Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
9
, No
.
6
,
Decem
ber
201
9
,
pp.
5277~
5285
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v9
i
6
.
pp5277
-
52
85
5277
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Hand de
tection
an
d seg
m
entation
usin
g s
m
art path tr
ac
kin
g
fingers a
s featu
res and
expert syst
em clas
sifier
Kha
le
d
N.
Yasen
1
, F
ahad
L
ayth
M
ala
ll
ah
2
, Lway F
aisa
l
A
b
dulraz
ak
3
,
Aso
M
ohamm
ad
D
arwesh
4
, Asem
Khm
ag
5
, Ba
r
aa T.
Sh
ar
eef
6
1
Depa
rtment of
Com
pute
r
Scie
n
ce
,
Cih
an
Univ
er
sit
y
-
Erbi
l
,
Erbil
4400
,
Ir
aq
2
Com
pute
r
and
I
nform
at
ion
Eng
i
nee
ring
,
Co
ll
e
g
e
of
E
lectr
oni
c
s
E
ngine
er
ing,
Nine
vah
Univer
si
t
y
,
Mos
ul,
Ira
q
3
Univer
sit
y
R
ese
arc
h
C
enter, Co
m
pute
r
Scie
n
ce
Depa
r
tment, Ci
h
an
Univer
si
t
y
Sl
emani
,
Slemani,
Ira
q
4
Depa
rtment of I
nform
at
ion
T
ec
h
nolog
y
,
Univ
ersi
t
y
of
Hum
an
De
vel
opm
ent
,
Sula
i
m
ani
,
Ir
aq
5
Com
pute
r
s
y
st
e
m
Engi
nee
r
ing,
Facul
t
y
of Engin
ee
ring
,
Univ
ersity
of
Z
awia,
Az
z
awia
,
L
ib
y
a
6
Depa
rtment
of
I
nform
at
ion
T
ec
h
nolog
y
.
Coll
ege
of
Inform
at
ion
T
ec
hnolog
y
,
Ahl
i
a
Univer
si
t
y
,
Mana
m
a,
Kingd
om
of
Bahrain
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Feb1
7
, 2
01
9
Re
vised
A
pr
2
,
201
9
Accepte
d
J
ul
21
, 2
01
9
Now
aday
s,
hand
gesture
re
cogni
t
ion
(HG
R)
is
g
etting
popul
ar
du
e
to
sev
eral
appl
i
ca
t
ions
such
as
remote
ba
sed
cont
rol
usin
g
a
hand
,
and
sec
urity
f
o
r
ac
c
ess
cont
rol.
One
of
the
m
aj
or
proble
m
s
of
H
GR
is
the
ac
cur
acy
lacking
hand
det
e
ct
ion
and
segm
ent
at
io
n.
In
thi
s
pape
r
,
a
new
al
go
ri
th
m
of
hand
det
e
ct
ion
wil
l
be
pre
sente
d
,
which
works
by
tracki
ng
finge
rs
sm
art
l
y
b
ase
d
on
the
p
la
nn
ed
pat
h.
The
tracki
ng
oper
a
t
ion
is
ac
complished
b
y
assum
ing
a
point
at
the
to
p
m
iddl
e
of
the
image
cont
a
ini
n
g
the
obje
ct
th
e
n
thi
s
point
slide
s
few
pix
els
down
to
be
a
ref
er
enc
e
point
the
n
br
anc
hing
int
o
tw
o
slopes:
le
ft
and
r
ight
.
On
the
se
sl
opes,
finge
rs
wil
l
be
sca
nned
to
e
xtra
c
t
fli
p
-
num
ber
s,
which
are
conside
r
ed
as
feature
s
to
b
e
class
ifi
ed
acc
ordingly
b
y
uti
lizing
the
exp
ert
s
y
s
te
m
.
Exp
eri
m
ent
s
were
c
o
nduct
ed
using
100
images
for
10
-
indi
vidual
cont
a
ini
ng
ha
nd
inside
a
cl
u
t
te
red
ba
ckgr
oun
d
b
y
using
Data
set
of
L
e
ap
Motion
an
d
Micr
osoft
Kinec
t
hand
ac
quisit
ions.
The
re
cor
ded
accura
c
y
is
depe
n
ded
on
th
e
complexi
t
y
of
th
e
Flip
-
Num
be
r
sett
ing
,
which
is
ac
hie
v
ed
96
%,
84%
and
81%
in
ca
se
6,
7
and
8
Flip_Num
ber
s
respe
ctively
,
in
which
thi
s
resul
t
ref
lects
a
h
igh
l
eve
l
of
finite
ac
cur
acy
in
compari
ng
wi
th exi
st
ing
t
ec
hniqu
es
.
Ke
yw
or
d
s
:
Ex
per
t sy
ste
m
Hand det
ect
io
n
Hand gest
ur
e
r
ecognit
ion
(HGR)
Hu
m
an
-
c
om
pu
te
r
interact
io
n
(H
CI
)
Segm
entat
ion
Copyright
©
201
9
Instit
ute of
Ad
v
ance
d
Engi
ne
eri
ng
and
Sc
ie
n
ce
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Faha
d
Lay
th
Ma
la
ll
ah
,
Com
pu
te
r
an
d Inform
at
ion
En
gin
ee
rin
g
,
Coll
ege
of
Ele
ct
ronics E
ng
i
ne
erin
g
,
Nine
vah Unive
rsity
,
ORCI
D:
0000
-
0001
-
60
67
-
7302
,
M
os
ul
, I
ra
q
.
Em
a
il
:
Fahad
.
m
al
al
la
h@
uoni
nev
a
h.
e
du.iq
1.
INTROD
U
CTION
In
or
der
to
en
han
ce
the
qu
al
it
y
of
li
fe
of
disabled
pe
ople
,
hu
m
an
-
com
pute
r
interact
ion
(H
CI
)
m
us
t
be devel
op
e
d
t
o
ac
hieve
t
he
a
forem
entioned
[1
]
. Ha
nd
gest
ur
e
rec
ogniti
on (
H
GR)
is a m
a
j
or to
pic
of H
C
I
tha
t
at
tract
s
resear
cher
s
in
diff
e
ren
t
fiel
ds
of
com
pu
t
er
vision
,
patte
rn
r
ecognit
ion,
an
d
m
achine
lear
ni
ng.
Hand
a
nd
hea
d
gest
ur
es
were
the
first
m
od
es
of
com
m
un
ic
at
ion
.
Us
ua
ll
y,
the
m
od
e
of
c
omm
un
ic
at
ion
is
verbal
an
d
non
-
ve
rb
al
.
I
n
te
r
m
s
of
non
-
ve
r
bal
com
m
un
ic
at
ion
,
it
ca
n
be
us
e
d
f
or
m
any
kinds
of
a
ppli
ca
ti
ons
su
c
h
as
a
viati
on
sim
ulati
on
,
3D
gam
ing
,
an
d
surveyi
ng.
O
n
the
m
os
t
popu
la
r
HC
I
to
ols
i
s
H
GR
te
ch
niques.
HG
R
syst
em
i
s
si
m
i
la
r
to
the
bio
m
et
ric
sy
stem
,
bio
m
e
tric
syst
e
m
s
are
consi
sti
ng
of
ba
sic
al
ly
the
fo
l
lowing
sta
ges:
in
pu
t
data,
pr
e
proce
ssing,
feat
ur
e
extracti
on
a
nd
sel
ect
ion,
a
nd
cl
assi
ficat
ion
sta
ges
,
as
du
ll
y
exp
la
ine
d
in
[
2
-
4].
Th
e
basi
c
sta
ges
of
de
sign
i
ng
t
he
H
GR
syst
e
m
are
com
pr
isi
ng
of
th
e
f
ollow
i
ng
:
data
acqu
isi
ti
on,
de
te
ct
ion
,
se
gm
e
ntati
on
a
nd
tr
ackin
g,
featu
re
extracti
on
a
nd
se
le
ct
ion
wi
t
h
the
final
st
age
is
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
6
,
Dece
m
ber
2
01
9
:
5277
-
5285
5278
the
ge
sture
re
cogniti
on
by
us
in
g
var
i
ou
s
cl
assifi
cat
ion
al
gorithm
s
[
5].
I
n
H
GR,
there
is
no
ne
ed
of
a
per
ip
her
al
de
vice
to
interact
with
the
com
pu
te
r
exce
pt
the
ca
m
era,
so
as
to
captur
e
th
e
fr
onte
d
vie
w
to
be
analy
zed
by
diff
e
ren
t
im
ag
e
processi
ng
an
d
arti
fici
al
intel
li
gen
t
to
ols
by
a
pr
ocess
or,
t
hen
act
io
ns
will
be
ta
ken
ac
co
rd
i
ngly
[6
]
.
On
e
of
the
cha
ll
eng
es
of
H
G
R
is
the
han
d
detect
ion
an
d
s
egm
entat
ion
preci
sel
y,
especial
ly
if
there
are
m
any
rand
om
ob
j
ect
s
bes
ide
ha
nd
obj
ec
t.
More
over
,
de
te
ct
ion
op
e
rati
on
will
be
m
or
e
diff
ic
ult
in
t
he
cas
e
wh
e
re
the
ra
nd
om
ob
j
ect
ha
vin
g
col
or
as
the
sam
e
as
han
d
colo
r.
In
t
his
pa
per,
a
new
al
gorithm
is
pr
es
ented
and
ex
plaine
d
on
ha
nd
ext
rac
ti
on
a
nd
segm
entat
ion
by
usi
ng
sca
nn
i
ng
the
obj
ect
im
age
from
le
ft
to
righ
t
an
d
from
top
to
bo
tt
o
m
in
or
der
t
o
scan
ho
w
m
any
flips
from
zero
pixe
l“
0”
as
black
to
on
e
pi
xel
“1”
as
wh
it
e.
In
t
his
rese
arc
h,
t
he
ha
nd
is
assum
ed
to
be
detect
ed
m
us
t
ta
ke
a
f
orm
or
te
m
plate
a
s
show
n
i
n
Fi
gure
1.
Othe
rw
ise
,
t
he
obj
ect
will
not
be
cl
assifi
ed
as
a
hand
ob
j
e
ct
.
This
is
the
assum
ption
of
the
curre
nt
al
gorithm
,
in
wh
ic
h
it
s
accuracy
de
pe
nds
on
the
Fli
p_N
um
ber
,
w
hi
ch
will
be
set
based
on
t
he
require
d
com
plexity
.
In
oth
e
r
w
ords
,
Fli
p_Nu
m
ber
is
con
si
der
e
d
a
s
the
de
gr
ee
of
the
com
plexity
,
la
rg
er
Fli
p_Nu
m
ber
is
set
,
bette
r
hand
obj
ect
is
pr
e
dicte
d.
M
ore
detai
ls
of
the
al
gorithm
will
be
i
ll
us
tra
te
d
in
the
m
eth
od
ology
sect
ion
a
nd
exp
e
rim
ents are
co
nducted
as
well
, to
test
t
he
correctne
ss
of
the pr
opos
e
d
al
gorithm
.
F
ig
ure
1. Ha
nd
te
m
plate
d
epe
nd
e
d f
or
t
he de
te
ct
ion
poi
nted
out as
5
-
fin
ger
It
is
w
or
t
h
to
m
ention
that,
the
idea
of
this
pap
e
r
as
flips
nu
m
ber
base
d
on
the
sm
art
pa
th
has
bee
n
adap
te
d
from
[
7
]
,
in
w
hich
th
e
ori
gin
w
ork
in
this
pa
pe
r
w
as
ex
plo
it
in
g
s
m
art
path
t
o
c
ount
num
ber
s
by
hand
gestu
re
as
0,1
,
2,3,4,
a
nd
5.
Accor
dingly
,
a
m
od
ific
at
ion
has
bee
n
done
to
outc
om
e
a
ne
w
ver
si
on
to
be
m
uch
m
or
e
su
it
able
for
our
propos
ed
w
ork
as
the
fo
ll
owin
g
hypothesis:
if
ther
e
are
five
fin
ge
rs
,
so
it
m
ean
s
it
is
a
hand
ot
herw
ise
it
is
no
t
a
hand
e
ven
if
t
her
e
a
re
f
our
f
ing
e
rs
are
po
i
nted
ou
t.
T
he
m
od
ific
at
ion
will
be
exp
la
ine
d
i
n
t
he
m
et
ho
dolo
gy
su
bse
ct
ion
.
The
pro
po
s
ed
m
et
ho
d
does
not
nee
d
data
t
rainin
g,
w
hich
is
an
adv
a
ntage
that
m
akes
the
syste
m
reli
able
fo
r
e
m
bed
de
d
syst
e
m
s
and
li
gh
twei
ght
dev
ic
es
.
Howe
ver
,
the
wea
k
po
i
nt
of
this
a
lgorit
hm
that
t
he
hand
detect
ion
wor
ks
f
or
on
ly
fi
ve
fin
ger
s
are
point
ed
ou
t,
f
or
i
nst
ance,
a
pe
rson
has
a
cut
fi
ng
e
r,
it
m
igh
t
not
be
w
orki
ng
pro
pe
rly
,
or
it
re
quires
to
in
f
or
m
the
s
yst
e
m
ad
m
inistr
at
or
,
la
te
r
on,
to
c
ha
ng
e
t
he
par
a
m
et
er
of
the
de
te
ct
ion
f
ro
m
five
to
f
our
by
si
m
ply
chan
ging
the
fi
lp
num
ber
.
The
orga
nizat
io
n
of
this
pa
pe
r
is
as
fo
ll
ow
s;
Sect
ion
2
re
vi
ews
li
te
rat
ur
e
r
el
at
ed
to
hand
detect
ion,
Sect
ion
3
exp
la
in
s
the
m
et
ho
dolo
gy
of
the
pro
pos
ed
te
c
hn
i
qu
e
with
te
sti
ng
a
nd
an
al
ysi
s,
Sect
ion
4
de
s
cribes
the
exp
e
rim
ent
;
Sect
ion
5
pre
sents
the
res
ults
and
discuss
i
ons.
Fin
al
ly
,
the
con
cl
usi
on
of
this
researc
h
and
it
s
po
s
sible
fu
t
ur
e
work a
re
pr
ese
nted
i
n
Sect
i
on 6
.
2.
LIT
ERATUR
E REVIE
W
Hand
detect
io
n
idea
s
for
p
re
vious
w
orks
are
li
st
ed
i
n
t
his
sect
io
n
with
thei
r
m
et
ho
do
l
og
ie
s
a
nd
at
tribu
te
s.
F
or
instance,
in
[
8
]
,
h
an
d
detect
ion
is
desig
ne
d
accor
ding
to
ha
nd
m
otio
n
bas
ed
on
FIFO
to
detec
t
foregr
ound
ha
nd
a
nd
non
-
ha
nd
in
form
at
ion
.
This
idea
is
based
on
the
se
ver
al
co
ns
ec
utive
dif
fer
e
nce
i
m
ages
thr
ough
the
F
I
FO
a
nd
path
ov
e
rlap
,
an
d
t
hen
t
he
ou
t
pu
t
is
com
bin
ed
with
KC
F
(
Ke
rn
el
iz
ed
C
orre
la
ti
on
Fil
te
r
)
on
H
O
G
in
orde
r
to
i
m
pr
ov
e
the
tra
ckin
g.
A
nothe
r
work
is
prese
nted
in
[
9
]
,
thi
s
work
c
onsist
s
of
4
ste
ps
: hea
d
det
ect
ion
ope
rati
on
, back
pro
j
ect
ion
, ha
nd
ro
ta
ti
on
,
an
d
the
n
ha
nd
detect
ion.
Her
e
, th
e hum
an
hea
d
inf
o
rm
at
ion
as
colo
r
is
us
e
d
to
be
assist
a
nc
e
of
hand
R
OI
s
detect
io
n,
by
us
in
g
th
e
featur
e
ext
racti
on
a
s
a
histogram
of
or
ie
nte
d
gra
dient
(H
O
G)
featur
e
a
nd
Sup
port
Vecto
r
Ma
chine
(
SV
M
)
as
a
cl
assifi
er.
I
n
2016
,
a
ha
nd
wit
h
wr
ist
detect
io
n
m
et
ho
d
f
or
unobt
r
us
ive
hand
gest
ur
ei
s
re
portedi
n
[
10
]
,
the
ope
rati
on
is
i
m
ple
m
ented
b
y
usi
ng
a
he
ad
m
ou
nte
d
disp
la
y
(
HM
D
)
wh
e
re
loc
at
es
in
uppe
r
body
a
rea
of
a
us
e
r,
and
a
dep
t
h
ca
m
era
unde
r
a
n
HMD
t
o
e
xtrac
t
the
s
hap
e
co
nt
ext
featu
res
a
nd
S
VM
f
or
th
e
cl
assifi
er
.
A
nothe
r
hand dete
ct
ion
us
in
g
facial
in
f
or
m
at
io
n
is pr
e
sented
i
n
20
16
in
[
11
]
, here
de
te
ct
ion
of a f
ac
e is the f
irst st
e
p
t
o
pick
up
the
fac
e
colo
r
s
o
a
s
to
be
us
e
d
for
re
gions
of
inter
e
st
(RO
I)
e
xtrac
ti
ng
to
detect
ha
nds,
s
pecial
ly
ha
nd
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Han
d detec
ti
on and se
gm
e
nta
t
ion
us
i
ng smar
t pa
t
h
tr
ackin
g f
ing
ers
as f
eat
ur
es
and…
(
Kha
le
d
N.
Y
as
en
)
5279
sk
in
c
ol
or
a
nd
face
s
kin
c
olor
are
al
m
os
t
sim
il
ar,
her
e,
su
c
cessf
ul
detect
ion
rate
is
up
t
o
92
%
.
Anothe
r
ha
nd
detect
ion
pr
e
se
nted
recently
in
20
18
in
[
12
]
,
w
hich
is
bas
ed
on
t
he
co
nv
olu
ti
onal
ne
ur
a
l
networ
k
as
a
dee
p
le
arn
in
g,
T
his
te
chn
iq
ue
is
ba
sed
on
the
a
r
chite
ct
ur
e
of
YO
L
O
by
util
iz
ing
the
s
patia
l
-
trans
fer
c
onne
ct
ion
(S
TC)
betwee
n
hi
gh
-
le
vel
la
ye
rs
an
d
lo
w
-
l
evel
la
ye
rs,
th
e
m
ulti
-
scal
e
featur
es
from
diff
ere
nt
la
ye
rs
can
be
aggre
gated
f
or
detect
ing
the
ha
nds.
A
nothe
r
work
f
or
ha
nd
detect
ion
ba
se
d
on
sta
ti
sti
cal
l
earn
in
g
trai
ning
way
is
introdu
ce
d
i
n
[
13
]
,
i
n
w
hi
ch
this
idea
w
as
te
ste
d
by
Using
Mi
cros
of
t'
s
Kinect
se
ns
or
dataset
,
w
hich
i
s
the
sam
e
database
of
the
pro
posed
w
ork
in
this
pa
pe
r
as
well
,
here
featu
res
for
sta
ti
sti
cal
le
a
rn
i
ng
wh
ic
happ
roxi
m
at
es
with
a
Harr
-
li
ke
featu
re
wit
h
the
hel
p
of
A
dabo
os
t
sta
ti
sti
cal
le
a
r
ning,
gets
t
he
trai
ning
m
od
el
.
Fu
rthe
rm
or
e,
idea
of
hand
detect
io
n,
w
hic
h
is
use
d
an
e
xten
de
d
histo
gram
o
f
ori
ented
gr
a
dient
s
(HOG)
m
od
el
nam
ed
sk
in
co
lor
histo
gr
am
of
ori
ente
d
gr
a
dients
(
SC
HOG
)
is
prese
nte
d
in
[
14
]
to
c
onstr
uct
a
hum
an
hand
detect
or,
fi
rstly
,
featu
res
bas
ed
on
SCH
O
G
are
e
xtracted
by
com
bin
in
g
HOG
with
s
ki
n
c
olor
cues,
the
n
s
up
port
vecto
r
m
a
chine
(SVM)
a
lgorit
hm
is
us
ed
for
trai
ning
t
he
dataset
a
nd
finall
y,
this
m
et
hod
is
ver
ifie
d
on
the
te
sti
ng
da
ta
set
fo
r
the
SCHOG
featu
res
.
T
he
h
a
nd
is
al
so
detect
ed
in
2014
i
n
[
15
]
,
by
e
m
plo
yi
ng
a
cor
ner
detect
or
to
fig
ur
e
ou
t
the
pro
ble
m
of
the
finge
r
fr
a
gm
ent
occu
r
red
durin
g
hand
detect
ion, t
he p
ro
ces
s
of t
his det
ect
or
is s
hr
i
nkin
g
the
RO
I
i
nto
a
m
uch
sm
al
le
r
range
w
hile per
form
ing
corner
detect
ion.
A
no
ther
w
ork
of
hand
detect
io
n
util
iz
ed
s
kin
colo
r
filt
erin
g
m
et
ho
d
base
d
on
sk
i
n
c
olor
range
m
od
el
ed
in
Y
CbCr
c
olor
s
pa
ce
as
in
[
3
,
16
]
.
It
is
w
or
t
h
t
o
add
that,
t
he
pr
opos
e
d
m
et
ho
d
desi
gn
e
d
acc
ordi
ng
to
the
c
olor
sk
i
n
m
et
ho
d
as
a
first
ste
p
t
hen
s
econdly
,
e
xam
i
ning
the
obj
ect
that
de
pendin
g
on
Fli
ps
_N
um
ber
s
wh
ic
h wil
l be e
xp
la
ine
d
i
n
t
he
n
e
xt s
ub
sect
io
n.
3.
THE
PROPO
SED
METHO
D
The
pr
ocess
st
arts
by
extracti
ng
fr
am
es
by
fr
am
e
fr
om
the
vid
e
o
stream
t
o
be
processe
d
separ
at
el
y.
The
sta
ge
s
as
a
blo
c
k
diagr
am
of
t
he
pr
opos
e
d
ha
nd
detect
ion
a
re
s
how
n
i
n
Fig
ure
2.
A
ft
er
extr
act
ing
fra
m
e,
searchi
ng
ope
r
at
ion
for
s
kin
colo
r
ba
sed
on
ranges
of
RG
B
colo
r
s
pace
will
be
sta
rte
d.
The
ra
ng
es
of
red,
gr
ee
n
a
nd
blu
e
s
are
m
od
el
ed
es
pecial
ly
for
wh
it
e
pe
op
le
.
I
n
oth
e
r
wor
ds
,
these
ra
ng
es
a
re
no
t
a
ppli
cab
le
to
dark
s
kin
c
olor
pe
ople
,
th
us
searchi
ng
oper
at
ion
ba
sed
on
the
ra
ng
e
of
the
s
kin
c
olo
r
is
i
m
ple
m
ented
an
d
il
lustrate
d
as a
bor
der
e
d box a
rou
nd the
ha
nd as in
Fig
ur
e
3(
1).
Fig
ure
2. Me
th
odology st
e
ps
of h
a
nd
detect
ion an
d segm
entat
ion
The
ra
ng
es
f
or
each
col
or
s
pa
ce
(R
GB)
a
s
red,
gr
ee
n
a
nd
bl
ue
a
re
sta
te
d
belo
w
,
these
ra
ng
are
m
odel
ed
especial
ly
f
or t
he
c
olor s
kin
:
103
<
re
d
_
c
olor
_
range
<
159
74
<
g
re
en
_
co
lor
_
range
<
103
43
<
blue
_
col
or
_
range
<
98
Af
te
r
wa
rd,
seve
ral
substage
s
of
pre
-
processi
ng
i
n
te
rm
s
of
i
m
age
proces
sing
s
uc
h
as
m
edian
filt
erin
g
a
nd
rem
ov
ing
s
ome
obj
ect
that
t
heir
are
as
are
up
t
o
30
0
pix
el
s
(as
ver
y
sm
a
ll
ob
j
ect
a
rea)
,
t
hen
ap
pl
yi
ng
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
6
,
Dece
m
ber
2
01
9
:
5277
-
5285
5280
i
m
age
dilat
e
m
or
phologica
l
filt
er
s
in
ord
er
to
sm
o
ot
hly
m
ake
the
obj
ect
c
onnecte
d.
Fig
ur
e
3(
c
)
de
picts
the
af
or
em
entione
d
ste
ps
of
i
m
age
processi
ng,
as
it
is
s
how
n
t
hat
there
are
on
ly
5
ob
j
ect
s
in
Fig
ur
e
3(
c
)
,
on
e
of
them
surel
y
is
the
hand.
Fi
nally
,
a
bo
rd
e
r
is
dr
a
w
n
a
rou
nd
t
he
hand
for
t
he
o
rigi
na
l
i
m
age
fr
am
e
to
set
the
ta
rg
et
re
gio
n
of
inte
rest
(
ROI)
by
extrac
ti
ng
the
f
our
bor
der
po
i
nts,
then
plo
tt
ed
to
the
or
i
gin
al
im
age
as
sh
ow
n
in
Fi
gur
e 1(a
).
(a)
(b)
(c)
Figure
3. Pict
ori
al
il
lustrati
on of the
m
et
ho
dolo
gy steps
as i
m
ages
3.1.
Prop
os
ed
de
te
cted
algorith
m
On
ce
the R
OI
(
obj
ect
)
is acces
sed, s
te
ps o
f
th
e
pro
posed
al
gorith
m
are
as
f
ollow
i
ng
:
1.
Get P
oin
t
(
x,y)
i
n
T
op Mi
ddle
of Ob
j
ect
I
m
age,
an
d n
ot at to
p
m
idd
le
of the
obj
ect
.
2.
Sli
din
g d
own
of this
point
qu
a
rterly
to be
refe
ren
ce
point
(
R
ef
_Pn
t
).
3.
Dra
w rig
ht li
ne
slo
p
f
r
om
(
Ref
_P
nt
).
4.
Dr
a
w
le
ft
li
ne
slop f
ro
m
(
Ref
_P
nt
).
5.
Me
rg
e
le
ft & ri
gh
t
slo
ps
to
b
e
a
sm
art
scann
e
d path.
6.
Ca
lc
ulate
Fli
p_
Nu
mber
base
d on
ly
on the
s
cann
e
d path
P
seu
do
-
c
od
e
de
scribin
g 1 &
2
for
e
xtracti
ng the
Ref
_P
nt
i
s as foll
owin
g
:
[
x_
m
axy_m
ax
]
=size(
img)
;
top_ref_
point_
y=fl
oo
r(
y
_max
/2)
;
top_ref_
point_
x=1;
y_ref=to
p_ref
_poi
nt_
y
;
qu
rt
_top
_x=
fi
x
(
(
x_m
ax/8))
;
f
or
i=
1:q
ur
t
_top_
x
for
j=1:y
_max
if
((
img(i
,j)
==
1))
x_ref=i;
e
nd
e
nden
dre
f_pnt=[
x_ref
y_ref
]
;
wh
e
re
Ref
_P
nt
is
the
po
int
tha
t
has
the
trajector
ie
s
(
x
)
and
(
y
)
to
be
con
side
red
as
a
ref
e
re
nce
of
the
br
a
nc
hing
the
tw
o
sl
o
ps
l
eft
an
d
rig
ht.
Figure
4
de
pic
ts
al
l
the
five
obj
ect
s
t
hat
a
r
e
possible
to
r
epr
ese
nt
ha
nd
as
these
obj
ect
s
passe
d
thr
ough
the
sk
i
n
colo
r
filt
er.
Nex
t
sta
ge,
e
xam
ining
operat
ion
is
app
li
ed
base
d
on
the p
r
opos
e
d
al
gorithm
to
extract
the
tr
ue
hand
obj
ect
a
m
on
g
the
oth
e
rs.
I
t
is
obvi
ous
that
al
l
obj
ec
ts
hav
e
t
he
red
ci
rcle
sy
m
bo
l,
w
hich
is
deem
ed
as
the
to
p
po
i
nt
th
en
by
getti
ng
dow
n
a
rou
nd
ha
lf
-
qua
rterly
,
th
e
blu
e
ci
rcle
sym
bo
l
is
con
si
der
e
d
as
the
Ref
_P
nt
.
At
this
poi
nt,
the
tw
o
sl
op
s
are
br
a
nc
hing
le
ft
an
d
rig
ht.
The
rea
so
n
f
or
br
a
nc
hing
is
im
po
rtant
to
ful
ly
dissect
the
obj
ect
ty
pe
an
d
to
co
un
t
how
m
any
flips
t
he
obj
ect
has
durin
g
the
scan
ning
operati
on.
Af
te
r
that,
point
3
of
the
al
gorithm
is
sp
eci
fyi
ng
t
he
rig
ht
slo
ps
.
Firstl
y,
the
ide
a
of
extracti
ng
th
e
rig
ht
slo
p
is
by
increm
enting
one
(
y
)
to
bot
h
rows
an
d
c
ol
um
ns
to
get
n
e
w
e
xtracted
sc
ann
e
d
rig
ht
slo
p
as
s
how
n
in
Fi
gur
e
4
with
gree
n
colo
r
slo
p
,
t
he
al
gorithm
pro
gr
am
m
i
ng
as
pse
udo
-
co
de
is
sh
ow
n
unde
r
the:
right
_s
lo
p
pse
ud
o
-
c
od
e
.
Sec
ondly
,
ste
p
4
of
the
pro
po
se
d
te
ch
nique
is
to
dr
a
w
a
le
ft
slop
sta
rtin
g
from
Ref
_P
nt
go
i
ng
dow
n
to
the
le
ft
e
nd
of
the
im
age
as
s
how
n
i
n
Fi
gur
e
4
of
t
he
pink
colo
r
sl
op
,
the
idea
of
extracti
ng the l
eft slop
is
b
y d
ecrem
enting
one to im
age col
um
ns
a
nd
inc
r
e
m
enting on
e t
o
the
rows
to
ge
t
new
extracte
d sca
nned
left slo
p
,
t
he
al
go
rithm
p
rogr
am
m
ing
as
p
se
udo
-
co
de
i
s sho
wn
le
ft
_sl
op
pseud
o
-
co
de
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Han
d detec
ti
on and se
gm
e
nta
t
ion
us
i
ng smar
t pa
t
h
tr
ackin
g f
ing
ers
as f
eat
ur
es
and…
(
Kha
le
d
N.
Y
as
en
)
5281
(1)
(2)
(3)
(4)
(5)
Figure
4. De
pi
ct
ing
Ref
_P
nt
with left a
nd
righ
t
slo
ps
as
sm
art p
at
h of
a
n
i
m
age obj
e
ct
right_sl
op
pse
udo
-
co
de
righ
t
_s
lo
p_
x
=
0;
ri
gh
t
_s
lo
p_y
=0;
rx
=0;ry
=0;si
ze_slop
=1;
fo
r y=
1:size
_sl
op
e
:(y_
ma
x
-
y_ref)
-
1
y
_n
ew
=y_ref
+y;
ry
=ry+
1;
ri
gh
t
_s
lo
p_
y(
ry)
=y_
new;
x
_n
ew
=x_ref
+y;
rx
=rx+
1;
ri
gh
t
_s
lo
p_
x(
rx)
=
x_
new
;
e
nd
plo
t(
righ
t
_s
l
op_x,ri
ght_slo
p_
y)
;
le
ft _
slo
p
pseu
do
-
code
size
_s
lo
p=1
;
le
ft
_s
lo
p_x=
0;
le
ft
_s
lo
p_y=
0;
lx
=
0;
ly
=0;
le
ft
_s
ize
=y
_max
-
le
ngth(
rig
ht_
sl
op_y)
;
w
hile (
le
ft
_s
i
ze~=0)
le
ft
_s
ize
=lef
t_size
-
1;
lx
=lx
+1;
x
_n
ew
=x_ref
+lx;
le
ft
_s
lo
p
_x(l
x
)
=
x_n
ew
;
ly
=ly+
1;
y
_ne
w=y_ref
-
ly
;
le
ft
_s
lo
p_
y(
ly
)
=y_
new;
en
d
pl
ot(
le
ft
_s
lop_x,lef
t_slo
p_y)
;
Af
te
r
s
pecifyi
ng
the
rig
ht
an
d
le
ft
slop
an
d
m
erg
in
g
them
to
be
one
sca
nned
(sm
art)
path,
now
it
is
re
ady
to
extract
the
Fli
p_
N
umber
base
d
sca
nn
e
d
pat
h.
Her
e
,
to
gu
aran
te
e
dr
a
wing
slo
ps
with
out
m
issi
ng
an
y
oth
e
r
fin
ger
of
the
ha
nd
to
be
dete
ct
ed.
I
n
this
sit
uation
,
the
ide
a
is
extracti
ng
the
Fli
p_
N
umb
er
from
the
rig
ht
and
le
ft
slop
s,
a
nd
then
a
dd
s
t
hem
tog
et
he
r
to
be
the
final
Fli
p_
Numb
er
of
the
obj
ect
.
T
he
Fli
p_
N
umber
is
de
fine
d
as
the
pi
xel
bri
gh
t
ness
cha
nging
f
ro
m
“0”
t
o
“1”
or
“
1”
t
o
“0”
if
a
ny
cha
ng
i
ng
happe
ne
d,
t
he
n
a
co
unte
r
will
be
inc
rem
ent
by
on
e,
e
ve
ntu
al
ly
,
this
counter
will
be
repres
ented
as
th
e
Fli
p_
N
umber.
T
he
two
ps
eu
do
-
co
de
s
of extracti
ng th
e
Fli
p_
N
umber
of
bo
t
h
ri
gh
t a
nd left slo
ps ar
e b
el
ow:
Fl
ip_
N
um
ber
fro
m
righ
t
sl
op
fst_v
alue
=
obj_im
g(ref_pnt
(
1,
2)
,
ref_pnt
(
1,
1))
;
[
x
_num,y
_num
]
=
size(
obj_i
mg
)
;fl
ip_num_right=0;
for p=1:l
ength(
right_slop_x
)
-
2
i
fx
_num>
righ
t_
slop_x(
p)
&
&
y
_num>
right_slop_y
(
p)
i
f
(
fst_v
alu
e ~=
obj_i
mg(
right_slop_x
(
p)
,
right_sl
op_y(
p)))
fst_
val
ue=
obj
_img(
right_slop
_x(
p)
,
right_slop
_y(
p))
;
f
li
p_num_r
igh
t=fl
ip_num_righ
t+1;
end
end
end
Fl
ip_
N
um
ber
fro
m
left
slo
p
fl
ip_num_
le
f
t=0;
fst_v
a
lue
= ob
j
_img(
ref_pnt
(
1,
2
)
,
ref_pnt
(
1,
1))
;
for k
=1:
le
ngth
(
le
ft
_slop_x)
-
2
ifx_num>
le
ft
_s
lop_x
(
k)
&
&
y
_
n
um>
le
ft
_slop_y(
k)
if (
fst_v
a
lue ~
= obj
_img(
le
ft
_s
lop_x
(
k)
,
le
ft
_slo
p_y(
k))
)
fst_v
a
lue
= ob
j
_img(
le
ft
_slop_x
(
k)
,
le
ft
_slop_y(
k))
;
f
li
p_num_le
ft
=f
li
p_num_left+1;
end
end
end
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
6
,
Dece
m
ber
2
01
9
:
5277
-
5285
5282
Now, sum
m
at
i
on b
et
ween
flip
_
n
u
m_
left
an
d
fli
p
_
n
u
m_
right
to
pr
oduce
Fli
p_
N
umber
featu
re as
:
fl
ip_
num
ber
=fl
ip_num
_left
+fl
ip_num
_r
ig
ht;
Nex
t,
cl
assifi
c
at
ion
is
respo
nsi
ble
to
detect
wh
ic
h
obje
ct
is
hand
an
d
no
n
-
hand,
as
ass
ume
d
the
ha
nd
to
be
detect
ed
is
sh
ow
n
in
Figure
1,
wh
ic
h
has
fi
ve
fin
gers
are
po
i
nted
out.
A
cco
rd
i
ng
l
y,
that
i
m
age
if
thi
s
pro
po
se
d
al
go
rithm
is
app
li
ed
to
it
,
the
Fli
p_Nu
m
ber
m
us
t
equ
al
to
10
flips.
Howev
e
r,
10
flips
are
chall
eng
i
ng
be
cause
s
om
eti
m
es
pro
blem
ris
es
relat
ed
to
i
m
age
processi
ng
filt
ering
a
nd
sk
i
n
col
or
i
ng
search
.
Ther
e
f
or
e,
dec
reasin
g
this
c
ha
ll
eng
e
10
t
o
8
or
6
Fli
p_N
um
ber
to
be
de
te
ct
ed
is
pr
e
f
erab
le
.
H
ow
e
ve
r,
on
ce
decr
easi
ng the
chall
enged
Flip_N
um
ber
, th
e
False
Accept
(
FA
)
w
il
l
be
inc
reased
.
An
ex
pe
rt
syst
e
m
is
def
i
ned
as
a
c
om
pu
te
r
syst
e
m
that
em
ula
te
s
the
de
ci
sion
-
m
akin
g
a
bili
ty
of
a
hum
an
exp
e
r
t.
In
r
ule
-
base
d
ex
per
t
syst
em
s,
f
orwa
rd
chai
ning
in
fer
e
nce
te
ch
niques
is
use
d
in
this
rese
arch.
The
do
m
ai
n
knowle
dge
is
r
epr
ese
nted
by
a
set
of
I
F
-
T
HEN
in
orde
r
to
pr
oduce
r
ules
an
d
t
he
data
is
represe
nted
by
a
set
of
facts
about
the
cur
re
nt
sit
ua
ti
on
,
w
hic
h
is
rep
rese
nte
d
by
featur
e
nam
ed
the
Fli
p_num
ber
.
T
he
i
nf
e
ren
ce
en
gin
e
m
ust
decide
w
hen
the
r
ules
m
us
t
be
e
xec
uted.
Fo
r
wa
rd
chai
ni
ng
is
us
e
d
in
this
pa
per
becau
s
e
of
the
sim
il
ari
ty
to
the
m
e
t
hodolo
gy
that
dep
e
nds
on
dat
a
-
dr
ive
n
re
aso
ning
.
The
reas
on
i
ng
sta
rts
fr
om
the
known
data
an
d
procee
ds
f
orward
with
that
data.
Each
ti
m
e,
on
ly
the
top
r
ule
is
execu
te
d,
a
nd
wh
e
n
exec
ute
d,
the
ru
le
ad
ds
a
new
fact
to
the
databa
se.
A
ny
ru
le
can
be
execu
te
d
only
on
ce
.
The pse
udo
-
co
de of
the e
xper
t sy
stem
is sh
own
b
el
ow
:
if
(f
l
ip_
nu
mb
er
>=
10)
disp
(
'H
an
d Dete
ct
ed
'
)
;
el
se
disp
(
'
N
o d
et
ec
ti
on'
)
;
end
As
it
is
show
n
in
Fig
ur
e
4,
there
a
re
five
obj
ect
s
gen
e
ra
te
d
duri
ng
pre
proc
essi
ng
a
nd
filt
ering
t
he
i
m
age,
the
al
gorithm
sh
oul
d
be
ap
pl
ie
d
to
al
l
generate
d
ob
j
ect
s
to
extract
the
Fli
p_
Nu
m
ber
f
eat
ur
e
an
d
on
l
y
on
e
obj
ect
is
pr
e
dic
te
d
by t
he
e
xpe
rt syst
em
as the
true
h
a
nd am
ong
t
he othe
rs.
4.
E
X
PERI
MEN
T
To
e
valuate
t
he
propose
d
m
et
ho
d,
ex
pe
r
i
m
ents
hav
e
be
en
c
onduct
ed
on
100
im
ages,
w
hic
h
is
exp
l
oited
a
da
ta
set
na
m
ed
Dataset
of
Le
ap
Moti
on
an
d
Mi
cro
s
of
t
Kinect
ha
nd
a
cqu
isi
ti
ons
issued
in
Un
i
ver
s
it
y
of
Padova
(
Ital
y),
2014
[
17
]
.
T
he
siz
e
of
ea
c
h
im
age
her
e
is
Kinect
c
olor
m
ap
(12
80
x
960).
These
im
ages
are
r
gb.
png
e
xt
ensio
n.
T
his
da
ta
set
con
ta
ins
10
ge
sture
s
in
ge
ner
al
,
of
di
f
fer
e
nt
ha
nd
ge
sture,
li
ke
the
ass
umpti
on
of
the
c
urre
nt
resea
rch
i
s
to
detect
a
ha
nd
with
a
sty
le
of
five
fin
gers
po
i
nt
ed
out,
s
o
that,
on
e
gest
ur
e
w
hi
ch
is
la
beled
ge
sture
9
in
this
database
is
sel
ect
ed
only
f
or
the
10
in
div
i
dual
s.
I
n
ot
her
w
ords
,
te
n
rgb.p
ng
im
ages
ha
ve
been
us
ed
for
eac
h
ind
ivi
du
al
,
t
otall
y
10
0
im
ages
exp
l
oited
in
th
e
te
st.
It
is
wo
r
th
t
o
m
ention
that
i
n
this
al
gorithm
there
is
no
ne
ed
to
tr
ai
n
im
a
ges.
J
us
t
te
st
a
ny
im
age,
so
that
al
l
the
10
0
i
m
ages
are
us
e
d
for
te
sti
ng
to
extrac
t
the
total
acc
ur
acy
.
U
su
al
ly
,
in
the
ver
ific
at
ion
or
identi
ficat
ion
com
par
iso
n,
there
a
re
t
wo
po
s
sible
e
rror
s
to
be
m
easur
e
d:
False
Acce
pt
Ra
te
(F
AR
)
e
rror,
w
hic
h
re
s
ults
f
ro
m
the
f
orge
d
tem
plate
that
a
ccepte
d
by
a
syst
e
m
falsel
y
du
ri
ng
te
sti
ng
a
nd
t
he
seco
nd
error
is
False
Re
j
ect
ion
Ra
te
(F
RR
),
wh
ic
h
res
ults
from
the
gen
ui
ne
te
m
plate
that
the
co
m
pu
te
r
pr
e
dic
te
d
it
wrongl
y
[
18
-
20
]
.
Ge
ner
al
ly
,
the
overall
accuracy
of
the
pro
posed
syst
e
m
is
cal
culat
e
d
by
subtract
i
ng
t
he
ave
rag
e
error
rate
f
r
om
10
0%
as in
(1):
2
%
100
%
F
R
R
FAR
A
c
c
ur
ac
y
(1)
Howe
ver,
this
researc
h
has
no
FA
R
er
r
or
,
si
nce
there
a
re
no
f
orge
ha
nd
i
m
ages
in
this
exp
e
rim
ent
to
be
te
ste
d.
T
he
refor
e
,
FA
R
is
con
si
der
e
d
to
be
zer
o.
H
ow
e
ve
r,
FRR
is
us
ed
f
or
t
he
te
sti
ng
to
assess
the
rec
ogniti
on
rate,
beca
use
these
ha
nd
i
m
ages
are
co
ns
i
de
red
a
s
ge
nu
i
ne
te
m
plates.
I
n
case
th
ey
are
wrongly
rec
ognized
by
the
propose
d
syst
em
,
then
t
he
FRR
increases
.
T
he
equ
at
i
on
s
t
hat
are
us
e
d
t
o
m
easure
the accu
racy
of this re
searc
h
a
re in (
2) an
d (3):
%
%
100
%
F
R
R
A
c
c
u
r
a
c
y
(2)
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Han
d detec
ti
on and se
gm
e
nta
t
ion
us
i
ng smar
t pa
t
h
tr
ackin
g f
ing
ers
as f
eat
ur
es
and…
(
Kha
le
d
N.
Y
as
en
)
5283
%
1
0
0
_
_
Re
_
_
%
A
t
t
e
m
p
t
T
r
u
e
T
o
t
a
l
j
e
c
t
F
a
l
s
e
T
o
t
a
l
F
R
R
(3)
Ma
tl
ab
2016
b
as
a
w
orksta
ti
on
has
bee
n
us
e
d
i
n
this
exp
e
rim
ent
insta
ll
ed
into
a
com
pu
te
r,
w
hich
has
the foll
owin
g
c
har
act
erist
ic
s c
or
e
2due
, 4
G
-
RAM.
5.
RESU
LT
S
A
ND
DI
SCUS
S
ION
The
repor
te
d
r
esults
in
t
his
re
search
fall
into
two
ty
pe
s:
pic
torial
an
d
sta
ti
sti
cal
resu
lt
s.
The
pictorial
resu
lt
is
dep
ic
t
ed
in
Fig
ur
e
5,
in
w
hich
it
co
ntains
5
im
ages,
the
first
on
e
in
Fig
ur
e
5(a)
,
wh
ic
h
is
the
de
te
ct
ed
hand
obj
ect
as
ROI
s
uccessful
ly
.
In
Fig
ur
e
5(b
)
il
lustrate
s
the
outp
ut
of
th
e
RGB
colo
r
sk
in
sea
rch
bas
ed
on
RGB
ra
ng
e
d
a
s
af
or
em
entione
d
in
t
he
m
et
ho
dolo
gy
sect
io
n.
It
is
cl
ear
th
at
con
ta
ins
m
a
ny
ob
j
ect
s
an
d
noise
,
after
rem
ov
ing
noise
us
in
g
s
om
e
pr
ep
r
oces
sing
to
ols
s
uc
h
as
m
edian
a
nd
m
or
phol
og
i
cal
filt
er
the
r
esult
is
il
lustrate
d
in
Figure
5(
c
),
he
re
it
is
cl
ear
f
inall
y
con
ta
ins
two
ob
j
ect
s
only
,
in
w
hich
s
ur
el
y
on
e
of
th
e
m
is
the h
a
nd a
nd th
e o
the
r
is a
no
n
-
ha
nd
obj
ect
.
(a)
(b)
(c)
(d)
(e)
Figure
5. A
pply
ing
the
pr
opose
d
al
gorithm
o
f
a
dataset
sa
m
ple
i
m
age and
it
s
pictorial
re
su
lt
s
Af
te
r
a
pp
ly
in
g
the
pr
op
os
e
d
al
gorithm
to
c
ount
the
Fli
p_Number
a
nd
e
xam
ine
the
ob
j
e
ct
s,
wh
ic
h
a
re
extracte
d
in
th
e
im
age,
resu
lt
s
will
be
ou
tc
om
e
ei
ther
“N
o
Detect
ion
”as
sh
own
in
the
Figure
5(d
)
be
cause
,
it
is
cl
ear
the
Fli
ps
_N
umber
is
2,
w
hich
does
not
sat
isfy
the
ex
per
t
syst
e
m
con
diti
on
to
be
an
noun
ced
as
the
obj
ect
is
non
-
ha
nd
.
O
r
“Hand
detec
te
d”
resu
lt
as
dep
ic
te
d
Fig
ur
e
5(
e
)
d
ue
to
the
cl
earn
e
ss
of
the
Fli
ps
_N
um
ber
is
10
in
wh
ic
h
sat
isfie
s
the
exp
ert
sy
stem
con
diti
on
t
o
be
anno
unced
as
the
obje
ct
is
the
ha
nd
t
hat
c
on
ta
in
s 5
fi
nge
rs.
In
te
rm
s
of stat
ist
ic
al
resu
lt
s
as
s
how
n
in
Tab
le
1
,
i
n
cas
e
the
Fli
p_N
umber
is
6,
w
hich
is
in
dicat
ed
to
the
easy
pr
edict
io
n,
beca
us
e
due
to
the
i
m
age
captur
i
ng
a
nd
filt
er
that
is
no
t
al
l
the
five
fin
ge
rs
will
be
a
ppeare
d
,
t
her
e
f
or
e
,
pe
rm
issi
b
le
is
al
lowed
to
be
6,
7
or
m
igh
t
be
8
as
le
ss
per
m
issi
ble.
H
ow
e
ve
r,
i
n
cas
e
of
decr
ea
sin
g
the
rigi
d
of
th
e
ha
nd
pr
e
dicti
on
that
m
eans
False
Re
j
ect
io
n
Ra
te
will
be
inc
reas
ed,
this
m
eans
that
obj
ect
s
m
igh
t
be
hand
t
ru
ly
,
a
nd
the
s
yst
e
m
w
rong
ly
re
j
ect
s
them
du
e
to
the
ri
gid
co
ndit
ion
of
the
te
st
an
d
decisi
on
m
aking
.
Howe
ver,
100
im
ag
es
with
a
cl
uttered
bac
kgr
ou
nd
go
t
accuracy
97%
and
84%
in
ca
se
Fli
p_
Nu
m
be
r
set
to
6
an
d
7
res
pecti
vely
.
In
case
Fli
p_
num
ber
is
set
up
to
8
and 10,
t
he res
ult
is rec
orde
d i
n
Ta
ble 1
as
w
el
l. Th
e acc
ura
cy
is 81% a
nd
48% r
e
sp
ect
ive
ly
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
6
,
Dece
m
ber
2
01
9
:
5277
-
5285
5284
Table
1.
Acc
uracy
r
ep
or
te
d w
it
h
f
our
cases
6,7, 8 an
d 1
0 fli
p_num
ber
It
is
no
ti
ced
t
hat
with
la
r
ge
r
Fli
p_N
um
ber
is
set
ti
ng
up
,
le
ss
accurac
y
is
reco
rd
i
ng
.
Howe
ver
,
the
rec
ogniti
on
rate
is
low
es
pecial
ly
with
10
Fli
p_N
um
ber
,
becau
se
the
sel
ect
ing
s
pecifica
ti
on
of
t
he
ha
nd
obj
ect
beco
m
es
rigi
d
a
nd
cha
ll
eng
in
g.As
t
he
auth
or’s
pe
rspect
ive,
the
m
ost
su
it
able
Fli
p_
Nu
m
ber
i
n
te
r
m
s
of
si
m
plici
t
y
is
7
or
8,
as
in
th
e
m
idd
le
betwe
en
seve
rin
g
and
easy
co
nd
it
io
n.
Be
ca
us
e
so
m
eti
m
es
us
ers
have
a
cut
-
fi
ng
e
r
or
m
is
-
counting
as
the
slo
p
is
no
t
passing
th
r
ough
the
fi
nge
rs
due
to
ha
nd
ro
ta
ti
ng.
How
ever,
the
opposit
e
ca
se
that
m
igh
t
an
ob
j
ect
is
not
a
hand
bu
t
it
ha
s
Fli
p_N
um
ber
as
10
or
m
or
e,
s
uch
t
his
ca
se
is
dep
ic
te
d
i
n
Fig
ur
e
6,
in
w
hich
this
obj
ect
ap
pear
e
d
du
rin
g
cond
ucting
t
he
exp
e
rim
ent
as
hand
bu
t
it
is
ind
ee
d
a
non
-
ha
nd
obj
ect
.
T
his
is
consi
der
e
d
a
s
the
wea
k
poin
t
of
t
he
propo
s
ed
al
go
rithm
.
In
the
sam
e
tim
e,
the pr
opos
e
d
al
gorith
m
is su
it
able f
or the li
ghtwei
ght
dev
ic
es, as
t
her
e
is no trai
ning
data
for
the
pre
dicti
on.
Figure
6. O
bje
ct
classi
fied
as
a h
a
nd b
eca
us
e
of c
on
ta
ini
ng
10 f
li
p_
nu
m
ber
s
In
te
rm
s
of
c
om
par
ison
wit
h
se
ver
al
publ
ished
wor
ks
r
ega
rd
i
ng
ha
nd
detect
ion
,
T
a
ble
2
li
sts
the
recent
w
orks
with
their
m
et
ho
do
l
ogie
s
com
par
ed
with
the
pro
po
se
d
w
ork.
It
is
cl
ear
that
fr
om
the
accur
aci
es
li
ste
d
in
Table
2
,
the
propos
ed
al
gorithm
resu
lt
can
offe
r
knowle
dge
co
ntributi
on
by
a
new
m
et
ho
dolo
gy
with a
n
acce
pt
able rec
ogniti
o
n rate
.
Table
2.
T
he
a
ccur
acy
of the
pro
po
se
d
m
et
ho
dolo
gy c
om
par
ed
w
it
h rece
nt se
ver
al
hand
d
et
ect
io
n
w
or
ks
No
.
Metho
d
o
lo
g
y
Accurac
y
Citatio
n
/Year
1
Han
d
Seg
m
en
tatio
n
by
d
eep learnin
g
: CNN
(Co
n
v
o
lu
ti
o
n
al Neural
Net
wo
rk)
96%
[
21
]
/ 20
1
9
2
Han
d
Reco
g
n
itio
n
:
Skin
and
W
rist
De
tectio
n
by
PCA
+
Euclid
ean d
istan
ce.
93%
[
22
]
/ 20
1
9
3
Han
d
Seg
m
en
tatio
n
and
Fing
ertip Tr
acki
n
g
f
ro
m
Dep
th
Ca
m
e
ra
I
m
ag
es
U
sin
g
Deep
Co
n
v
o
l
u
tio
n
al Neu
ral
Net
wo
rk an
d
M
u
lti
-
task
SegNet.
83%
[
23
]
/ 20
1
9
4
Prop
o
sed
m
e
th
o
d
96%
6.
CONCL
US
I
O
N
W
it
h
an
im
pr
ovem
ent
hap
pe
ned
in
c
om
pu
te
r
visio
n
an
d
m
achine
le
arn
i
ng
in
the
fiel
ds
relat
ed
to
hu
m
an
-
c
om
pu
te
r
interact
io
n,
hand
detect
ion
researc
hes
are
bec
om
ing
i
m
portant
am
ong
researc
he
rs.
I
n
this
pap
e
r,
a
new
al
gorithm
has
been
pro
posed
and
te
ste
d
s
o
as
to
pr
e
dict
hand
ver
ses
non
-
hand
ob
j
ect
in
a
n
i
m
age
that
co
nt
ai
ns
a
c
om
ple
x
backg
rou
nd.
The
operati
on
is
kicke
d
off
by
searchi
ng
on
colo
r
s
kin
obje
ct
s,
then
e
xam
inatio
n
operati
on
f
or
eac
h
obj
ect
is
perf
or
m
ed
by
the
pro
po
s
ed
al
gorithm
,
wh
ic
h
is
pr
e
dicti
ng
a
ref
ere
nce
poi
nt
(Ref
_Pnt)
in
the
obj
ect
the
n
drawi
ng
a
ri
gh
t
li
ne
slo
p
from
Re
f_
Pn
t
an
d
drawi
ng
a
le
ft
li
ne
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Han
d detec
ti
on and se
gm
e
nta
t
ion
us
i
ng smar
t pa
t
h
tr
ackin
g f
ing
ers
as f
eat
ur
es
and…
(
Kha
le
d
N.
Y
as
en
)
5285
slop
from
Re
f_
P
nt
the
n
m
erg
in
g
them
to
be
a
sm
art
scann
e
d
path
.
Fin
al
ly
,
com
pu
ti
ng
the
Fli
p_N
um
ber
,
wh
ic
h
is
ba
sed
on
ly
on
t
he
s
cann
e
d
path
,
a
ct
s
th
e
featu
re
of
t
his
syst
em
.
The
c
onduct
e
d
ex
pe
rim
ents
wer
e
perform
ed
by
us
in
g
10
0
ha
nd
i
m
ages
or
i
gi
nated
f
ro
m
ran
dom
10
-
in
di
vidual
ta
ken
f
ro
m
dataset
na
m
ed
Dataset
of
Lea
p
Moti
on
a
nd
Mi
cro
s
of
t
Kine
ct
hand
ac
quisi
ti
on
s.
The
pe
rfor
m
ance
of
t
he
propose
d
al
gorit
hm
is
up
t
o
84%
and
81%
in
case
the
Fli
p_
Nu
m
ber
feat
ure
is
7
an
d
8
res
pecti
vely
.
F
or
t
he
f
uture
work,
the
pro
posed
a
lgorit
hm
m
igh
t
be
de
velo
pe
d
by
en
han
ci
ng
t
he
accu
racy
by
add
i
ng
a
nothe
r
exam
ining
id
ea
to
boos
t t
he han
d object
resu
lt
a
nd incl
ud
i
ng dar
k
s
k
in
h
a
nd
detect
ion
as
w
el
l.
REFERE
NCE
S
[1]
Pati
l,
N.M.
and
S.
Pati
l,
“
Re
v
i
ew
on
real
-
ti
me
EMG
acqui
si
ti
on
and
hand
gesture
rec
ognition
system
,”
in
El
e
ct
roni
cs,
Co
m
m
unic
at
ion
an
d
Aerospace
T
echnolog
y
(ICEC
A),
2017
In
te
rn
a
ti
onal c
onf
ere
n
c
e
of. IE
EE
,
2017
.
[2]
Mala
llah,
F.L
.
,
e
t
a
l.,
“
A
Review
of
Biom
et
r
ic
T
emplat
e
Prote
ct
i
on
Techni
ques
f
or
Online
Hand
writt
en
Signat
ur
e
Applic
a
ti
on
,
”
Int
ernati
onal
Revi
e
w
on
Computers
and
Soft
ware
(
I.
RE
.
CO.
S
.
)
,
8(1
2)
,
2013.
[3]
Mala
llah,
F.
L.,
et
al.,
“
H
y
b
rid
Hand
-
Dire
ctiona
l
Gesture
s
For
Biom
et
ric
Bas
ed
On
Area
Feat
ur
e
Ext
r
ac
t
ion
And
Expe
rt
S
y
st
em
,”
Journal
of
Theoretical
&
Appl
ie
d
Information
Te
c
hnology
,
95(23)
,
2017
.
[4]
Mala
llah,
F.
L.,
et
al
.
,
“
Irre
ver
sible
Biom
et
ri
c
Te
m
pla
te
Prot
ec
t
ion
b
y
Tr
igo
nom
et
ric
Function
,”
Int
ernati
ona
l
Re
v
ie
w
on
Comp
ute
rs
and
Sof
twa
re
(
IRE
COS)
,
11:
p.
1138
-
1146
,
2
016
.
[5]
Sonkus
are
,
J.S.,
et
al.
“
A
rev
ie
w
on
hand
gesture
rec
ognit
ion
syst
em
,”
in
Com
puting
Comm
unic
at
i
on
Control
and
Autom
at
ion
(IC
CUBEA),
2015
I
nte
rna
ti
ona
l
Con
fer
ence
on
,
I
EEE
,
2015
.
[6]
Kaur,
H.
and
J.
Rani
.
“
A
rev
i
ew
:
Study
of
vario
us
te
chni
ques
of
Hand
gesture
r
ec
ogni
ti
on
,”
in
Pow
er
El
ectroni
c
s,
Inte
lligen
t
Contr
ol
and
En
erg
y
S
y
stems
(ICPEIC
ES),
IE
EE Int
ern
at
ion
al
Conf
ere
n
ce
on
,
I
EEE
,
201
6
.
[7]
Sul
y
m
an,
A.A
.
,
et
a
l.,
“
Rea
l
-
T
ime
Num
eri
ca
l
0
-
5
Counti
ng
B
ase
d
On
Hand
-
Finger
Gesture
s
Rec
ogni
ti
on
,”
Journal
of
Theoretical
&
Appl
ie
d
Information
Te
c
hnology
,
95(13)
,
2017.
[8]
W
ang,
Y.
-
R.
,
W
.
-
H.
Li
n
,
and
L
.
“
Yang.
A
fast
hand
motion
det
e
ct
ion
based
on
FIF
O
,”
in
Mac
hi
ne
Le
a
rning
and
C
y
ber
n
etics
(IC
MLC),
2017
Internat
ion
al
Conf
er
enc
e
on.
IE
EE
,
2
017
.
[9]
Kim
,
J.,
J.
Baek,
and
E.
Kim
.
“
A
part
-
based
rotati
onal
inv
a
riant
hand
det
e
ct
ion
,”
in
Fuzzy
Th
eor
y
and
I
ts
Applic
a
ti
ons (i
F
UZZY),
2013
In
te
rna
ti
ona
l
Conf
ere
nc
e
on
.
IE
EE
,
2013
.
[10]
Oh,
J.Y.,
e
t
al
.
“
A
hand
and
wris
t
detec
t
ion
method
for
uno
btrusive
hand
g
esture
in
te
ract
io
ns
using
HMD
,”
in
Consum
er
Ele
ct
roni
cs
-
As
ia
(I
CCE
-
As
ia
),
IEEE
Int
ern
ationa
l Confere
nc
e
on
.
I
EE
E
,
2016
.
[11]
Kaur,
T.,
J.
Gam
bhir,
and
S.
Kum
ar.
“
Arduin
o
based
solar
p
owered
batt
ery
charging
system
for
rur
al
S
HS
,”
in
Pow
er
E
lectr
o
nic
s (IICPE)
,
20
16
7th
Ind
ia Int
e
rna
ti
on
al
Conf
er
enc
e
on.
IE
EE
,
2
016
.
[12]
Le
,
T
.
-
H.
,
et
a
l.
“
An
ef
f
ic
i
ent
hand
det
ec
t
io
n
method
based
on
conv
olut
i
onal
neural
ne
twork
,”
in
201
8
7thInt
ern
at
ion
al
S
y
m
posium
on
Next
Gene
ration
Elec
tron
ic
s (IS
NE)
,
I
EE
E
,
201
8
.
[13]
Li
,
H
.
,
et
al.
“
Hands
det
ection
b
ased
on
stati
sti
c
al
le
arning
,”
in
Computati
onal
I
nte
lligen
ce
and
Design
(
ISCID
)
,
2012
Fifth
In
te
r
nati
onal
S
ymposium on
,
IEEE
,
2
012
.
[14]
Meng,
X.
,
J.
Lin,
and
Y.
Di
ng
.
“
An
e
xt
end
ed
HO
G
model:
SCHO
G
for
human
hand
de
te
c
tion
,”
in
S
y
st
ems
and
Inform
at
ic
s (ICS
AI),
2012
In
te
rn
at
ion
al
Conf
ere
n
ce
on
.
I
EEE
,
201
2
.
[15]
W
ang,
Y.
-
R.
,
W
.
-
H.
L
in,
and
L.
Yang.
“
An
imp
rove
d
hand
d
et
e
ct
ion
b
y
employing
corne
r
d
et
e
c
tor
,”
in
Mac
h
ine
Le
arn
ing
and
C
yber
netics
(ICML
C),
2014
Inte
rn
a
ti
onal Confe
r
ence
on
,
IE
EE
,
201
4
.
[16]
Mala
llah,
F.L
.
,
et
al
.
,
“
Vision
-
B
ase
d
Contro
l
B
y
Hand
-
Dire
ct
ion
al
Gestur
es
Con
ver
ti
ng
To
Voi
c
e
,”
In
te
rnationa
l
Journal
Of
S
cien
ti
fic
&
Technol
ogy
R
ese
arch
.
7(
7
).
[17]
Marin,
G.,
F.
Dom
ini
o,
and
P.
Zanutti
gh.
“
Hand
gesture
rec
ognitio
n
wit
h
le
ap
moti
on
and
ki
nec
t
de
vi
c
es
,”
in
Im
age
Proce
ss
ing
(ICIP),
2014
IEEE
In
t
ern
ational Conference on.
IE
EE
,
2014
.
[18]
Mala
llah,
F.L
.
,
et
al.,
“
Online
h
andwri
tten
signa
ture
rec
ogn
it
ion
b
y
le
ngth
norm
al
i
za
t
ion
using
up
-
sam
pli
ng
and
down
-
sam
pli
ng
,”
Inte
rnat
ional
J
ournal
of
C
ybe
r
-
Sec
urit
y
and
Dig
it
al Forensic
s (
I
JCSDF
),
4
(1):
p
.
302
-
313
,
2015
.
[19]
Mala
llah,
F.
L.,
et
al.,
“
Non
-
Inve
rti
bl
e
Onlin
e
Signat
ure
B
i
om
et
ric
T
empla
te
Protection
v
ia
Shuffling
an
d
Tri
gonom
et
r
y
Tr
ansform
at
ion
,”
In
te
rnational
Jour
nal
of
Computer
Appl
ic
a
ti
ons
,
98
(4)
,
2014.
[20]
Alkassar,
S.
,
e
t
a
l.,
“
Robust
scle
r
a
r
ec
ogni
ti
on
s
y
s
te
m
wi
th
nove
l
scl
era
segm
ent
ation
and
va
li
da
ti
o
n
te
chn
ique
s
,”
IEEE
Tr
ansacti
ons
on
Syste
ms
,
Man
,
and
Cybe
rne
ti
c
s: Sy
stems
,
47
(3)
,
p
.
474
-
486
,
20
17
.
[21]
Ron
y
,
M.R.
A
.
H.
and
M.M.
Al
am.
“
Hand
gesture
rec
ogni
ti
on
using
image
s
egme
ntation
an
d
dee
p
n
eural
net
work
,”
in
E
leventh
Int
ern
atio
nal
Confer
ence
on
Mac
hine
Visi
on
(ICMV
2018
)
,
Inte
rn
ational
Socie
t
y
for
Opt
i
cs
and
Photoni
cs
,
2
019
.
[22]
Yada
v,
K.
,
et
al.,
“
Hand
Gesture
Rec
ogni
ti
on
u
sing
Im
prove
d
Skin
and
W
rist
Dete
c
ti
on
Algor
it
hm
s
for
India
n
Sign
,”
Journal
o
f
Ne
twork
Comm
unic
ati
ons
and
E
merging
Technologies
(
JNCE
T)
www.
jn
ce
t
.
org
,
9(
2)
,
2019.
[23]
Ngu
y
en
,
H.
-
D.
and
S.
-
H.
Kim
,
“
Hand
Segme
ntat
ion
and
Fi
nge
rtip
Tr
ack
ing
from
Depth
Came
ra
Images
Us
in
g
Dee
p
Convo
lut
io
nal
Neural
Net
w
ork
and
Mul
ti
-
ta
sk
SegNet
,”
arXi
v
pre
prin
t arXiv:1901.03465,
201
9.
Evaluation Warning : The document was created with Spire.PDF for Python.