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.
8
,
No.
6
,
D
ece
m
ber
201
8
, pp.
50
1
4
~
50
2
0
IS
S
N:
20
88
-
8708
,
DOI:
10
.11
591/
ijece
.
v
8
i
6
.
pp
50
1
4
-
50
2
0
5014
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Compari
son Anal
ysis of G
ait Class
ification
f
or Hum
an Moti
on
Identific
atio
n
U
si
ng Embe
dd
ed Comp
uter
Agun
g Nu
groho J
at
i,
A
s
tri
Noviant
y, N
anda
Sep
tiana,
Le
ni Widi
a N
as
u
tion
Depa
rtment
o
f
C
om
pute
r
Engi
n
e
eri
ng,
School
of Electrical E
ng
in
ee
ring
,
Te
lkom
Univer
sit
y
,
Indo
nesia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ja
n
26
, 2
01
8
Re
vised
Jun
2
8
, 201
8
Accepte
d
J
ul
22
, 2
01
8
In
thi
s
pape
r,
it
will
be
discussed
about
compa
rison
bet
wee
n
t
wo
kind
s
of
cl
assifi
ca
t
ion
m
et
hods
in
ord
er
to
improve
sec
u
rity
s
y
s
te
m
base
d
of
hum
an
gai
t
.
Gai
t
is
on
e
of
biometr
ic
m
e
thods
which
ca
n
be
used
to
ide
nt
if
y
per
son
.
K
-
Nea
rest
Neig
hbour
has
par
a
ll
ell
y
implemen
te
d
with
Supp
ort
Vec
to
r
Mac
hine
for
c
la
s
s
if
y
ing
hum
an
g
ai
t
in
sam
e
basi
c
s
y
stem.
Gen
eral
l
y
,
s
y
s
te
m
has
bee
n
buil
t
u
sing
Histogram
and
Princi
pa
l
Com
ponent
Anal
y
sis
for
gai
t
det
e
ct
ion
and
it
s
fea
tur
e
ex
tra
c
ti
o
n.
The
n
,
th
e
resu
lt
of
th
e
sim
ula
tion
show
ed
tha
t
K
-
Nea
r
est
Neighbour
is
slower
in
proc
ess
ing
and
le
ss
accura
t
e
tha
n
Support
Vec
tor
Mac
hine i
n
gait
cl
assifi
ca
t
ion.
Ke
yw
or
d:
Gait
Reco
gn
it
ion
Hu
m
an
Moti
on
Identific
at
io
n
K
-
Near
e
st Nei
ghbour
Suppor
t
V
ect
or Mac
hin
e
Copyright
©
201
8
Instit
ute of
Ad
v
ance
d
Engi
ne
eri
ng
and
Sc
ie
n
ce
.
Al
l
rights
reserv
ed
.
Corres
pond
in
g
Aut
h
or
:
Agu
ng Nu
groho
Jat
i
,
Dep
a
rtm
ent o
f C
om
pu
te
r
E
ng
i
neer
i
ng,
School
of Elec
tric
al
Engineer
ing
,
Tel
kom
Un
ive
rsity
,
Jl. Tele
kom
un
ikasi Te
ru
sa
n
B
uah
Ba
tu,
Ban
dung, 4
0257
,
I
ndonesi
a
.
Em
a
il
:
agu
ngnj
@
te
lko
m
un
ive
rsity
.ac.id
1.
INTROD
U
CTION
Ever
y
sin
gle
pe
rson
in
the
w
or
l
d
has
uniq
ue
featur
es
wh
ic
h
is
diff
e
ren
t
f
ro
m
oth
ers.
I
n
eng
i
neer
i
ng,
it
can
be
us
e
d
as
a
key
f
or
identific
at
io
n
or
in
sec
ur
it
y
s
yst
e
m
.
Gen
e
ra
ll
y,
it
’s
cal
le
d
as
bi
om
et
ric
syst
e
m
.
Ma
ny
i
m
ple
m
entat
ion
of
biom
et
ric
syst
e
m
us
e
pa
rt
of
hu
m
an
body,
but
it
’s
sti
ll
ver
y
rar
e
bio
m
et
ric
syst
e
m
us
e
hu
m
an
m
otion.
Mo
st
of
them
,
m
ai
nly
us
ed
face
r
ecognit
ion
s
uc
h
as
t
o
def
i
ne
ge
nd
e
r
of
hum
an
identific
at
ion
[
1].
So
in
this
pa
per,
it
will
be
disscuse
d
a
bout
m
et
ho
ds
f
or
i
den
ti
fyi
n
g
hu
m
an
featur
e
ba
sed
on
hu
m
an
m
otion
especial
ly
u
sin
g hu
m
an
gait.
Gait
is
def
ined
as
the
way
an
ind
ivid
ual
org
anism
walk,
in
cl
ud
in
g
hum
an.
And
eve
ry
hum
an
has
a
un
i
qu
e
gait
as
a
beh
a
vio
ral
char
act
e
risti
c
wh
ic
h
is
diff
e
r
ent
from
each
oth
er
.
It
de
pe
nd
s
on
w
ei
ght
,
le
g’
s
le
ng
th
a
nd
siz
e,
an
d
postu
re
of
body
[
2
]
.
St
ud
yi
ng
gait
co
m
bin
es
m
or
e
than
a
s
ubj
e
ct
,
li
ke
m
edical
st
ud
ie
s
,
ps
yc
holo
gy, bi
ology, a
nd m
ot
ion
a
naly
sis [
3
]
. I
t st
r
ongly p
r
ov
e
s that
gait c
an be
us
e
d
in
bi
om
e
tric
s syst
em
.
To
rec
ognize
a
nd
cl
assify
the
hu
m
an
gait,
there
a
re
tw
o
ki
nd
of
a
ppr
oac
h.
Fir
st
one,
it
’s
base
d
on
m
ot
ion
a
naly
sis
w
hich
nee
ds
at
te
ntion
of
w
ho
le
hum
an
body
m
ov
em
ent.
To
ease
t
hese
m
et
ho
ds,
we
c
an
us
e
captu
red im
age
o
f
hum
an
sil
houttes.
Sec
ond, it
’s
base
d o
n
f
eat
ur
e a
ppr
oac
h wh
ic
h
on
ly
ne
eds
to
p
ay
att
ention
in
the
sp
esi
fic
par
t
of
hum
an
body,
es
pec
ia
ll
y
the
m
ov
ing
par
t
li
ke
knees
an
d
hing
es
[
4]
,
[5
]
.
In
oth
er
researc
h, it
’s
s
ai
d
that t
her
e t
wo app
ro
ac
h o
f gait
an
al
ysi
s, spati
al
f
eat
ure
and tem
po
ral f
eat
ur
e a
naly
sis [
6
].
Com
par
ing
t
he
or
i
gin
al
hu
m
an
body
m
od
el
with
sim
ulatio
n
m
od
el
of
ga
it
featur
e
ca
n
be
us
e
d
t
o
know
so
m
eone’s
gait
cha
rac
te
risti
c.
Si
m
ul
at
ion
m
od
el
con
ta
in
s
a
gait
cy
cl
e
wh
ic
h
is
a
per
io
dic
co
nd
it
io
n
from
a
fo
ot
ste
pp
i
ng
in
a
gro
und
unti
l
ano
t
her
t
urns.
T
he
cy
cl
e
is
di
vid
e
d
int
o
tw
o
par
t
s
of
c
onditi
on,
w
hic
h
are
cal
le
d
sta
nc
e
and
s
wing.
A
sta
nce
is
des
cribe
d
w
hen
only
a
fo
ot
ste
ppin
g
in
a
gro
und,
an
d
swi
ng
def
i
nes
wh
e
n
a
fo
ot swi
ng
in
g
i
n
th
e ai
r.
Ge
ner
al
ly
, a c
yc
le
co
ns
ist
s
of 60% sta
nce
and 40%
sw
i
ng
[
7
].
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
& C
om
p
Eng
IS
S
N:
20
88
-
8708
Compari
son A
na
ly
sis
of
Ga
it
Cl
as
sif
ic
ation
f
or
Hum
an M
otion
I
den
ti
fi
cation
…
(
Ag
ung
Nug
r
oho Jati
)
5015
It
need
s
s
om
e
ste
ps
base
d
on
i
m
age
pr
oces
sing
f
or
identi
fyi
ng
hu
m
an
ga
it
.
First,
m
oti
on
m
us
t
be
est
i
m
at
ed
in
a
lot
of
nu
m
ber
in
a
n
area
an
d
m
us
t
be
li
m
i
te
d
by
hum
an
anatom
ic
con
st
raint
in
hiera
rc
hical
m
od
el
.
It’s
nee
ded
t
o
extr
act
the
fix
ed
hu
m
an
bo
dy
par
am
et
ers.
I
n
th
e
next
ste
p,
the
f
orm
of
hum
an
gait
will
be
analy
zed
by
us
in
g
ed
ge
sha
rpness
to
repr
esentat
e
le
gs
area.
He
uri
sti
c
m
od
el
is
us
ed
to
est
i
m
at
e
the
per
i
od
and
will
be
co
m
par
ed
w
it
h
uniq
ue
f
or
m
of
gait
cy
cl
e
as
sh
o
w
n
in
Fi
gur
e
1
[8
]
.
A
nd
th
e
la
st,
that
un
i
qu
e
f
or
m
will
b
e
recog
nized by cl
assifi
cat
ion
m
et
ho
ds. T
o
ease,
w
e
c
an use
HOG a
nd PC
A for e
xtr
act
ing
gait feat
ur
e
.
Figure
1.
Peri
odic
gait cy
cl
e
[
8
]
Histo
gr
am
of
Gr
a
dient
(
HOG)
is
a
n
al
go
rithm
wh
ic
h
is
use
d
t
o
detect
th
e
sp
esi
fic
obj
e
ct
in
a
n
sta
ti
c
i
m
age.
HOG
util
iz
es
colour
changin
g
of
each
pi
xel
with
oth
e
rs
in
th
e
dig
it
al
i
m
ag
e
and
well
-
kn
own
as
gr
a
dient
[
9
]
.
Wh
il
e
P
rincipa
l
Co
m
po
ne
nt
An
al
ysi
s
(P
C
A
)
is
us
e
d
to
ide
ntify
the
patte
r
n
an
d
e
xpress
t
hem
to
the
ot
her
form
in
or
der
t
o
s
how
t
he
diff
e
re
nces
a
nd
the
si
m
il
arities.
PCA
is
oft
en
us
e
d
as
featu
re
ex
tract
io
n
m
et
ho
d
[
10
].
I
nd
ee
d,
PC
A
ha
s
eve
r
been
use
d
t
o
e
xtract
t
he
featur
e
s
f
r
om
chest
X
-
ray
i
m
age
an
d
prov
i
ded
m
or
e than
95%
accu
racy [
11
]
.
In
this
pa
per,
will
be
fo
cu
se
d
only
in
cl
ass
ific
at
ion
m
e
tho
ds.
The
cl
assi
ficat
ion
m
et
ho
ds
ha
ve
bee
n
com
par
ed
in
this
pa
per
a
re
K
-
near
e
st
N
ei
ghbour
(
KNN)
a
nd
Sup
port
V
ect
or
M
achine
(SVM)
.
This
com
par
ison
is
need
e
d
f
or
our
fu
t
ur
e
wor
k
in
im
ple
m
e
ntati
on
of
Hum
an
Id
e
ntific
at
ion
Syst
em
b
ased
on
e
m
bed
de
d
c
om
pu
te
r.
The
s
yst
e
m
accuracy
in
the
m
ai
n
pro
blem
of
gai
t
recogn
it
io
n.
Be
sides,
it
’s
ne
eded
to
know
ho
w
m
u
ch
res
ources
use
d
f
or
c
om
puta
ti
on
.
S
VM
is
chosen
beca
us
e
the
c
on
ce
pt
of
m
ulti
cl
as
s
one
againts
on
e
c
an
dec
rease
t
he
cl
assifi
cat
ion
e
rro
r
[
2
]
.
On
t
he
ot
her
hand,
K
NN
is
a
si
m
ple
m
eth
od
i
n
cl
assifi
cat
ion
wh
ic
h
us
e
s
th
e
con
ce
pt
of
le
arn
i
ng
by
an
al
og
y
an
d
on
l
y
cl
assifi
es
based
on
m
os
t
m
at
ched
featur
e
s [1
2
].
This
pa
pe
r
will
b
e
di
vid
e
d
int
o
f
our
sect
io
ns
wh
ic
h
in
the
first,
disc
us
se
d
a
bout
the
re
vie
w
of
hum
an
gait,
the
m
ai
n
f
ocu
s
of
t
he
discuss
i
on,
a
nd
s
om
e
of
relat
ed
w
orks
.
In
the
sec
ond
s
ect
ion
,
sho
wn
of
t
he
te
chn
iq
ues
use
d
i
n
the
re
sear
ch.
The
re
su
lt
will
be
prov
i
ded
in
the
sec
ti
on
fou
r.
A
nd
in
t
he
la
st,
w
il
l
be
pr
ese
nted
the
c
on
cl
us
io
n
a
nd t
he fut
ur
e
w
ork of t
he researc
h.
2.
RESEA
R
CH MET
HO
D
In
ge
ne
ral,
gait
ide
ntific
at
ion
process
can
be
descr
i
bed
as
f
ol
lows
.
Ge
ne
ral
Pr
oc
ess
Desig
n
as
s
how
n
Figure
2.
Im
age
aqc
uisit
ion
i
s
the
fir
st
ste
p
to
get
dig
it
al
im
age.
I
n
this
s
te
p,
cam
era
will
captur
e
vid
e
o
an
d
div
ide
s
into
f
r
a
m
es.
Fr
am
es
will
be
entere
d
in
to
ne
xt
ste
p,
cal
le
d
prep
r
ocessin
g.
In
pre
processin
g,
i
m
ages
will
be
proces
sed
to
inc
reas
e
i
m
age
qu
al
it
y,
no
ise
re
duc
ti
on
,
tra
ns
f
orm
the
i
m
age
into
ot
her
f
orm
at,
an
d
determ
ine
the
par
t
will
be
ob
s
er
ved.
The
pr
oc
esses
are
sta
rt
ro
m
gr
a
ysc
al
ing
,
thre
s
ho
l
ding,
bac
kgr
ound
su
bst
racti
on, cl
os
in
g, an
d
se
gm
entat
ion
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
87
08
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
201
8
:
50
1
4
-
50
2
0
5016
Figure
2. Ge
ne
ral Process
D
e
sign
Feat
ur
e
ext
rac
ti
on
is
pro
pos
ed
to
get
th
e
i
m
po
rtant
i
nfo
rm
ation
s
from
the
im
age
to
de
fine
t
he
diff
e
re
nces
f
orm
oth
ers.
T
o
e
xtract
them
,
Pr
incipal
Com
po
nen
t
A
naly
sis
(
PCA)
is
us
e
d
i
n
this
resea
rch.
PCA,
wh
ic
h
fir
st
wa
s
introd
uce
d
by
Pearson
(19
01)
a
nd
H
otell
in
g
(
1933)
,
is
de
sign
e
d
to
c
olle
ct
the
uniq
ue
f
eat
ur
es
from
the
i
m
age
[1
3
]
.
The
un
iqu
e
feat
ur
e
s
can
be
fou
nd
by
transfor
m
ing
the
i
m
age
fr
om
hig
h
dim
ension
al
data
into
lo
w
on
e
.
It’s
ob
ta
i
ned
from
a
set
trai
nin
g
data
of
the
im
age
[1
4
].
T
o
sim
pli
fy
the
pro
blem
,
each
acqu
i
red im
age
w
as
rezise
d
in
to
80
x80 pixel
i
m
age.
PCA
ste
ps
will
b
e
pro
vid
e
d
in
the
T
a
ble
1
bel
ow.
Table
1.
PCA
Pr
oc
esses
f
or
F
eat
ur
e E
xtracti
on
No
Step
An
n
o
tatio
n
1
Matr
ic
es
Ad
ju
st
m
e
n
t
To trans
f
o
r
m
the t
r
ain
in
g
datas
et into
m
a
tri
ces (64
0
0
x
2
4
0
),
then
calculate
t
h
e
m
ean
(64
0
0
x
1
)
an
d
g
et adju
sted
data (64
0
0
x
2
4
0
).
240
,
6400
2
,
6400
1
,
6400
2
,
2
1
,
2
240
,
1
2
,
1
1
,
1
...
...
...
...
...
...
...
...
...
...
...
...
u
u
u
u
u
u
u
u
u
2
Co
v
ariance M
at
ric
es
Calcu
latio
n
Calcu
latio
n
of
Co
v
ariance bas
ed
on
giv
en
f
o
r
m
u
las :
Mean
2
4
0
1
1
,
1
k
k
u
m
u
,
wh
e
re
u
is av
era
g
e vecto
r,
and
m
is n
u
m
b
e
r
o
f
m
a
tri
ces.
Da
ta
Adju
st
= (
u
-
u
)
6
4
0
0
2
1
u
u
u
u
Co
v
ariance
m
at
rix
Co
v = da
ta
No
rma
l
T
*
d
a
ta
No
rm
a
l
An
d
r
esu
lt is cov
ar
ian
ce
m
at
rix with
(
2
4
0
x
2
4
0
)
sized
.
3
E
ig
en
v
ecto
r
an
d
E
ig
en
v
alu
es
Calcu
latio
n
E
ig
en
vector
(64
0
0
x
2
4
0
)
an
d
eig
en
va
lu
es
(1x
2
4
0
)
f
ro
m
co
v
ariance
m
atric
es.
4
Princip
le Co
m
p
o
n
en
t
Extractio
n
Eigen
vector
is so
rted b
ased
on
eig
en
v
a
lu
es
f
ro
m
hig
h
est to
lowes
t
.
Fo
r
clas
sif
icatio
n
is
u
sed
24
0
PC (
Princip
al Co
m
p
o
n
en
t)
Re
su
lt
data
from
featur
e
e
xt
racti
on
was
use
d
t
o
cl
assify
an
d
rec
ogniz
e
gait
a
nd
it
s
owne
r.
A
s
exp
la
ine
d
befo
re,
m
ai
n
fo
cus
in
this
researc
h
is
the
com
par
iso
n
betwe
en
SV
M
an
d
K
N
N
f
or
cl
assifi
c
at
ion
and ide
ntific
at
ion m
et
ho
d.
H
oweve
r, t
hey s
ha
red sam
e d
at
a an
d
te
ste
d by t
he
sam
e scenari
os
.
2.1.
Sup
po
r
t Vect
or M
achi
ne (
S
V
M) f
or Gait
C
l
as
sif
ficat
i
on
SV
M
will
def
i
ne
the
di
visor
betwee
n
tw
o
ki
nd
of
cl
asses
in
the
i
nput
s
pa
ce
know
n
as
hype
rlane.
Hype
rlane
ca
n
be
fi
nd
out
by
cal
culat
e
it
s
m
arg
in
a
nd
fin
d
m
axi
m
u
m
va
lue.
T
her
e
is
s
upport
vector
def
i
ne
d
a
s
cl
os
est
data
fo
r
eac
h
cl
ass
.
It
can
be
us
e
d
ke
rn
el
a
pproach
to
get
the
m
.
Ker
nel
is
def
ine
d
as
f
un
ct
ion
f
or
m
app
in
g data f
eat
ur
e
from
the origin
d
im
ension into
o
t
her fe
at
ur
e
with
high
est
d
im
ension
[
1
5
].
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
& C
om
p
Eng
IS
S
N:
20
88
-
8708
Compari
son A
na
ly
sis
of
Ga
it
Cl
as
sif
ic
ation
f
or
Hum
an M
otion
I
den
ti
fi
cation
…
(
Ag
ung
Nug
r
oho Jati
)
5017
Ther
e
are
t
wo
kind
of
S
VM
f
un
ct
io
ns
i
n
cl
a
ssific
at
io
n.
Th
e
first
is
nee
de
d
to
def
i
ne
the
cl
asses
by
us
in
g
trai
ning p
r
ocess
.
It’
s p
urp
os
e
is
t
o
fi
nd
w
(w
ei
gh
t
),
b
(
bias
),
a
nd
s
upport
m
achine
f
ro
m
each
cl
ass
.
Th
e
n
in
the
te
sti
ng
proces
s,
pa
ram
e
te
rs
w
hich
have
been
obta
ine
d
f
ro
m
trai
ning
pro
cess
incl
ude
in
the
cal
c
ul
at
ion
as in
pu
t t
o
m
ake d
eci
sio
n func
ti
on
.
(1)
Sign
val
ue
fro
m
pr
ocesses
a
bove
is
the
re
su
lt
f
ro
m
te
sted
data
[
1
6
].
R
efering
t
o
the
trai
ning
proces
s,
it
s
i
m
ple
m
entat
io
n
us
es
Ga
us
si
an
kernel
a
nd
Sygm
a
value
6000.
I
n
t
he
previ
ou
s
res
earch
,
it
s
hows
t
he
m
axi
m
u
m
acc
ur
acy
with
opti
m
u
m
co
m
pu
ta
ti
on
tim
e.
As
we
know
that
ke
rn
el
ty
pe
us
ed
is
ver
y
i
m
po
rtant
for
al
l of
syst
em
p
erfor
m
ance.
2.2. K
-
Ne
arest Neig
hb
ou
r
(
KNN) f
or
G
ai
t
Cl
as
sific
at
i
on
K
-
Near
e
st
Nei
ghbour
(KN
N)
is
well
-
known
super
vise
d
le
a
rn
i
ng
al
gorith
m
,
wh
ic
h
cl
ass
ifie
s
obj
e
ct
s
base
d
on
the
m
ajo
rity
cl
ass
of
feat
ur
es
f
r
om
k
a
m
ou
nt
of
cl
os
est
neig
hbou
r.
K
NN
is
base
d
on
le
ar
ni
ng
by
analo
gy
co
nce
pt,
w
he
re
le
ar
ning
data
a
re
descr
i
bed
by
n
dim
ension
al
nu
m
erical
at
t
rib
utes.
Eac
h
of
t
hem
represe
nts
a
point
i
n
n
dim
ensio
nal
s
pace
.
E
uclidean
distance
form
ula
is
use
d
to
ca
lc
ulate
the
dis
ta
nce
betwee
n qu
e
ry
data an
d
le
a
rn
i
ng d
at
a
[
1
2]
,
[
17
]
,
[
18
].
n
i
i
i
e
u
c
q
p
Q
P
D
1
2
)
(
,
(2)
KNN
cl
assifi
c
at
ion
accu
racy
dep
e
nds
on
s
i
m
i
la
rity
m
eas
ur
em
ent
us
age
and
value
of
k.
Be
f
ore
i
m
ple
m
entat
io
n,
ha
ve
been
t
est
ed
s
om
e
vari
nace
of
k
i
n
orde
r
to
get
the
best
acc
uracy
.
They
a
re
1,
3,
5,
7.
And for
im
plem
entat
ion
, use
d k=1 beca
us
e
this val
ue has
b
est
acc
ur
acy
.
3.
RESU
LT
S
A
ND AN
ALYSIS
In
this
sect
io
n,
it
is
exp
la
ined
the
resu
lt
s
of
te
sti
ng
sce
nar
i
os.
The
par
am
eter
s
w
hich
is
m
easur
e
d
a
nd
analy
zed a
re sy
stem
accur
at
ion, c
om
pu
ta
ti
on
ti
m
e, light d
e
ped
e
ncy
for
ac
cur
at
io
n, an
d d
ist
ance of cam
era.
3.1. Sys
tem
Acc
ura
cy
Ba
sed
on
prev
iou
s
sect
ion,
t
her
e
are
t
wo
ty
pe
of
cl
assifi
cat
ion
m
et
ho
ds.
T
hey
are
Suppo
rt
Vect
or
Ma
chine
(SV
M)
an
d
K
-
Nea
rest
N
ei
ghbo
ur
(
KNN).
In
S
V
M,
it
has
be
en
chosen
so
m
e
va
riables
value
,
sigm
a
6000
with
Gaussi
an
kernel.
T
hen,
in
K
N
N
m
et
ho
d
t
her
e
’s
giv
e
n
k=
1
f
or
cl
assifi
cat
ion
.
It’s
base
d
on
the
no
n
real
tim
e
te
stin
g
f
or
each
m
et
hod
in
or
der
to
ta
ke
the
be
st
perf
or
m
ance
f
or
e
ach
cl
assifi
cat
ion
m
et
hod.
Com
par
ison o
f
KNN an
d SV
M Acc
ur
at
io
n
as sho
wn in Fi
gure
3.
Figure
3. Com
par
is
on of
K
N
N
a
nd S
VM Ac
cur
at
io
n
In
t
he
fig
ur
e
a
bove,
sho
wn
that
cl
assifi
cat
ion
us
i
ng
S
VM
was
m
or
e
acc
ur
at
e
t
han
K
N
N.
T
he
te
sti
ng
was
done
by
usi
ng
118
data,
and
us
e
d
6
cl
a
sses.
S
VM
be
c
a
m
e
m
or
e
accu
rate
beca
us
e
it
cl
assifi
ed
by
find
i
ng
best
hype
rlane
to
div
i
de
data
a
nd
base
d
on
“
m
ul
ti
cl
ass
on
e
againts
one”
c
on
ce
pt.
Wh
il
e
KNN
only
us
e
d
k=
1
wh
ic
h
m
eans it o
nly com
par
es
w
it
h
le
ss
am
ou
nt
of
neig
hbour
hood
data.
*
)
(
b
x
x
a
y
x
f
i
i
i
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
87
08
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
201
8
:
50
1
4
-
50
2
0
5018
3.2. C
omp
u
t
ati
on
Ti
me
Com
pu
ta
ti
on
ti
m
e
analy
sis
was
perf
or
m
ed
in
order
to
know
the
ave
rage
of
processi
ng
ti
m
e.
Be
sides
,
it
’s
need
e
d
to
com
par
e
between
S
VM
an
d
K
NN
wh
at
’
s
faster
in
pr
oc
ess.
This
m
e
asur
em
ent
us
e
d
sam
e
conditi
on
with
pr
evi
ous
te
st
wh
ic
h
was
use
d
sigm
a
60
00
and
Gau
s
sia
n
kernel
for
SVM
,
and
k=
1
f
or
KNN.
The res
ult i
s shown
i
n
Fi
gure
4
belo
w.
Figure
4. Com
par
is
on of
K
N
N
a
nd S
VM C
om
pu
ta
ti
on
Ti
m
e
SV
M
giv
es
fa
ste
r
res
ult
tha
n
K
NN
beca
us
e
in
S
VM
on
ly
process
es
vec
tor,
w
hile
K
N
N
c
om
par
es
so
m
e
neighb
our
data
t
o
fi
nd
sim
il
arities.
SV
M
needs
a
ver
a
ge
17,
9
s
to
com
pu
te
t
hem
,
and
K
N
N
nee
ds
aver
a
ge 2
2,51
s.
3.
3.
Li
gh
t
D
e
p
endenc
y
In
orde
r
to
a
na
ly
ze
the
li
gh
t
eff
ect
in
proc
essing,
this
te
s
t
was
pro
pose
d.
T
he
te
st
us
e
s
six
cl
asses
com
e
fr
om
si
x
pe
rs
on
gait.
Each
cl
ass
use
s
fou
rthy
tr
ai
nn
in
g
data
and
ei
gh
t
te
st
data.
T
he
res
ul
t
on
ly
represe
nt
the
r
eal
tim
e
te
sti
n
g
,
m
eans
com
e
f
ro
m
ei
gh
t
t
est
data.
Cl
assifi
cat
ion
m
et
ho
d
par
am
et
ers
sti
ll
use
sam
e
with
oth
e
r
te
st.
For
S
V
M
sti
ll
us
e
Ga
us
sia
n
Ke
r
nel
with
6000
si
gm
a,
wh
il
e
K
N
N
us
es
k=
1.
A
ccur
at
io
n
in
Day
Ligh
t
Me
asur
em
ent
as
sh
ow
n
in
F
igure
5.
Acc
ur
at
io
n
in
Ni
ght
Ligh
t
Me
as
ur
em
ent
as
sh
own
in
Figure
6.
Figure
5. Acc
urat
ion i
n Day L
igh
t M
eas
ur
em
ent
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
& C
om
p
Eng
IS
S
N:
20
88
-
8708
Compari
son A
na
ly
sis
of
Ga
it
Cl
as
sif
ic
ation
f
or
Hum
an M
otion
I
den
ti
fi
cation
…
(
Ag
ung
Nug
r
oho Jati
)
5019
Figure
6. Acc
urat
ion i
n Nig
ht
Ligh
t M
eas
ur
e
m
ent
Ba
sed
on
t
he
te
st,
it
’s
sho
wn
that
ge
ner
al
ly
higher
l
um
inance
will
decr
eas
e
the
accu
racy.
The
high
e
r
li
gh
t
intensit
y
will
m
ake
the
data
ha
ve
m
or
e
no
ise
th
an
no
rm
al
con
diti
on.
Be
si
de
s,
la
m
p
will
pro
du
c
e
sh
a
dows. It m
akes i
n
the
n
i
gh
t
, accuracy
is
de
crease.
3.4.
E
ff
ec
t of
Ca
m
era
Posi
tion
s
In
this
te
st,
kn
own
t
hat
po
sit
i
on
an
d
distac
e
of
cam
era
fro
m
the
obj
ect
in
flue
nced
the
a
ccur
acy
.
It’
s
cause
d
by
patte
rn
of
capt
ur
e
d
fr
am
es.
Cl
os
er
obj
ect
f
ro
m
the
ca
m
era
c
auses
capt
ur
e
d
i
m
age
con
ta
in
le
ss
backg
rou
nd
but
not
fu
ll
obje
ct
in
the
f
ra
m
e.
W
hile
fur
ther
distance
m
akes
the
obje
ct
in
the
f
ra
m
e
siz
ed
sm
a
ll
er.
So
,
it
needs
op
ti
m
u
m
distance
. Res
ul
t can b
e
desc
ri
bed in t
he gra
phic
b
el
ow.
Figure
5. Acc
urat
ion i
n Day L
igh
t M
eas
ur
em
ent
4.
CONCL
US
I
O
N
Ba
sed
on
the
i
m
ple
m
entat
ion
,
te
sti
ng,
a
nd
a
naly
sis
can
c
onstr
ucted
so
m
e
co
nclusi
on
s
of
the
stu
dy.
Ov
e
rall
,
S
VM
is
bette
r
tha
n
KNN
for
cl
ass
ify
ing
hum
an
m
ot
ion
by
usi
ng
gait
of
the
m
.
It
can
be
s
how
n
i
n
accuracy,
c
ompu
ta
ti
on
tim
e,
and
s
om
e
eff
ect
of
cha
ng
e
d
c
onditi
on
p
aram
et
ers.
Alto
ugh,
hu
m
an
identific
at
ion
by
us
i
ng
gait
has
t
oo
m
any
con
st
raints
to
be
co
ns
ide
red.
Act
ually
in
im
ple
m
entat
io
n,
it
need
s
m
or
e
par
am
et
ers
to
m
ake b
et
te
r
ac
cur
acy
.
In
the
f
ur
t
her
sta
ge
of
stu
dy,
the
resu
lt
fro
m
this
stud
y
c
an
be
co
ns
ide
r
ed
to
be
im
pl
e
m
ented
in
sing
le
boar
d
co
m
pu
te
r,
su
ch
a
s r
asp
berrr
y pi,
p
cd
uino board
, b
eagle
bone,
e
tc
as a s
ecur
it
y
m
on
it
or
in
g
syst
e
m
.
It
can
be
integ
r
at
ed
to
our
m
ov
em
ent
pr
edict
ion
syst
em
wh
ic
h
has
bette
r
r
esult
in
accura
cy
[1
9].
Alth
ough,
it
needs m
or
e m
e
asur
em
ent, es
pe
ci
al
ly
in
resou
rce all
ocati
on a
nd m
anag
em
ent.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
87
08
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
201
8
:
50
1
4
-
50
2
0
5020
REFERE
NCE
S
[1]
M
.
J
.
Bara
ni
,
K
.
Faez
,
F
.
Jalili
.
“
Im
ple
m
ent
at
ion
of
Gabo
r
Filt
ers
Com
bine
d
with
Bina
r
y
Fe
at
ure
s
for
Gend
er
Rec
ognition
”
.
Inte
rnational
Jo
urnal
of
El
ec
tri
cal
and
Compu
te
r
Engi
nee
ring
(
IJE
CE)
.
Vol
4
(1)
:
108
-
115
.
Februa
r
y
2014
.
[2]
Darm
a
Putra. Biom
et
ric
s S
y
s
te
m
.
Yog
y
ak
art
a
:
A
ndi
Publisher
.
20
09.
[3]
Anil
Jain,
Ruud
Boll
e,
Shara
t
h
Pankat
i.
“
Biom
et
ric
s
Personal
Ide
nti
fi
cation
in
Networke
d
Socie
t
y
”
.
Kluwer
Aca
demic
Publishers.
[4]
David
Pos
t.
G
ai
t
Anal
y
sis Re
v
iew
.
Dow
nloa
ded
from
ww
w.nd
.
edu/
~dpos
t/
IntS
y
st
/re
port1
.
pdf.
20
06.
[5]
Chira
s
Ben
Abd
el
kad
er,
Ross
G.
Cutl
er,
d
an
Larr
y
S.
Dav
is.
“
Gait
Re
cogni
t
io
n
using
Im
age
Self
Sim
il
ia
ri
t
y
”
.
Hindawi
Publish
ing
Corpora
ti
on.
2004.
[6]
Zha
ng
Zha
ox
ia
n
g,
Maodi
Hu,
Yunhong
W
ang.
“
A
survey
of
Ad
vanc
es
in
Biom
et
ri
c
Gait
Re
cog
nit
ion
”
.
Journal
.
China
:
Be
iha
ng
Univer
sit
y
.
[7]
Spence
r,
N
.
M.
“
Pos
e
Inva
ria
n
t
Gait
Ana
l
y
s
is
and
Rec
onst
ruc
ti
on
”
.
Fa
cul
t
y
of
Engi
ne
ering,
Scie
n
ce
an
d
Mathe
m
at
i
cs,
Sc
hool
of El
ec
tron
i
cs
and
Com
put
er
Scie
n
ce,
Univ
er
sit
y
of
South
amton.
2005
.
[8]
Nash,
J.M.,
Ca
r
t
er,
J
.
N.,
and
M.
S.,
Nixon
.
“
D
y
n
amic
fe
at
ur
e
ext
r
ac
t
ion
v
i
a
the
ve
loc
ity
tr
houg
h
tr
ansform
”
.
Pat
te
r
n
Rec
ognition
Le
t
t
ers
.
vol
18
pp
10
35
-
1047.
1997.
[9]
Fathurra
hm
an,
I
rwan.
“
Im
ple
m
ent
at
ion
of
Te
x
t
Dete
c
ti
on
and
Rec
ognition
b
y
Us
ing
Histogram
of
Orien
te
d
i
n
Android
Plat
for
m
”
.
Final
Proje
ct
.
B
andung:
Sc
hool
of
Elec
t
rical
and
Info
rm
atics,
Insti
tut
T
ek
nologi
Bandung
.
2012.
(unpubl
ished)
[10]
Dian
Esti
Prat
iwi.
“
Im
ple
m
ent
ati
on
of
Face
Re
co
gnit
ion
b
y
using
(Princi
pal
Com
ponent
Anal
y
sis
)
”
.
Fina
l
Project.
Yog
y
ak
art
a
:
El
e
ct
ri
ca
l
and
Instr
u
m
ent
at
ion
Program
,
Univer
si
ta
s
Gadja
h
Mad
a. 2
013.
(unpubl
ishe
d)
[11]
Roopa
H,
As
ha
T
.
“
Feat
ur
e
Ex
tr
ac
t
ion
of
Chest
X
-
ra
y
Im
age
s
an
d
Anal
y
s
is
Us
ing
PC
A
and
kPC
A
”
.
Int
ernati
ona
l
Journal
of
Elec
t
rical
and
Computer
Eng
ine
ering
(
IJE
CE)
.
Vol
8
No 5.
Oc
tobe
r
2
018.
[12]
Wa
h
y
u
Hid
a
y
a
t.
“
K
-
Nea
rest
Neig
hbor
Im
ple
m
entati
on
for Landsc
ape
Im
ag
e
bas
ed
on
Colour
Fea
tu
re
and
T
ext
ur
e
”
.
Final
Proj
ec
t
.
B
a
ndung:
Insti
tut
T
eknol
ogi
Telko
m
.
2007.
(unpub
li
shed)
[13]
J.S
.
Butz
er
,
A
.
P
.
H
.
Butl
er,
P.J
.
Bones.
“
Medipix
Im
agi
ng
Eva
l
uat
ion
of
Da
ta
se
ts
with
PC
A
”
.
Image
and
Vi
sio
n
Computing
New
Zealand
IE
EE In
te
rnational
Conf
ere
nce P
roc
ee
di
ngs
.
2008;
23
:
1
-
6.
[14]
J.A
.
Le
e
,
M.
Verl
e
y
sen.
“
Uns
uper
vised
Dime
nsional
i
t
y
R
ed
uct
ion:
Overvi
e
w
and
Rec
ent
Advanc
es
”
.
IEEE
Inte
rnational
Jo
i
nt
Conf
ere
nce o
n
Neural
N
et
wor
ks
.
2010;
1
-
8.
[15]
Ching
-
Chih
Sai
,
You
-
Zhu
Che
n,
Ching
-
W
en
Li
ao
.
“
Inte
ra
ct
i
ve
Emotion
Re
cogni
ti
on
using
Support
Vec
tor
Mac
hine
for
Hum
an
-
Robot
Inte
rac
t
ion
”
.
IE
EE
I
nte
rnational
Co
nfe
renc
e
on
Syste
m,
man
and
C
ybe
rnet
ic
s
.
2009
;
407
-
412.
[16]
Yang
-
Guang
Li
u
,
Qi
Chen
,
Rui
-
Z
hao
Yu.
“
Ext
ra
ct
Candi
da
te
s
of
Support
Vec
tor
Tr
ai
ning
Set
”
.
IEEE
Inte
rnat
iona
l
Confe
renc
e
on
Mac
hine Learni
ng
and
Cyb
erne
t
ic
s
.
2003
;
5
:
319
9
-
3202.
[17]
J.L
.
Henr
y
.
A
.
“
K
-
Nea
rest
Neigh
bour
Method
for
Mana
ging
the
E
volut
ion
of
a
Learni
ng
Base
”
.
IE
EE
Inte
rnationa
l
Confe
renc
e
on
C
omputati
onal
Int
el
li
g
ence
and
M
ult
imedi
a
App
li
c
ati
ons
.
2001
;
4:
357
-
361.
[18]
Zhongj
ia
n
Song,
Qing
W
u,
Chun
m
ing
Xia.
“
Patte
rn
rec
ogn
it
ion
o
f
Finger
-
Motion
s
base
d
on
Diffu
s
ioa
n
Maps
and
Fuzz
y
K
-
Nea
r
est
Neighbour
C
la
s
sifie
r
”
.
I
EE
E
Int
ernati
onal
Confer
enc
e
on
Signa
l Proce
ss
ing
.
201
2;
2:
1207
-
1212.
[19]
AN
Jati
,
L
.
Novam
iz
anti,
M
.
B
.
Praset
y
o
,
A
.
R
.
Putra.
“
Ev
al
ua
ti
o
n
of
Moving
Object
Detect
ion
Methods
base
d
on
Gene
ral
Purpos
e
Single
Board
C
om
pute
r
”
.
Indon
esian
Journal
of
Elec
tri
cal
Enginee
ring
and
Co
mputer
Scienc
e
(
IJE
ECS)
.
Vol
4
(1)
:
123
-
129
.
A
pril
2015
.
Evaluation Warning : The document was created with Spire.PDF for Python.