
ligent
S
y
st
ems
and
Te
chno
log
y
.
201
1;
2(3):
27:1
–
27:
27.
[31]
Akbar
IA,
Ru
m
agi
t
AM
,
Uts
unom
i
y
a
M
,
M
orie
T,
Ig
asa
ki
T.
Three
dro
ws
ine
ss
ca
t
egor
ie
s
assess
m
ent
b
y
el
e
ct
roe
n
ce
ph
al
o
gra
m
in
drivi
ng
sim
ula
tor
enviro
nm
ent
.
Proce
edings
of
the
39
th
Annual
Int
ern
ati
onal
Conf
ere
nc
e
of
the IEEE Eng
ine
er
ing
in
Medi
ci
ne
and
Biol
og
y
So
ci
e
t
y
(
EMBC).
Jeju
.
2017
:
2
904
–
2907.
[32]
Gu
y
on
I,
W
eston
J,
Barnhi
ll
S,
Vapni
k
V.
Gene
sele
ction
for
cance
r
class
ifi
c
at
io
n
using
support
vec
tor
m
ac
hin
es
.
Mac
hine L
ea
rni
ng.
2002;
46
(1)
:
389
–
422.
[33]
Ebra
himi
F,
Se
t
are
hdan
SK
,
Na
ze
ran
H.
Autom
at
i
c
slee
p
stagi
n
g
b
y
sim
ultaneo
us
ana
l
y
sis
of
E
CG
and
respir
a
t
or
y
signal
s in
long
e
pochs.
Biom
edic
al
Sign
al
Proc
es
sing a
nd
Con
trol.
2015;
18:
69
–
7
9.
[34]
Jac
kson
ML,
Raj
S,
Croft
RJ,
Ha
y
l
e
y
AC
,
Dow
ne
y
LA,
Kenn
ed
y
GA
,
How
ard
ME.
Slow
ey
el
id
cl
osur
e
as
a
m
ea
sure
of
dr
iver drowsiness a
nd
it
s r
elati
onship
t
o
per
form
ance.
Tra
ffi
c
In
jur
y
Pr
eve
nt
ion. 2016;
17(3):
251
–
257.
[35]
Ma’touq
J,
Al
-
N
abul
si
J,
Al
-
Ka
z
wini
A,
Ban
i
y
as
sien
A,
Al
-
Haj
I
ss
a
G,
Moham
ma
d
H.
E
y
e
bl
inking
-
base
d
m
et
hod
for
detec
ti
ng
dr
i
ver
drows
ine
ss
.
Journal
of
M
edica
l
Engi
n
ee
r
ing &
T
ec
hnolog
y
.
2
014;
38(8):
416
–
419.
[36]
W
ang
X,
Xu
C
.
Drive
r
drows
ine
ss
d
et
e
ct
ion
base
d
on
non
-
i
ntrusive
m
et
r
ic
s
conside
ring
in
divi
dual
spec
ifics.
Acc
ide
n
t
Ana
l
y
s
is &
Prev
ent
ion
.
2016;
95:
350
–
3
57.
[37]
Hu
S,
Zh
eng
G.
Drive
r
d
rows
ine
ss
det
e
ct
ion
wi
th
e
y
e
li
d
re
la
t
ed
par
amet
ers
b
y
support
vector
m
ac
hine
.
Exp
ert
S
y
stems
with
A
ppli
c
at
io
n
.
2009
;
36(4):
7651
–
76
58.
Evaluation Warning : The document was created with Spire.PDF for Python.