TELKOM
NIKA
, Vol. 11, No. 10, Octobe
r 2013, pp. 6
135 ~ 6
142
ISSN: 2302-4
046
6135
Re
cei
v
ed Ap
ril 1, 2013; Re
vised July 1
4
, 2013; Accept
ed Jul
y
24, 2
013
A Gait Recognition System using GA-based C-SVC and
Plantar Pressure
Yanbei Li, Lei Yan*, Hua Qian
Schoo
l of T
e
chnol
og
y, Beij
in
g F
o
restr
y
Un
ive
r
sit
y
Beiji
ng 1
0
0
083
, China, +
86-1
0
-62
338
27
9
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: mark_
y
an
lei
@
bjfu.e
du.cn, l
e
i
y
a
nbfu@
16
3.com
A
b
st
r
a
ct
In order to co
nduct the g
a
it recog
n
itio
n system,
a w
i
rele
ss in-sho
e w
earab
le pl
antar
pressur
e
acqu
isitio
n system b
a
sed o
n
ATmeg
a
1
6
an
d 8 FSR s
ens
ors w
a
s appli
e
d to data acq
u
i
s
ition for the g
a
its
w
h
ich co
nsist
of stand
in
g, w
a
lkin
g, j
u
mpi
n
g a
n
d
go
in
g
u
p
stairs. An
d fo
ur vo
lunt
eers (
2
fe
ma
les
an
d
2
m
a
les) were
invited
in this
researc
h
to collect
the
press
u
re infor
m
atio
n. MATLAB and LIBSVM wer
e
app
lie
d to co
n
duct al
l al
gor
ith
m
s pr
op
osed
b
y
this st
udy. G
enetic A
l
g
o
rith
m (GA) w
a
s us
ed to s
e
t the b
e
st
tunin
g
(pen
alty
) para
m
eter an
d the best (ga
m
ma) of G
aus
s radial b
a
sis k
e
rne
l
(RBF
) for C-support vec
t
or
classificati
on (
C
-SVC)
mod
e
l
and th
e GA-
base
d
C-SVC
w
a
s obtain
e
d
.
A dataset n
a
med
‘
t
r
a
in-
d
a
t
a
’
,
contai
nin
g
80
0
sets of pressure data w
a
s used to
train t
he GA-base
d
C-SVC as the
algor
ith
m
of gait
recog
n
itio
n. F
i
nally
a t
e
sting
dataset
c
onta
i
n
i
ng
40
0 sets
of pressur
e
dat
a
w
a
s app
lie
d to
test the a
l
gor
it
h
m
of gait reco
gnit
i
on ca
lle
d GA-base
d
C-SVC. T
he accura
cy
of this GA-based C-SVC w
a
s
98% for stand
i
n
g
,
91% for w
a
lki
n
g, 82% for go
in
g-upsta
irs an
d 97% for
ju
mpi
n
g. In gener
ally
speak
ing, a b
e
tter GA-based
C-
SVC w
a
s obtai
ned i
n
this rese
arch.
Ke
y
w
ords
:
C-
SVC; GA; Plantar pressu
re; Gait recognition
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
In rece
nt years, a growi
n
g
need for a f
u
ll r
ang
e of visual
surveill
a
n
ce a
nd mon
i
toring
system
s in secu
rity-sen
sitive
environme
n
t
s
su
ch a
s
banks, airp
o
r
t
s
and human i
dentificatio
n at
a dist
a
n
ce ha
s gain
ed increasi
ng intere
st from co
m
p
uter re
se
arch
ers. Gait reco
gnition is onl
y
a
recogni
zin
g
tech
nolo
g
y which can be detecte
d and
measu
r
ed at a dist
ance. It overcome
s the
chall
enge
s
whe
n
usin
g
physical o
r
close
cont
a
c
t su
ch as
finger pri
n
t recognitio
n
, face
recognitio
n
b
y
it
s inhere
n
t biological ch
ara
c
teri
stics
su
ch a
s
; non
-invasive, no
n-cont
a
c
t, hid
i
ng
and di
sgui
sin
g
dif
f
eren
ce, far-dist
an
ce re
cog
n
ition an
d
so on.
1.1. Pre
v
iou
s
Appr
oach
e
s
to Gait
Rec
ognition
Although gait
recognition i
s
a ne
w re
se
arch
field, there have
bee
n som
e
studi
es an
d
resea
r
che
s
.
Curre
n
tly, gai
t re
cognitio
n
approa
che
s
can
be
mainl
y
cla
ssifie
d
i
n
to two
cl
asses:
motion-ba
sed
methods a
n
d
model-ba
s
ed
methods.
Model
-ba
s
ed
method
s ai
m to mod
e
l h
u
man
body b
y
analyzin
g the pa
rts
of b
ody su
ch
as
hand
s, to
rso, thig
hs, le
gs, a
nd feet
and
perfo
rm model matchi
ng
in ea
ch
frame
of a wal
k
in
g
seq
uen
ce to
measure these para
m
eters.
As a typical
example of model
-ba
s
ed
approa
ch
e
s
o
f
gait reco
gni
tion, Cuna
do
et al. [1
]
con
s
id
ere
d
l
egs a
s
an i
n
terlin
ked
pe
ndulum,
and
gait
sig
natu
r
es
were
de
rived from t
h
e
freque
ncy co
mpone
nts of the variation
s
in the inc
lina
t
ion of huma
n
thighs. The
s
e features
were
analyzed u
s
i
ng the p
h
a
s
e
-
wei
ghted
Fo
urie
r Mag
n
itu
de Spe
c
tru
m
for gait
re
co
gnition. John
son
and Bobi
ck [2] used a
c
tivity-spe
cific sta
t
ic body par
a
m
eters for ga
it recog
n
ition without direct
ly
analyzi
ng th
e dynami
c
s
of gait p
a
tterns. Yam
et
al. [3] first
u
s
ed
ru
nnin
g
and
wal
k
ing
to
recogni
ze
pe
ople. They th
en explo
r
ed t
he rel
a
tion
shi
p
betwe
en
walkin
g and
ru
nning that
was
expre
s
sed
as a m
appin
g
b
a
se
d o
n
p
h
a
s
e
modul
atio
n. In ad
dition
, Cun
ado
et
al. [4] used t
h
igh
joint trajecto
ri
es a
s
feature
s
.
Re
cently, in literature [5], a simple b
u
t e
ffi
cient gait re
cog
n
ition alg
o
rithm u
s
ing
spatial
-
temporal sil
h
ouette an
alysis
wa
s
pro
p
o
se
d by
Wa
ng an
d Ta
n. This can b
e
con
s
id
ered
“not
rocket
scie
n
ce” as it may b
e
very easy to unde
rsta
nd
the techn
o
log
y
in it.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : 613
6 –
6142
6136
The advantages of model
-based ap
proaches are that they o
ffer the ability to derive gait
sign
ature
s
di
rectly from mo
del paramete
r
s.
As for the m
o
tion-ba
se
d a
p
p
roa
c
h
e
s of
gait re
co
gniti
on, Ji
we
n Lu
and E
r
hu
Zh
ang [2
1
]
prop
osed a
gait re
cog
n
ition metho
d
usin
g mult
ipl
e
gait features rep
r
e
s
ent
ation ba
se
d
on
Indepe
nde
nt Comp
one
nt Analysi
s
(I
CA)
and
geneti
c
f
u
zzy
suppo
rt vector ma
chi
ne
(GFSVM) for
the purp
o
se
of gait recogn
ition at a distance.
Hayfro
n-Acqua
h et al. [22] descri
bed an a
u
to
matic
gait recogniti
on method u
s
ing the gene
ralize
d
sy
mme
try operato
r
, and Phillip
s et al.
[23] applied
a baseline al
gorithm b
a
se
d on sp
atial-t
e
mpo
r
al si
l
h
o
uette co
rrel
a
tion to the ga
it identificatio
n
probl
em. The
s
e auth
o
rs
should b
a
si
call
y be on “t
he same wavel
e
ngth”
in
the u
nderstan
ding of
gait recogniti
on, as they are able to app
r
oach it from different points of view.
1.2. Plantar Pressur
e
Ac
quisition Sy
stem
Plantar
pre
ssure
a
c
qui
sitio
n
sy
stem i
s
o
ne of
th
e m
o
st imp
o
rtant
a
pplication te
chniqu
es
in gait re
cogn
ition, clinical foot-pai
n treat
ment
and foo
t
wear d
e
si
gn
fields an
d ha
s also be
com
e
a po
we
rful t
ool for biom
ech
ani
cal
re
search. T
h
is wide
appli
c
at
ion could
m
a
ke
a “lay man”
equate it to
the pre
s
e
n
t day ipad
s. Numerou
s
sy
stems have
b
een d
e
velop
ed for G
r
o
u
n
d
Rea
c
tion Fo
rce (GRF
) a
s
sessment an
d gait-ph
ase de
tection.
In-sh
o
re pla
n
tar pressu
re acq
u
isition
sy
stem, wh
ich is capa
b
l
e of simulta
neou
sly
measuri
ng G
R
F-i
ndu
ce
d p
l
antar force
s
and dete
c
ti
ng
gait-pha
se
s
of human, ha
s eme
r
ge
d as an
attractive alternative for
grou
nd mo
u
n
ted force pl
atform due t
o
several ou
tstandin
g
me
rits
inclu
d
ing
po
rtability, flexibility and g
r
eat
co
nvenie
n
ce
. Therefore, rese
arche
r
s i
n
this field
can
mak
e
more heads
or tail out of it.
As for the
nu
mber
of force
sen
s
o
r
s to
b
e
used, man
y
authors ha
ve prop
osed
different
numbe
rs to reco
gni
ze the
gait pha
se; fo
r insta
n
ce, [8] use
d
four to
identify the g
a
it pha
se
s wi
th
a cl
assification al
gorith
m
, and
ha
s
ob
tained
goo
d
results. An
d
in mid
-
20
00
s, acco
rding
t
o
literature [7], Faivre emp
l
oyed eight sen
s
o
r
s
in t
he in-sho
e plantar p
r
e
s
sure sy
stem
and
Flexiforce se
nso
r
(T
eksca
n
Inc., USA)
or FSR
se
nsor (Inte
r
lin
k Electro
n
ics, USA) have b
e
e
n
comm
only used.
Different
sen
s
ing
prin
cipl
e
s
have al
so
b
een
widely ex
plore
d
for in
-sho
e planta
r
pre
s
sure
system. S
p
rin
g
ele
m
ent
s
with strain
gau
ges we
re
co
mmonly u
s
e
d
to me
asure
vertical
re
acti
on
and
she
a
r fo
rce
s
. F-scan
(Tekscan In
c., USA)
utilize
s
force
sen
s
it
ive re
sisto
r
s.
A Novel p
e
d
a
r
system (Nov
el USA, Inc.) that captu
r
ed dy
nami
c
in-sh
oe te
mporal and
spatial p
r
e
s
sure
distrib
u
tion
s
were utilized
for dynami
c
g
a
it stabilit
y an
alysis [9], gait
re
cog
n
ition [
10], and alte
red
gait cha
r
a
c
teristics du
ring
runnin
g
[11]. Both sy
stem
s (F-scan an
d
Novel pe
dar)
use
d
ele
c
tri
c
a
l
wire
s to
con
n
e
ct in
-shoe
sensors
and
d
a
ta a
c
qui
sitio
n
sy
stem
aro
und th
e
wai
s
t, whi
c
h
ca
used
inco
nvenien
ce an
d di
scom
fort du
ring
strenuo
us ex
ercise
s. A wi
rel
e
ss st
ru
cture
shoe-i
n
teg
r
ate
d
sen
s
o
r
syste
m
was
develo
ped for gait
a
nalysi
s
and real-time fee
d
back [12].
2. Proposed
Algorithms
The stu
d
ie
s of gait recog
n
ition and its key tech
nol
ogie
s
have a
n
importa
nt a
c
ad
emic
signifi
can
c
e
a
nd p
r
a
c
tical
v
a
lue. So
in th
is p
ape
r, a
g
a
it re
cog
n
itio
n sy
stem
usi
ng GA
-ba
s
e
d
C-
SVC and pla
n
tar pressu
re
was int
r
odu
ced.
2.1. Genetic
Algorithm
Geneti
c
Algo
rithm (GA) [
6
], a ki
nd o
f
self-a
dapte
d
glo
bal o
p
timization
p
r
o
bability
sea
r
ching alg
o
rithm
by
si
mulating cre
a
t
ures’ gen
e a
nd evaluatio
n
in natural e
n
vironm
ent was
develop
ed a
nd investigat
ed by profe
s
sor
Holla
nd and his
stud
ents (e.g. De
Jon
g
) in Michigan
University in
197
5. Rece
ntly, Genetic Algorit
h
m
h
a
s bee
n su
cce
ssfully app
lied
to
va
rio
u
s
optimizatio
n
probl
em
s an
d it co
ntain
s
4 ge
neral
steps; initiali
zation, se
lection, recombi
n
ation
(cro
ssover) a
nd mutation.
The gen
eral algorith
m
of Geneti
c
Algorithm is sho
w
n
in Figure 1.
Step 1
G
e
n
e
rate
initial
p
opulatio
n ran
domly
with a
fixed nu
mb
er
of individ
u
a
ls
and
encode e
a
ch individual.
Step 2
Cal
c
ulate the fitn
ess of ea
ch
indivi
dual, a
nd then j
udg
e this p
opul
ation by
Optimizatio
n
Standard. If it is fit,
the ca
lculation is ove
r
. Or turn to step 3.
Step 3
Sele
ct the ‘good pa
rents’ via fitness.
Step 4
Ge
nerate the offspri
ng throu
gh th
e recombi
nati
on of the parents.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Gait Recog
n
ition System
using GA
-ba
s
ed
C-SVC a
nd Plantar Pressure (Yan
b
e
i Li)
6137
Step 5
Ge
nerate the offspri
ng throu
gh th
e mutation of the pare
n
ts.
Step 6
G
e
t a
ne
w p
opul
ation that
co
mp
rise
s
the
offspring
from
step4
and
ste
p
5 a
nd
return to s
t
ep 2.
Figure 1. The
general flow
cha
r
t of Gene
tic Algorithm
2.2. C-Suppo
rt Vector Cla
ssifica
tion
The Su
ppo
rt
Vector Cl
assi
fication
(SVC), a
bra
n
ch o
f
Suppo
rt Ve
ctor Ma
chin
e
(SVM)
whi
c
h is
ba
se
d on extrem
el
y well develo
ped ma
chi
ne
learni
ng the
o
ry named Statistical
Lea
rnin
g
Theo
ry (SLT), is a promi
n
ent cla
ssifie
r
that
has be
e
n
pro
p
o
s
ed b
y
Vapnik an
d
his co
-worke
rs
(Co
r
te an
d Vapni
k, 1995;
Vapnik, 19
95,
1998
) [19].
C-Su
ppo
rt Vector
Cla
ssifi
cation (C-SVC) [20
], a part of SVC, can
be ch
ara
c
te
ri
zed a
s
a
sup
e
rvised le
arnin
g
algo
rithm cap
able o
f
solving nonli
near
cla
ssifi
cation pro
b
lem
s
.
Gene
rally sp
eaki
ng, a
c
cording to the L
agra
nge fu
nction, Karu
sh–
K
uhn–T
ucke
r (KTT)
con
d
ition
s
, st
rong
du
ality theore
m
an
d
Slater
co
n
d
i
t
ions [2
0], given a trainin
g
data
s
et
with
instan
ce
-lab
e
l
pairs
1
{(
,
)
}
m
ii
i
xy
, where
n
i
xX
is an i
nput
vector
and
{1
,
1
}
i
y
, the
initial algorith
m
model of C-su
ppo
rt Vect
or Cla
s
sificati
on (C-SVC) is:
11
1
1
(,
)
2
mi
n
mm
m
ij
i
j
i
j
j
ij
j
yy
K
x
x
s
.t.
1
0
,
0
,
1
,
..
.,
m
ii
i
i
yC
i
m
(1)
In Eq. (1)
i
denote
s
La
g
r
ang
e multip
liers,
(,
)
ij
Kx
x
that can expan
d the linea
r
probl
em
s to
the non
-line
a
r
problem
s i
s
called
ke
rnel fun
c
tion,
and
C
rep
r
e
s
e
n
ts the tuni
ng
(pen
alty) parameter. Th
us the algorithm
of C-SVC is:
Step 1
Giv
en a trai
ni
ng data
s
et
with in
sta
n
ce
-lab
el pa
irs
1
{(
,
)
}
m
ii
i
xy
whe
r
e
n
i
xX
is an inp
u
t vector an
d
{1
,
1
}
i
y
.
Step 2
Chose a prop
er tu
ning pa
ram
e
ter
0
C
and a ke
rn
el function
(,
)
ij
Kx
x
.
Step 3
Obtain
**
*
1
(
,
...,
)
T
m
v
i
a Eq. (1).
Step 4
Sele
ct a vector of
*
from 0 to
C
, and then cal
c
ul
ate
**
()
j
ii
i
j
b
yy
Kx
x
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : 613
6 –
6142
6138
Step 5
Const
r
uct the d
e
ci
sion functio
n
()
s
g
n
(
()
)
f
xg
x
where
**
1
()
(
)
m
ii
i
i
g
xy
K
x
x
b
.
3. Materials
and Method
s
3.1. In-Shoe W
ear
able Plant
a
r Pres
su
re
Acq
u
isitio
n Sy
stem
In order to ef
fectively obt
ain the dif
f
erent
types of plant
ar p
r
e
s
sure such as st
andin
g
,
wal
k
ing, jump
ing and g
o
ing
up
st
airs a
s
the trainin
g
a
nd testing d
a
t
a
, this p
a
p
e
r
desi
gne
d an in-
sho
e
we
ara
b
l
e
plant
a
r
pressure mea
s
u
r
e
m
ent system.
Acco
rdi
ng to the literature
s
mentio
ned
above, this st
udy
cho
s
e F
S
R se
nsors provide
d
by Interlink El
ectro
n
ics in USA
as dat
a a
c
qui
sition
sen
s
ors.
Force Sen
s
in
g Re
sisto
r
s (FSR) a
r
e pol
ymeric thi
ck f
ilm (PTF) de
vices
which exhibit a
decrea
s
e i
n
resi
stan
ce
wi
th an incre
a
se in the
force applie
d to
the active su
rface. Its fo
rce
sensitivity is
optimized for
use
i
n
hu
man
touch
control
of ele
c
troni
c
device
s
. FSRs a
r
e n
o
t loa
d
cell
s or st
rain
gauge
s, thou
gh they have simila
r pro
p
e
r
ties [17].
As illustrated
in Figure 2,
the heel, metatarsals a
nd
hallux are th
e prima
r
y re
gion
s to
bear the
bod
y weig
ht [18]
. Therefore,
this p
ape
r was de
signe
d to
have
fo
ur
force sen
s
o
r
s
config
ure
d
at
the Heel, M
e
ta 2
nd
, Meta 1
st
, and
Hall
ux for e
a
ch foot to obtai
n
plantar force
and
detect gait p
hase.
Th
e sensors
we
re
pa
ckage
d i
n
two
p
a
irs
of insole
an
d the
in
sole
wa
s
adhe
red
to a
pair
of shoe
s, a
s
sho
w
n i
n
Figu
re
3 a
nd Fig
u
re
4.
A total of eig
h
t FSR
sen
s
ors
were appli
ed
in this experi
m
ent.
Figure 2. Plantar pressu
re
distributio
n [13]
Figure 3. The
packag
e
of
sensors in the sho
e
s
Figure 4. The
position of sensors in insole
This
re
se
arch u
s
ed
a
kin
d
of mi
cro
c
o
n
trolle
r na
me
d ATmeg
a16
that is
a lo
w-p
o
wer
CMOS 8
-
bit
microcontroll
er ba
se
d on t
he AVR e
nha
nc
e
d
RISC archite
c
ture as the
main chip
of
this data
acq
u
isition
syste
m
be
cau
s
e
o
f
its 10-
bit succe
ssive
ap
proximatio
n
analo
g
to dig
i
tal
conve
r
ter
(A
DC)
whi
c
h i
s
con
n
e
c
ted t
o
an
8-cha
n
nel Analog Multiplexer that
allows 8
single
-
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Gait Recog
n
ition System
using GA
-ba
s
ed
C-SVC a
nd Plantar Pressure (Yan
b
e
i Li)
6139
ende
d voltag
e inp
u
ts con
s
tructed
from t
he pi
ns
of
P
o
rt A, and the
singl
e-e
nde
d voltage input
s
refer to 0V (G
ND
).
Therefore thi
s
rese
arch
d
e
sig
ned th
e
data a
c
qui
siti
on prog
ram
usin
g C l
ang
uage
by
Cod
e
Visio
n
AVR
to
turn the
an
alog pl
antar pressu
re information
to digital
informatio
n by t
he
ADC in ATme
ga16.
As for the
dat
a uplo
ading, t
he wi
rele
ss serial m
odule
named X
L10
5-23
2 was
utilized to
uploa
d the da
ta to the com
puter in this
study.
In summ
ary,
the wh
ole
sy
stem b
oard
of dat
a a
c
q
u
i
s
ition
contai
n
ed: a mi
crocontrolle
r
named
ATme
ga16, a wirel
e
ss se
rial
m
o
dule nam
e
d
XL105
-23
2
,
a
linea
r com
p
ensation circuit
for s
e
ns
ors
with 10
k
re
sisto
r
s, and a volta
ge-stabili
zing
filter circuit.
The procedu
re for data a
c
quisitio
n
wa
s as follo
ws;
colle
cting a
n
a
log si
gnal v
i
a FSR
sen
s
o
r
s, then
obtaining dig
i
tal signal through AT
me
g
a16, and final
ly uploading t
he digital sig
nal
to the comput
er by wirele
ss serial m
odul
e.
3.2. Data
Ac
quisition
In ord
e
r to
ge
t more
universal a
nd
effici
ent data
of pl
antar
pre
s
su
re, this
re
sea
r
ch m
ade
use
of four
volunteers; t
w
o fe
m
a
le
s
and two mal
e
s to
co
ndu
ct the d
a
ta
acq
u
isitio
n of
this
experim
ent. The inform
atio
n of the volunteers i
s
a
s
sh
own in Ta
ble
1
.
Table 1. Gen
e
ral info
rmati
on of volunte
e
rs
Gende
r
Age
Height
Weight
Feet
size
Female
23
167 cm
55 kg
39 cm
Female
22
160 cm
48 kg
37 cm
Male
20
174 cm
60 kg
41 cm
Male
22
167 cm
65 kg
40 cm
Four
kind
s
of general g
a
its; standi
n
g
, wa
lki
ng, jumping a
n
d
going up
st
airs
we
re
desi
gne
d to be re
co
gni
ze
d by the algo
rithms
of gai
t recognitio
n
introdu
ce
d in
this pap
er
an
d it
wa
s de
cide
d to acqui
re 1
0
00 group
s of data for
ea
ch
kind of gait,
namely; 250
sets of
standi
ng-
data, 250
set
s
of wal
k
in
g-data, 250
set
s
of ju
mpi
ng-data and
250
sets of g
o
in
g-up
stai
rs-dat
a
were obtain
e
d
by each v
o
luntee
r. And once the
data of plant
ar pressu
re
wa
s se
nt to the
comp
uter, th
e data
wa
s saved to a te
xt file vi
a a control p
r
o
g
ra
m that wa
s d
e
sig
ned
by this
study throu
g
h
Micro
s
oft Vision C++ 6.0.
Therefore
ea
ch d
a
taset contain
s
one
thous
and
su
b data
s
ets f
o
r sta
ndin
g
, wal
k
ing,
jumping, an
d going
-up
s
tairs wa
s obtai
ne
d from the four voluntee
rs.
Then thi
s
research
sele
cte
d
200
data
ra
ndomly
fro
m
each
data
s
et respe
c
tively
to
ma
ke
a traini
ng
dat
aset
call
ed ‘train-d
ata’ a
n
d
100
data
f
r
o
m
ea
ch
data
s
et to
get the
testing
data
s
et
named ‘te
s
t-d
a
ta’.
3.3. Simulation of GA-Ba
se C-
SV
C for Gait Recognition
Whe
n
u
s
ing t
he C-SVC m
odel (Eq.
(1)), two pro
b
lem
s
were
co
nfro
nted; ho
w to cho
o
se
the optimal p
enalty param
eter
C
for C-SV
C, and ho
w to set the best kernel pa
rameters. To
impleme
n
t propo
sed
app
roach, this
re
sea
r
ch
u
s
e
d
the RBF
ke
rnel fu
nctio
n
for the
C-SVC
becau
se the
RBF kernel f
unctio
n
can
analyze
hig
h
e
r-dimen
s
io
n
a
l data a
nd
requires th
at
only
two paramet
ers,
C
and
(gam
ma) be defin
ed. Therefore the
Geneti
c
Algorithm
wa
s used to
optimize the
para
m
eters
C
and gamm
a
in this re
sea
r
ch.
MATLAB, LIBSVM [15] and
LIBSVM-Faruto
Ulti
m
a
te Version
[14]
were applied to
simulate the
algorith
m
of gait reco
gnitio
n
namely GA-base C-SV
C.
Acco
rdi
ng to
the de
sig
ned
pro
g
rams of
algor
ith
m
s, b
e
ca
use
of usi
ng
GA
for se
arching
the best
C
and best gamm
a
, this paper u
s
ed 10
0 as the maximum
number of g
eneration
s
, 20
as the
si
ze
of popul
ation
,
0.9 as th
e
rate of
in
di
viduals to
be
sele
cted, 0.
7 as the rate of
recombi
natio
n, and 0.025 as the
rate of
mutation. The sea
r
ch rang
e of penalty
para
m
eter
C
an
d
is from
1 t
o
10
0. And t
he
sele
ctio
n
method
s of t
he GA
are
Ran
k
-ba
s
ed
Fitness Assi
gnment a
nd
Stocha
stic
Universal Sam
p
ling. The fitness fun
c
tion
of this Gen
e
tic
Algorithm is t
he accu
ra
cy of 5-Fold
Cro
ss Valid
ation
of C-SVC.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : 613
6 –
6142
6140
K-Fold
Cro
s
s Validation
(K
CV) i
s
one
of
the mo
st p
o
p
u
lar
re
sam
p
li
ng te
chni
que
s, which
is effective, reliable an
d si
mple [16].
This resea
r
ch made a n
u
meral label
to repr
esent
all the types of gait, thus ‘1’ was
cho
s
e
n
to re
pre
s
ent ‘sta
n
d
ing’; ‘2’ rep
r
ese
n
ted ‘wal
king’; ‘-1’ for ‘jumping’ an
d ‘-2’ for ‘goi
ng-
upstai
r
s’.
So gen
erally
sp
ea
king, i
n
order to
est
abli
s
h a GA-ba
s
e
d
C-SVC syste
m
of
gait
recognitio
n
preci
s
ely, the following m
a
in
step
s (a
s sho
w
n in Figu
re
5) mu
st be prece
ded;
Figure 5. System architect
u
re
s of
GA-b
ase
d
C-SV
C of gait reco
gn
ition
Firstly o
b
tain
the d
a
ta of
plantar p
r
e
s
sure
and
cho
s
e th
e ‘trai
n
-data’ a
nd ‘te
s
t-data’.
Secon
d
ly get
the o
p
timal
para
m
eters
o
f
C-SV
C by
t
he G
eneti
c
Al
gorithm
an
d t
r
ain it
to g
e
t ‘
t
he
bes
t GA-bas
ed C-SVC ’. Finally s
t
art th
e
gait reco
gniti
on throu
gh th
e ‘test-data’.
4. Experimental Re
sults
The ‘train
-dat
a’ wa
s an
800
8
dataset made o
f
800 sub d
a
taset
s
whi
c
h
contai
ned 2
0
0
data fro
m
‘
s
ta
nding’,
200
d
a
ta from
‘wal
king’,
200
dat
a from
‘jumpi
ng’ a
nd
200
d
a
ta from
‘goi
n
g
-
upstai
r
s’; a
n
d
‘test-data’
was a
400
8
data
s
et that com
p
ri
se
d 100 d
a
ta from ‘stan
d
ing’
, 100 data
from ‘wal
kin
g
’
, 100 data from ‘jumpin
g’
and 100 d
a
ta from ‘goin
g
-
up
stairs’ after selectin
g the
experim
ental data.
The b
e
st
wa
s 1
3
.645
2 a
n
d
the
be
st ga
mma
wa
s 0.
0029
564,
after utili
zing
th
e GA fo
r
optimizin
g as
sho
w
n in Fig
u
re 6.
The final
cl
a
ssifi
cation
m
odel
ba
sed
o
n
GA-ba
s
ed
C-SVC u
s
ing
the b
e
st
parameters
and ‘fin
al trai
n-data’
was d
one
by the
al
gorithm
de
sig
ned
by this re
sea
r
ch th
rou
g
h
MATLAB
a
n
d
LIBSVM-Faruto Ultimate Vers
ion. An
d t
he
res
u
lt of final gait re
c
ognition based
on the final
GA-
based C-SVC was a
s
sho
w
n in Table 2.
From th
e ex
perim
ental
re
sult of final
gait re
co
gniti
on sho
w
n i
n
Table
2, it can
be
con
c
lu
ded
th
at this
re
sea
r
ch
obtai
ns
a bette
r GA
-base
C-SVC for g
a
it recognition,
and
the
accuracy of
the classifier is more than 90%
except for ‘going-up
stairs’
becau
se of
its
compli
catio
n
among oth
e
r
three gait.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Gait Recog
n
ition System
using GA
-ba
s
ed
C-SVC a
nd Plantar Pressure (Yan
b
e
i Li)
6141
Figure 6. The
result of pa
ra
mete
rs optimi
z
ation of C-S
V
C by GA
Table 2. The
result of final gait recogniti
on
Gaits True
nu
mber
of gaits
Classified number
of gaits
Accuracy
o
f
the t
r
ained
GA-based C-SV
C
Standing 100
98
98%
Walking
100
91
91%
Jumping 100
97
97%
Going-upstairs
100
82
82%
5. Discussio
n
s and Co
nc
lusions
Usi
ng the Ge
netic
Algo
rith
m for p
a
ram
e
ters
optimization of SVM is a gre
a
t method of
this study
. It
may not be an easy con
c
e
p
t to underst
a
nd as one will
need to “bea
t brains out”,
but
one can ma
ke head
s or t
a
i
l
s out of it for future stu
d
ies.
Although o
u
r re
cog
n
ition
accuracy i
s
comp
a
r
atively
high a
nd e
n
coura
g
ing, it
still does
not give much con
c
lu
sio
n
about gait
s
.
In ord
e
r to
g
e
t a better cl
assificatio
n
o
f
‘going
-u
p
s
t
a
irs’
o
r
mo
re
com
p
licate
gait, an
improve
d
GA
or othe
r alg
o
rithms for
p
a
ra
meter
optimi
z
ation su
ch
as p
a
rti
c
le
swa
r
m optimizatio
n
(PSO) o
r
more experim
ent
al dat
a shoul
d be appli
ed i
n
the SVC in future stu
d
ies.
Of course oth
e
r ke
rnel fun
c
tion
s of SVM such as Sigmoid kern
el, S
p
line kern
el
, Fourier
kernel al
so can be cho
s
en
to get the classifi
cation mo
del if further rese
arche
s
wil
l
go on.
In addition,
-su
ppo
rt vector cla
s
sifica
tion (
-SVC)
can
be trie
d
as the i
n
itial
cla
ssifi
cation
model in th
e future works an
d the p
a
ramete
r
co
uld be optimized by othe
r
optimizatio
n algorith
m
.
Beside
s SVM
, there are m
any other ki
n
d
s of algo
rith
ms whi
c
h
ca
n be applie
d to gait
recognitio
n
.
Examples in
clud
e; BP n
eural
net
wo
rk, the
algo
ri
thm of L
o
ca
lity Prese
r
ving
Projections (LPP), the algorithm of S
t
atistica
l Shape Analysi
s
(SSA), and t
he algorithm
of
combi
n
ing
static and dyna
mic biom
etrics. The bottom
line in this re
sea
r
ch is that
somethin
g was
reali
z
ed from the con
c
e
p
t used.
Finally, Princi
pal Com
pon
e
n
t Analysis (PCA) or
Fa
ct
or Analysi
s
(FA) can b
e
a
pplied to
pre
-
processin
g
the datas
ets in furthe
r st
udie
s
.
Ackn
o
w
l
e
dg
ement
This research
would like to ackno
w
le
dge
the
four volun
t
eers
who a
r
e
studying in Beijing
Fore
stry
University an
d ou
r thanks al
so
go to t
he a
n
o
n
ymous
refe
rees fo
r thei
r thoro
ugh
revi
ew
and con
s
tru
c
tive comme
nts.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : 613
6 –
6142
6142
Referen
ces
[1]
D Cu
nad
o, MS
Ni
xo
n, JN C
a
r
t
er.
Using
ga
it as a
bio
m
etric, via
p
has
e-w
e
i
ghted mag
n
itu
de
sp
ectra
. in:
Proc.1
st
Int. Conf. Audio- a
nd
Vide
o-bas
ed Bi
omet
ric Perso
n
Authenticati
o
n
.
1997, pp. 95-
102.
[2]
A Johnso
n
, A Bobick.
A mu
lti
-
view
metho
d
for gait reco
gnit
i
on usi
ng static
body par
a
m
et
ers
. in: Proc.
3
rd
Int. Conf. Audi
o- and V
i
de
o-bas
ed Bi
ome
t
ric
Person Aut
hentic
atio
n.
20
01; pp. 30
1-31
1.
[3]
C Yam, M
Ni
xon, J
Carter.
On
th
e re
la
ti
on
sh
ip
o
f
h
u
man
wa
l
k
in
g a
nd ru
nn
i
n
g
:
au
tom
a
ti
c pe
rson
identification by gait
. in: Proc. Int. Conf. Pattern Rec
ogn
itio
n. 2002; I: 287-
290.
[4]
D Cun
ado, M
S
Nixon, JN
Carter. Autom
a
tic e
x
traction
and descr
ipti
on of huma
n
gait mod
e
ls fo
r
recog
n
itio
n pur
poses.
Co
mput
/Vis. Image U
n
derstan
din
g
. 2
003; 90(
1): 1-4
1
.
[5]
Lia
ng W
a
ng, T
i
eni
u T
an, Hua
z
hon
g Ni
ng, W
e
imin
g
H
u
. Sil
hou
ette Ana
l
ys
is-Base
d
Gait
Reco
gniti
o
n
for Human Identification.
IEEE Transactio
n
s
on Pattern
Analys
is an
d Machi
ne Intel
l
i
genc
e
. 200
3;
25(1
2
).
[6]
Xi
ao
pin
g
W
a
n
g
, Limin
g
Cao
.
Genetic Algo
rithm-theori
e
s,
appl
icatio
n a
nd soft
w
a
re i
m
pleme
n
tatio
n
.
Xi
’a
n Jiao T
ong Univ
ersit
y
Pr
ess; 2002.
[7]
A F
a
ivre, M D
aha
n, B Parra
tte, G Monnier
, In
strumented
shoes
for p
a
thol
ogic
a
l
gait
assessme
n
t
.
Mechanics Research Comm
unications
. 20
04
; 31: 627-6
32.
[8]
Y Okaza, N Yoshi
da, K Oguchi. Using s
o
le
pressure si
gn
als to ana
l
y
ze
w
a
lki
ng p
o
stur
e.
ICRA200
6
.
200
6: 197
–2
00
.
[9]
ED Lem
aire, A
Bis
w
as, J Ko f
i
na
n.
Pla
n
tar p
r
essure
para
m
eters for dyn
a
m
ic
gait sta
b
il
ity ana
lysis
. in:
Proc. 28th Ann
u
. Int. Conf. IEEE Eng. Med. Bi
ol. Soc. (EMBS), Ne
w
Y
o
rk
, 2006: 44
65-4
468.
[10]
P Catalfamo,
D Moser, S G
houss
a
yni, D
E
w
i
n
s. Detecti
on of gai
t ev
e
n
ts using a
n
F
-
Scan in-s
hoe
pressur
e
meas
ureme
n
t s
y
ste
m
.
Gait Posture
. 2008; 2
8
(3): 420-
426.
[11]
PW
Kong, HD
Heer. W
eari
n
g
the F
-
Scan
mobil
e
in-s
ho
e
pressure m
e
a
s
ureme
n
t s
y
st
em alters g
a
it
character
i
stics duri
ng run
n
i
ng.
Gait Posture
. 200
9; 29(1): 14
3-14
5.
[12]
SJ Morris, JA
Paradis
o
.
Shoe-i
n
teg rated
sensor system for w
i
reless
gait analys
is and re
al-ti
m
e
feedb
ack.
in: Proc. 24th Annu. Conf. Annu. Fall M
eet. Biomed. Eng.
Soc. EMBS/BMES Conf.,
Housto
n, 200
2, 2468
–2
469.
[13]
Clin
ical
and R
e
search So
lutio
n
s, http://
w
w
w
.
tekscan.com.
[14]
F
a
ruto, yan
g
Li,
LIBSVM-fa
r
utoUltimat
eVe
r
si
on
a to
ol
b
o
x
w
i
t
h
im
ple
m
ents for S
u
pport V
e
ctor
Machi
nes bas
e
d
on li
bsvm, 20
11, http://
w
w
w
.
matlabsk
y
.com
.
[15]
Chihc
h
ung Chang, Chihje
n Lin, LIBSVM: a libr
a
ry
for Su
ppor
t Vector Machines, 2001,
http://
w
w
w
.
csi
e
.
ntu.edu.t
w
/~
cj
l
i
n/libsvm.
[16]
Cross–V
a
li
dati
on (Ana
l
y
sis S
e
rvices - Data
Mi
nin
g
), in: Microsoft SQL Server 20
08
Books Onli
ne
,
Microsoft Corp
oratio
n, 200
8.
[17]
F
S
R®
Integration Guid
e & Ev
alu
a
tion P
a
rts Catal
og
w
i
t
h
Sugg
ested El
ectrical Interfaces
, http://
w
w
w
.
Interlink
e
lectronics.com.
[18]
G Yuan, M
X
Z
han
g, Z
Q
W
a
n
g
, JH Z
h
a
ng. T
he
distri
b
u
tion
of foot pr
essur
e
a
nd
its infl
ue
nce factors
in
Chin
ese p
e
o
p
l
e
.
Chin
a J. Phys. Med. Rehab
i
l
. 2004; 2
6
(3): 156-
159.
[19]
Suju
n H
ua, Z
h
iro
ng S
un.
S
upp
ort Vector
Mach
ine
a
p
p
r
oach
for
prot
ein
subc
el
lul
a
r
loc
a
li
z
a
t
i
o
n
pred
iction
. 2
0
0
1
; 17(8): 72
1-7
28.
[20]
Nai
y
a
ng D
e
n
g
,
Yingji
e
T
i
an, Supp
ort Vector
Machi
nes-t
heor
ies, al
gorit
hm
s and c
onti
nuati
on. in:
Beiji
ng Sci
enc
e Press; 200
9.
[21]
Ji
w
e
n Lu, Erhu Zhang. Gait
recognition for human
id
entification based on I
C
A and fuzz
y
SVM through
multiple vie
w
s f
u
sion.
Pattern Recognition Letters.
2007; 28:
2401-
24
11.
[22]
JB Ha
yfro
n-Ac
qua
h, MS N
i
x
o
n
, JN C
a
rter.
A
u
tomatic gait
r
e
cog
n
itio
n by
s
y
mmetry an
aly
s
is
. in: Proc.
3
rd
Int. Conf. Audi
o- and V
i
de
o-bas
ed Bi
ome
t
ri
c Person Aut
hentic
atio
n. 20
01; 272-
27
7.
[23]
PJ Phi
lli
ps, S
Sarkar, I R
obl
edo, P
Grother
, K Bo
w
y
e
r
.
B
a
seli
ne
res
u
lts for the
ch
all
e
n
ge
prob
le
m o
f
hu
ma
n
ID usi
ng gait an
aly
s
is
. in: Proc.
Int. Conf.
A
u
tomatic F
a
c
e
an
d Gesture
Reco
gn
ition,
W
a
shin
gton, D
C
, 2002; 1
37-1
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.