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.
10
,
No.
3
,
June
2020
,
pp. 2
97
8
~
298
5
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v10
i
3
.
pp2978
-
29
85
2978
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
Human
ga
it
re
cognition u
sing pre
processi
ng
and cl
assific
ation techn
iqu
es
Sa
mer
K
ais Jameel,
Jih
ad
An
w
ar
Q
ad
ir
,
Moh
ammed
Hussein
Ah
m
ed
Depa
r
tment
o
f
C
om
pute
r
Scie
n
ce,
Univer
s
i
t
y
of
R
apa
rin
,
Ir
aq
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
J
ul
6
,
2019
Re
vised
N
ov 26
,
2019
Accepte
d
Dec
10, 201
9
Biom
et
ric
re
cog
nit
ion
s
y
stems
have
b
ee
n
attra
ct
ed
num
ero
us
rese
arc
h
ers
since
they
attempt
to
over
come
the
proble
m
s
and
fac
tors
wea
ke
ning
the
se
s
y
stems
includi
ng
proble
m
s
o
f
obtaining
images
ind
ee
d
no
t
app
ea
ring
the
resolut
ion
or
the
obje
c
t
completel
y
.
In
thi
s
work,
the
obje
c
t
m
ovement
rel
i
anc
e
w
as
con
sidere
d
to
d
isti
n
guish
the
hum
an
through
his/he
r
gai
t
.
Som
e
losing
fea
tur
es
proba
bl
y
wea
ke
n
the
s
y
stem’s
ca
pability
in
r
ec
ogni
zi
n
g
the
people,
hen
c
e,
we
propose
using
al
l
data
recorded
b
y
th
e
Kinec
t
sensor
wi
th
no
emplo
y
ing
the
feature
ext
ra
ct
ion
m
et
h
ods
base
d
on
th
e
li
teratu
r
e.
In
the
se
studie
s
,
coor
dina
t
es
of
20
point
s
are
re
cor
ded
for
e
ac
h
per
son
in
var
ious
gend
ers
and
age
s,
wal
king
with
v
ari
o
us
dire
c
ti
ons
a
nd
spee
ds
,
cre
a
ti
ng
8404
c
onstrai
nts.
More
over
,
p
re
-
proc
essing
m
et
hods
are
utilized
t
o
m
ea
sure
it
s
inf
l
uenc
es
on
the
s
y
stem
eff
iciency
through
t
esti
ng
on
six
t
y
pes
of
cl
assifie
rs
.
W
it
hin
the
propos
ed
appr
oac
h
,
a
note
worth
y
rec
o
gnit
ion
ra
te
was obt
ai
n
ed
r
eachi
ng
91%
with
out
ex
amining th
e
desc
r
ipt
ors
.
Ke
yw
or
d
s
:
Cl
assifi
cat
ion
Gai
t
rec
ogniti
on
Kinect se
nsor
Pr
e
processin
g
Copyright
©
202
0
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
:
Sam
er K
ai
s Ja
m
eel
,
Dep
a
rtm
ent
of
Com
pu
te
r
Scie
nce,
Un
i
ver
s
it
y o
f
Ra
par
in
,
Ra
nia, Ira
q
.
Em
a
il
:
sa
m
er.k
ai
s@uor.ed
u.k
rd
1.
INTROD
U
CTION
The
autom
at
ed
identify
ing
sy
stem
has
i
m
pr
ov
e
d
ba
rely
in
recent
deca
des
,
in
par
ti
cular
in
excessi
ve
safety
areas
e
.
g.
Air
ports
a
nd
banks
Bi
ome
tric
authe
ntica
ti
on
m
akes
use
of
orga
nic
or
be
ha
vio
ral
t
rait
s
to
ver
ify
th
e
i
den
t
ific
at
ion
of
a
pe
rson
[
1,
2].
I
n
the
m
idst
of
th
e
vio
la
ti
ons
tha
t
occur
co
ns
ta
nt
ly
in
m
os
t
par
ts
of
the
globe
a
m
ulti
plied
at
te
ntion
ha
s
bee
n
ta
ken
to
the
prev
e
ntio
n
of
te
rror
ist
at
ta
ck
s,
thr
oughout
bu
il
d
so
phist
ic
at
ed
a
nd
swift
syst
e
m
s
to
identify
the
hum
ans.
Ma
ny
bio
m
et
ric
te
chnolo
gies
ha
ve
em
erg
ed
f
or
identify
in
g
an
d
ve
rifyi
ng
pe
rsons
th
r
ough
analy
zi
ng
face
,
fin
gerp
rint,
pa
l
m
pr
int,
iris,
gait,
or
a
m
ix
ture
o
f
these trai
ts [
3
-
6].
Bi
om
e
tric
s
is
the
autom
at
ic
us
e
of
physi
ologica
l
or
be
hav
i
or
al
trai
ts
to
determ
ine
or
c
onfirm
the
identific
at
ion
of
a
pe
rson
,
the
ph
ysi
olog
ic
al
bio
m
et
rics
exam
ine
s
ph
y
siolo
gical
chara
ct
erist
ic
s
li
ke,
iris,
faces,
fin
gerpr
i
nts,
DNA,
a
nd
hand
geo
m
et
ry;
the
beh
a
vior
al
bio
m
et
rics
e
xam
ines
beh
a
vi
or
al
issues
,
s
uc
h
as
vo
ic
e,
sig
nature,
an
d
gait
[
1]
.
The
sig
nific
ance
of
c
om
pu
te
rized
i
den
ti
ficat
ion
of
pe
op
le
has
acce
l
erated
thr
oughout
the
previ
ou
s
deca
des,
pa
rtic
ularl
y
in
exce
ssive
secu
rity
ar
eas
su
c
h
as
ai
r
por
ts
an
d
banks
[
5,
7].
Each
c
ha
racter
has
disti
ng
uish
able
uniq
ue
t
rait
s
(
bio
m
et
ri
cs)
that
can
be
us
e
d
by
id
e
ntific
at
ion
syst
e
m
s
to
ver
ify
a
nd
pic
k
ou
t
the
pers
on
'
s
ide
ntit
y
[
8].
T
o
disti
nguish
uniq
ue
hum
ans
in
the
m
ann
er
they
s
troll
is
a
h
orri
ble
pro
je
ct
hu
m
ans
perform
ed
each
da
y.
Psycho
l
og
i
cal
stud
ie
s
[
4
,
8,
9]
ha
ve
s
hown
th
at
gait
sign
at
ur
es
bought
f
ro
m
vid
eo
ca
n
be
use
d
as
a
reli
able
cue
to
bec
om
e
awar
e
of
i
nd
i
viduals.
T
he
se
find
i
ngs
stim
ulate
d
researc
hers
in
com
pu
te
r
im
ag
inati
ve
an
d
presci
ent
to
ext
r
act
po
te
ntial
gait
sign
at
ures
f
ro
m
pictures
t
o
pic
k
ou
t
pe
ople
.
It
is
ch
al
le
ng
i
ng,
howe
ver,
to
lo
cat
e
idiosyncra
ti
c
gait
aspects
in
m
ark
er
-
le
ss
m
otion
seq
ue
nces,
wh
e
re
the
us
e
of
m
ark
ers
is
pr
e
ve
nted
due
to
the
fact
it
i
s
intru
sive
a
nd
no
t
su
it
able
in
com
m
on
place
gait
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
Hum
an gait
re
cogniti
on usi
ng
pr
e
pr
oce
ssin
g and cl
as
sif
ic
ation t
ech
niqu
es
(
Samer
K
ais
Jameel
)
2979
cognizanc
e
set
ti
ng
s.
Id
eal
ly
,
the
f
oc
us
feat
ures
extracte
d
from
ph
oto
s
be
i
nv
a
riant
to
el
e
m
ents
diff
e
ren
t
tha
n
gait,
s
uch
as
colo
r,
te
xture,
or
kind
of
cl
oth
in
g
[
9].
T
wo
Com
par
e
d
to
dif
fer
e
nt
bi
om
e
tric
m
e
tho
ds,
gait
consci
ousn
e
ss
offer
s
se
ver
al
un
i
qu
e
c
har
act
erist
ic
s.
The
m
os
t
app
eal
i
ng
c
har
act
erist
ic
is
it
s
un
ob
t
r
us
i
ve
ness,
wh
ic
h
do
es
no
longe
r
re
quire
obser
ve
d
s
ubje
ct
s’
at
te
ntio
n
and
co
operati
on.
In
ad
diti
on,
hum
an
gait
can
be
captu
red
at
s
om
e
distance
be
sides
requiri
ng
physi
cal
inf
orm
at
ion
f
ro
m
su
bject
s.
T
his
f
avora
ble
cha
ra
ct
erist
ic
has
br
il
li
ant
adv
a
ntage
s,
m
a
inly
wh
e
n
m
a
n
or
wo
m
an
f
act
s
su
ch
as
f
ace
i
m
age
are
confide
ntial
[
5,
8].
More
ov
e
r,
gait
recog
niti
on
presents
super
b
achieva
ble
f
or
co
ns
ci
ou
s
nes
s
of
lo
w
-
res
olu
ti
on
vi
deo
s
,
wh
e
r
e
oth
e
r
bi
om
et
ri
cs
te
chnolo
gie
s
m
ay
add
it
iona
ll
y
be
inv
al
id b
ecau
se
of
ins
uffici
e
nt
pix
el
s
to
identify
the
hu
m
an
top
ic
s
[9
]
.
T
he
com
m
on
f
ra
m
ewo
r
k
of
co
m
pu
te
rized
gai
t
fo
c
us
c
onsist
s
of
c
halle
nge
detect
io
n,
sil
houette
extracti
on,
f
unct
ion
e
xtracti
on,
f
un
ct
io
n
sel
ect
ion
,
an
d
cl
a
ssific
at
ion
.
O
nc
e
tran
sfe
rr
in
g
to
pics
are
ca
pt
ur
e
d,
hu
m
ans w
il
l be
d
et
ect
ed
and s
epar
at
e
d
f
r
om
t
he
im
age b
ac
kgr
ound
[5
]
.
This
pa
per
wa
s
aim
ed
to
dev
el
op
an
inte
gr
a
te
d
and
s
ophist
ic
at
ed
syst
e
m
fo
r
rec
ognizin
g
hu
m
ans
in
te
rm
s
of
a
set
of
points
f
ound
by
us
in
g
a
Kinect
se
nsor
,
pa
rtic
ularly
si
nce
not
re
quiring
a
cam
era
di
rectl
y
fo
c
us
e
d
on
the
hu
m
an
face
or
on
any
hum
an
bio
m
et
ric
li
ke
oth
e
r
syst
em
s
as
a
m
ajo
r
c
hal
le
ng
e
i
n
s
om
e
cases.
The
filt
ers
(Re
sam
ple,
Discreti
ze,
and
Sprea
d
s
ub
-
sam
ple)
wer
e u
se
d
as
a step
pre
-
proces
sing
t
o
treat
th
e
data
and
to
m
ini
m
ize
long
ti
m
e
co
ur
ses
w
hile
inc
reasin
g
syst
em
perf
or
m
ance,
wh
ic
h
is
m
easur
e
d
via
pa
ss
t
he
data
into
six
ty
pes
of
cl
assifi
ers
(
Sequentia
l
Mi
nim
a
l
Op
tim
izati
on
,
Decisı
on
Tree,
Naï
ve
Ba
ye
s,
Ra
nd
om
Tree,
Rule,
an
d
Ba
ye
s
Net).
Th
e
influ
e
nce
of
the
filt
ers
on
syst
em
eff
ic
ie
ncy
was
te
ste
d
to
unde
rstan
d
the app
ropr
ia
te
ness of fil
te
rs f
or su
c
h
cl
assi
fiers a
nd their
posit
ive ef
fects i
n
ide
ntifyi
ng s
yst
e
m
s.
2.
REVIEW
OF
LIT
ERATUR
E
Diff
e
re
nt
ways
of
rec
ognizin
g
pe
op
le
by
their
gaits
are
wide
m
entione
d
within
the
li
te
ratur
e
f
or
sever
al
ye
a
rs.
The
first
a
dd
psy
cho
l
og
ist
s
a
dm
inist
rated
th
is
sp
ace
i
n
1971,
once
J
ohans
so
n
c
onnected
li
gh
t
-
weig
ht
points
to
the
joints
of
people’s
bo
dies
duri
ng
a
dark
sp
ace
.
Pa
rtic
ip
ants
wer
e
t
hen
aske
d
to
r
un,
r
un,
or
ride
a
bicy
cl
e
[
10
,
11]
.
The
re
su
lt
s
prom
pt
that
ind
ivid
uals
will
acknow
le
dg
e
on
e
a
noth
er
by
their
in
div
id
ual
walkin
g
desig
ns.
The
bio
m
echan
ic
s
st
ud
ie
s
of
P
er
ry
et
al
.
[12],
Mu
rr
ay
[
13
]
a
nd
Winte
r
[
14
]
sem
ic
on
du
ct
or
diode
to
t
he
i
dea
that
gait
m
ay
be
a
cha
racteri
sti
c
an
d
pro
b
ably
in
di
vidual
at
trib
ute
of
a
n
i
nd
i
vid
ual.
Gait
recog
niti
on
m
ay
be
a
patte
rn
rec
ogniti
on
draw
bac
k.
Most
of
t
he
pr
evail
ing
gait
re
cogniti
on
a
ppr
oach
e
s
dep
e
nd
upon
A
N
a
naly
sis
of
t
he
bin
a
r
y
sil
houette
of
wal
king
pe
rsons
f
or
id
entifi
cat
ion
[
15
,
16]
.
Cutt
in
g,
et
al
.
[
17
]
st
ud
ie
d
hu
m
an
per
ce
pt
ion
of
gait
exp
l
oitat
ion
m
ov
in
g
li
ght
-
weigh
t
dis
play
s
(
ML
D
)
the
sam
e
as
th
at
e
m
plo
ye
d
by
Jo
hansso
n
a
nd
s
howe
d
hu
m
an
per
s
on
id
entifi
cat
ion
res
ults
[18]
and
gende
r
cl
assifi
cat
ion
r
esults
[
19
]
.
T
hey
sho
wed
that
hum
an
ob
ser
ve
rs
m
ay
determ
ine
ge
nder
with
m
or
e
or
le
s
s
seve
ntiet
h
accuracy
e
xp
l
oitat
ion
s
olely
the
visu
al
c
ues
f
r
om
MLD.
Bo
bick
an
d
J
ohnso
n
[
20]
cal
culat
e
fou
r
distances
of
hu
m
an
bodies,
pa
rtic
ularly
the
ga
p
betwee
n
the
pin
nacle
a
nd
f
oo
t,
the
ga
p
be
tween
the
pinn
acl
e
and
pel
vis,
the
ga
p
bet
wee
n
the
f
oo
t
a
nd
pe
lvis,
an
d
t
herefo
re
the
dista
nce
betwee
n
t
he
le
ft
foot
an
d
ri
gh
t
foot.
They
us
e
the
f
our
dista
nc
es
to
m
ake
2
t
ea
m
s
of
sta
ti
c
body
par
am
et
e
rs
a
nd
re
veal
t
hat
the
sec
ond
set
of
par
am
et
ers
are
add
it
io
nal
view
-
in
var
i
ant
com
par
ison
to
the
pr
im
ary
set
of
body
par
am
et
ers.
Give
n
the
flexi
bili
ty
of
hu
m
ans
to
s
po
t
per
s
ons
an
d
cl
assify
ge
nder
by
the
j
oi
nt
ang
le
s
of
a
wa
lking
s
ubj
e
ct
,
Rob
e
rt
Hu
tc
hings
Godd
a
r
d
[
21
]
devel
op
e
d
a
c
onne
ct
ion
ist
f
or
m
ula
f
or
gait
re
cogniti
on
ex
pl
oi
ta
ti
on
j
oi
nt
locat
ions
ob
ta
ine
d
f
ro
m
MLD.
H
ow
e
ver,
com
pu
ti
ng
joint
an
gles
fr
om
vid
eo
s
equ
e
nce
rem
ain
s
a
tough
drawb
ac
k,
al
tho
ug
h
m
any
trie
s
are
c
reat
ed
the
re
on
[
2
2
-
24]
.
T
her
e
is
a
var
ie
ty
of
l
ooks
ba
sed
m
os
tl
y
on
al
go
rith
m
s
fo
r
gait
an
d
act
ivit
y
recog
niti
o
n.
Dealer
a
nd
Da
vis
[
25]
us
e
d
s
el
f
-
co
rr
el
at
io
n
of
m
ov
ing
f
or
e
gro
und
obj
ect
s
to
te
ll
apar
t
walkin
g
hu
m
ans
f
ro
m
diff
ere
nt
m
ov
in
g
ob
j
ect
s
li
ke
c
ars.
P
olana
an
d
Nels
on
[
26
]
detect
ed
re
gula
rity
in
op
ti
cal
fl
ow
a
nd
us
ed
these
t
o
ack
nowled
ge
act
ivit
ie
s
li
ke
fr
ogs
jum
pin
g
a
nd
hum
an
walkin
g.
Ver
y
li
tt
le
an
d
Boyd
[27]
us
e
d
m
o
m
ent
opti
on
s
a
nd
re
gula
rity
of
f
or
e
gro
und
sil
houette
s
an
d
op
ti
cal
flow
to
spot
wa
lkers.
Nixon,
et
al
.
[
28
]
us
e
d
pr
i
nc
ipal
pa
rt
analy
sis
of
pictures
of
a
walki
ng
per
s
on
to
s
po
t
the
walke
r
by
gait.
S
hutl
er,
et
al
.
[29]
us
e
d
higher
-
or
der
m
ome
nts
su
m
m
e
d
ove
r
se
qu
e
nt
pictures
of
a
walkin
g
se
qu
ence
as
op
ti
ons
within
the
ta
sk
of
di
sti
ng
uis
hing
pe
rsons
by
thei
r
gait.
T
he
work
delineat
ed
durin
g
this
pa
per
is
cl
os
el
y
associ
at
ed
with
th
at
of
ver
y
li
tt
le
and
B
oyd
[30].
H
ow
e
ve
r,
rather
t
han
ex
plo
it
at
ion
m
ome
nt
descr
i
ptions
a
nd
re
gula
rity
of
the
w
ho
le
sil
houette
a
nd
op
ti
cal
flo
w
of
a
walke
r,
we
te
nd
to
di
vide
the
sil
houette
s
into
re
gions
a
nd
ci
pher
sta
ti
sti
cs
on
t
hese
r
eg
ion
s
.
W
e
te
nd
t
o
ad
d
m
or
e
stu
dy
the
capa
bili
ty
of
our o
ption
s
in
t
asks o
n
the
f
a
r si
de pers
on ide
ntific
at
ion
,
li
ke
g
e
nd
e
r
cl
assifi
cat
ion
.
3.
DA
T
AS
ET
Ther
e
are
a
fe
w
ga
it
sens
or
s
database
s,
a
nd
the
data
us
e
d
in
t
his
pa
pe
r
has
bee
n
coll
ect
ed
us
i
ng
a
Kinect
se
nso
rs
de
vice.
T
he
gait
was
recorde
d
f
or
a
gr
oup
of
volu
ntee
rs
of
49
perso
ns
,
9
of
whom
wer
e
wo
m
en
a
nd
th
e
rest
wer
e
m
en.
Each
pe
rs
on
wa
s
te
ste
d
f
or
wal
king
five
ti
m
es
fo
r
both
le
ft
a
nd
rig
ht
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.
10
, No
.
3
,
J
une
2020
:
29
7
8
-
298
5
2980
directi
ons,
in
fron
t
of
the
s
ens
or
at
a
90
-
degree
an
gl
e,
and
heig
ht
of
the
0.
6
m
et
e
rs
from
the
gr
ou
nd.
The
Kinect
sen
so
r
prese
nts
th
e
hum
an
sk
el
et
on
see
F
ig
ur
e
1
,
w
hich
pro
vi
des
value
s
of
X,
a
nd
Y
c
oor
din
at
es
for
20
po
i
nts
f
ro
m
head
to
t
oe
,
w
her
e
eac
h
per
s
on
wa
s
re
gi
ste
red
be
twee
n
90
a
nd
19
0
r
ecords,
w
hich
re
su
lt
s
in 84
04 r
ec
ord
s for
al
l perso
ns.
F
igure
1
.
A h
um
an
sk
el
et
on
by
K
inect
se
nso
r
4.
BACKG
ROU
ND
4
.
1.
Ba
yesia
n
network
s
Ba
ye
sia
n
netw
orks
a
re
a
ki
nd
of
pr
ob
a
bili
sti
c
grap
hical
m
a
nn
e
quin
t
hat
use
s
Ba
ye
sia
n
i
nf
e
ren
ce
f
or
chan
ce
c
om
pu
ta
ti
on
s
.
Ba
ye
sia
n
netw
ork
s
go
al
to
m
a
nn
e
quin
c
ondi
ti
on
al
de
pend
ence,
a
nd
the
refore
causati
on,
via
represe
nting
conditi
on
al
de
pende
nce
by
us
in
g
e
dg
e
s
in
a
directed
gr
a
ph.
T
hro
ugh
the
s
e
relat
ion
s
hip
s
,
on
e
ca
n
c
orre
ct
ly
hab
it
infe
ren
ce
on
the
r
andom
var
ia
bl
es
in
the
gr
a
ph
th
ro
ugh
t
he
us
e
of
factors
[
3
1
,
3
2
]
.
Two
it
is
al
s
o
reg
a
r
ded
as
"
belie
f
netw
ork
s"
or
"causal
ne
tworks"
are
grap
hical
fash
i
ons
f
or
represe
nting
m
ulti
var
ia
te
ch
a
nce d
ist
ri
bu
ti
ons
. Eac
h varia
bl
e X
i
is rep
res
e
nted
as a
ver
te
x
in a
directed
acy
cl
i
c
gr
a
ph (
"
da
g")
;
t
he
pr
ob
a
bili
ty
d
ist
rib
ution
(
1
,
2
,
3
,
…
,
)
is represe
nted
i
n fact
or
iz
e
d for
m
as f
ollow
s:
(
1
,
2
,
3
,
…
,
)
=
∏
(
1
|
П
)
=
1
w
he
re
П
is
t
he
s
et
of
ve
rtic
es
t
hat
a
re
'
s
pa
re
nts
i
n
the
gra
ph.
A
Ba
ye
sia
n
netw
ork
is
f
ully
sp
eci
fied
by
the co
m
bin
at
io
n of
:
The gra
ph str
uc
ture, i
.e., w
ha
t directe
d arcs
exist i
n
t
he gra
ph.
The pr
obabili
ty
table
(
1
|
П
)
f
or eac
h varia
ble
.
It
can
be
use
d
for
a
huge
ra
nge
of
ta
sk
s
to
ge
ther
with
pre
dicti
on
,
a
no
m
al
y
detect
ion
,
di
agnostic
s,
a
ut
om
ati
c
insig
ht,
rea
soni
ng, tim
e series p
re
dicti
on and
decisi
on m
aking
unde
r un
ce
rtai
nty [
29,
3
1
,
33
].
4
.
2.
Decisi
on
tree
algorithm
J48
J4
8
cl
assi
fier
is
an
easy
C4.
5
sel
ect
ion
tree
for
cl
assifi
cat
ion.
It
create
s
a
bin
ary
tree.
T
he
sel
ect
ion
tree
strat
egy
is
m
os
t
ben
e
fici
al
in
the
cl
assifi
cat
ion
pr
oble
m
[3
4
]
.
De
ci
sion
Tre
es
e
m
bo
dy
a
su
pe
rv
ise
d
cl
assifi
cat
ion
a
ppr
oach.
A
de
ci
sion
tree
is
a
si
m
ple
structu
re
the
place
no
n
-
te
rm
inal
node
s
sign
i
fy
chec
ks
on
on
e
or
e
xtra
at
trib
utes
an
d
te
r
m
inal
nodes
re
plica
te
sel
ect
ion
o
utc
om
es.
The
c
on
ce
pt got h
ere
f
ro
m
the n
orm
a
l
tree
sh
a
pe
w
hich
is
m
ade
up
of
a
r
oot
an
d
node
s
(t
he
po
sit
ion
s
w
her
e
pla
ces
branc
hes
di
vid
e)
,
br
a
nch
e
s
an
d
le
aves.
In
the
sam
e
way,
the
ch
oice
tree
c
on
sist
s
of
node
s
w
hich
sta
nd
for
ci
rcles
,
t
he
br
a
nc
hes
st
and
f
or
segm
ents
con
ne
ct
ing
the
node
s.
A
Decisi
on
Tree
beg
i
ns
f
ro
m
the
ro
ot,
strikes
dow
nwa
rd
an
d
norm
al
l
y
are
dr
a
w
n
f
ro
m
le
ft
to
rig
ht,
so
it
is
le
ss
dif
ficult
to
dr
a
w
it
.
T
he
node
f
ro
m
w
her
e
t
he
tree
be
gin
s
is
re
ferre
d
to
as
a
root
node
.
T
he
no
de
the
n
pl
ace
the
chain
ends
is
recog
ni
zed
as
the
“l
eaf”
node
.
Fro
m
ever
y
interi
or
node
(i.e.
no
longe
r
a
le
af)
m
ay
add
it
ion
al
ly
gro
w
out
two
or
m
or
e
br
anc
hes
i.e.
a
no
de
tha
t
is
no
w
no
t
a
le
af
node.
A
node
repres
ents
a
s
ur
e
at
trib
ute
w
hi
le
the
br
a
nc
hes
si
gn
i
fy
a
ra
nge
of
val
ues.
T
he
se
ra
ng
es
of
va
lues
act
as
pa
rtit
ion
po
i
nts
f
or
th
e
set
of
values
of
the
gi
ven
char
act
e
risti
c.
Figure
2
desc
r
ibes
the
str
uctur
e
of
a
tree
[29,
30]
.
J4
8
is
an
e
xtension
of
I
D3.
The
ad
diti
on
al
featur
es
of
J
48
a
re
acco
unti
ng
for
m
issi
ng
values
,
decisi
on trees
pru
ning,
c
onti
nuous
at
tribu
te
value ra
ng
es
, d
erivati
on
of
r
ules, etc.
[35, 3
6]
.
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
Hum
an gait
re
cogniti
on usi
ng
pr
e
pr
oce
ssin
g and cl
as
sif
ic
ation t
ech
niqu
es
(
Samer
K
ais
Jameel
)
2981
Figure
2.
Str
uc
ture of
a tree
[
34
]
4.
3
.
N
aïv
e
b
ayes
The
Nai
ve
Ba
ye
s
cl
assifi
er
is
a
si
m
ple
pr
obabili
sti
c
cl
assifi
er
base
d
on
app
ly
in
g
Ba
ye
s'
Theo
rem
with
str
ong
in
dep
e
ndence
as
su
m
ption
s,
w
hi
ch
ass
um
es
all
of
the
featu
r
es
are
eq
ually
ind
e
pe
ndent.
It
us
e
s
a
Ba
ye
sia
n
al
go
rithm
fo
r
the
total
pr
oba
bili
ty
pr
oced
ure,
the
pr
i
nciple
is
accor
ding
to
the
probabil
it
y
that
the
te
xt
belo
ng
s
to
a
cat
egory
of
pr
i
or
prob
a
bi
li
t
y,
and
the
te
xt
w
ou
l
d
be
as
sign
e
d
to
the
c
at
egory
of
poste
rio
r
pro
bab
il
it
y.
In
si
m
ple
te
rm
s,
a
naive
Ba
ye
s
cl
assifi
er
ass
um
es
that
the
presence
(
or
a
bse
nce)
of
a
pa
r
ti
cular
featur
e
of a
class i
s un
relat
ed t
o
the
presenc
e
(or
a
bs
e
nce) o
f
a
ny o
t
her
feat
ur
e
[
3
7
].
4.4
.
P
art
PA
RT
is
a
sep
arate
-
an
d
-
co
nq
uer
r
ule
le
ar
ne
r.
T
he
al
gorith
m
pr
oduci
ng
s
et
s
of
r
ules
cal
le
d
„decisi
on
li
sts‟
wh
ic
h
ar
e
pla
nn
e
d
set
of
r
ules.
A
ne
w
data
is
c
ompare
d
t
o
eac
h
r
ule
in
the
li
st
i
n
tu
r
n,
a
nd
t
he
it
e
m
is
assigne
d
t
he
cl
ass
of
t
he
first
m
a
tc
hin
g
r
ul
e.
P
ART
bu
il
ds
a
pa
rtia
l
C4.
5
decisi
on
t
ree
i
n
eac
h
it
erati
on
a
nd
m
akes th
e “
bes
t”
leaf into
a
rul
e
[
37
].
4
.5
. R
andom
trees
Ra
ndom
trees
hav
e
a
dd
e
d
via
Leo
Brei
m
an
and
A
dele
C
utler
[
38
]
.
T
he
r
andom
tree
is
a
tree
dr
a
w
n
at
ran
dom
fr
om
a
set
of
feas
ible
trees.
I
n
this
co
ntext
‘
‘at
rando
m
’’
abili
ty
that
each
tree
in
the
set
of
tim
ber
has
a
n
e
qual
da
ng
e
r
of
bein
g
sam
pled.
A
no
ther
way
of
a
nnounci
ng
this
is
that
t
he
dis
tribu
ti
on
of
tre
es
is
‘
‘
un
i
form
’’
.
Ra
ndom
tim
ber
can
be
gen
e
rat
ed
co
rr
ect
ly
an
d
the
com
bin
at
ion
of
la
r
ge
un
it
s
of
ra
ndom
t
i
m
ber
ty
pical
ly
le
ads
to
accurate
m
od
el
s.
The
al
go
rithm
can
deal
with
each
cl
assifi
c
at
ion
an
d
re
gressi
on
issues
[
39
,
40
].
4.6.
S
MO alg
orithm
The
Se
quentia
l
Mi
ni
m
al
Op
ti
m
iz
at
ion
(S
MO)
al
go
rithm
was
pr
opos
e
d
by
John
C
.
Plat
t
in
1998
a
nd
becam
e
the
fas
te
st
qu
ad
rati
c
pro
gr
am
m
ing
op
ti
m
iz
ation
al
gorithm
,
especial
ly
fo
r
li
near
SV
M
an
d
sp
a
r
se
data
perform
ance
[
41
]
.
SMO
al
gorithm
is
der
ive
d
by
ta
king
th
e
id
ea
of
the
de
com
po
sit
ion
m
et
ho
d
t
o
it
s
extrem
e
and
op
ti
m
iz
ing
a
m
ini
m
al
su
bs
et
of
just
two
po
ints
at
each
it
erati
on
.
The
powe
r
of
this
t
echn
i
qu
e
resid
es
in
the
fact
that
th
e
op
ti
m
iz
ation
pro
blem
fo
r
tw
o
data
po
i
nts
a
dm
i
ts
an
analy
ti
cal
so
luti
on
,
el
i
m
inati
ng
the
nee
d
to use a
n
it
erati
ve qu
a
drat
ic
pr
ogram
m
ing
op
t
i
m
iz
er as
par
t
of the al
gorith
m
[
42
,
35
].
5.
RESU
LT
S
A
ND
DI
SCUS
S
ION
The
dataset
that
m
entıoned
bove
hav
e
exam
ıned
w
it
h
six
cl
assi
fiers,
Seq
ue
ntial
Mi
ni
m
al
Op
ti
m
iz
ation
(S
MO
),
Naïv
e
Ba
ye
s,
Decis
ıon
T
ree
(J48
),
Ra
ndom
Tre
e,
Rule
(PAR
T),
a
nd
Ba
ye
s
Ne,
resp
ect
ively
.
T
he
r
esults
i
nd
i
cat
e
that
by
usi
ng
PA
RT
cl
as
sifie
r
the
syst
em
giv
es
the
be
st
perf
or
m
ance
wi
th
higher
accu
rac
y
as
s
hown
in
T
able
1.
Th
e
r
ules
ca
n
be
ge
ner
at
e
d
to
re
pr
esent
a
base
to
cl
assify
as
a
ne
gative
and
posit
ive
cl
assifi
er,
as
wel
l
as
extract
the
ru
le
s
from
the
non
-
pru
ne
d
tr
ee.
I
ns
te
ad
,
or
der
t
he
r
ules,
s
ub
s
et
ru
le
s
are
or
dered
(class
orde
r
ing)
an
d
com
pu
te
the
descr
i
pt
ion
le
ng
t
h
of
each
subset,
w
her
e
the
cl
asse
s
that
hav
e
sm
all
le
ng
th
giv
e
n
a
high
pri
ori
ty
.
PAR
T
is
easy
to
gen
e
rated
r
ule,
there
fore,
th
e
syst
e
m
gen
erat
e
s
430
ru
le
s
from
the
m
entioned
dataset
which
giv
e
s pow
e
r
t
o
the
syst
e
m
to
cl
assify
n
e
w
insta
nc
e
s
ra
pid
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.
10
, No
.
3
,
J
une
2020
:
29
7
8
-
298
5
2982
The
perfo
rm
ance
s
of
D
T
a
nd
the
ra
ndom
tree
wer
e
quit
e
good,
DT
has
ba
sed
on
in
form
at
ion
the
ory
,
therefo
re,
with
this
hu
ge
num
ber
o
f
at
trib
ute
s
an
d
dif
fer
e
nt
cl
ass
e
s
wit
h
va
rio
us
in
sta
nce
,
it
is
m
or
e
co
m
plex
to
cal
culat
e
the
gain
perfect
ly
in
order
t
o
s
plit
the
inst
anc
es.
Com
par
iso
n
with
t
he
ot
he
r
cl
assifi
ers,
t
he
SMO
had
go
od
pe
rfor
m
ance
,
wh
e
r
e,
SMO
is
resil
ie
nt
to
the
ove
rf
it
ti
ng
dataset
beca
us
e
it
depends
on
only
a
sm
all
nu
m
ber
of
points
in
the
dataset
.
Wh
e
reas
the
syst
e
m
giv
es
bad
pe
rfo
r
m
ance
wh
e
n
usi
ng
naï
ve
Ba
y
es
and
Ba
ye
s
net
cl
assifi
ers,
wh
e
re
t
he
Naï
ve
Ba
ye
s
cl
assifi
er
is
no
t
updatable
because
the
es
tim
a
tor
val
ue
chose
n
base
d
on
a
nal
yz
ing
trai
ni
ng
set
.
Naïve
net
cl
assifi
er
us
i
ng
di
ff
e
ren
t
ty
pe
s
of
searc
h
al
gorithm
s
and
qu
al
it
y
m
easur
es,
so,
with s
uc
h
a
hu
ge d
at
aset
the
s
yst
e
m
g
ives we
akn
e
ss
perfor
m
ance.
Me
an
Absol
ute
Erro
r
(M
AE)
are
cal
culat
e
d
f
or
eac
h
cl
assif
ie
r
as
s
how
n
in
T
able
1
to
m
easur
e
ho
w
cl
os
e
pr
e
dicti
on
is
to
even
t
ual
ou
tc
om
es,
wh
ic
h
re
pres
ent
the
aver
a
ge
of
the
abs
ol
ute
err
ors,
R
el
at
iv
e
Ab
s
olu
te
E
r
ror
(RA
E)
is
cal
c
ulate
d
as
well
in
or
der
to
m
easur
e
the
uncer
ta
inty
com
par
e
to
act
ual
val
ue
s
an
d
cl
arify
the
rela
ti
ve
in
act
ual
value
how
m
uch
s
p
ace
dose
error
ta
ke
up,
wh
ic
h
i
s
r
ep
re
sented
a
s
per
c
entage.
These
m
easur
e
m
ents
ind
ic
at
e
that
the
SMO
cl
assifi
e
r
ha
s
the
highest
RAE
that
the
re
st
cl
assifi
ers,
wh
ic
h
exp
la
in
ho
w
th
is
cl
assifi
er
ch
oo
s
es
on
ly
the
sm
a
ll
sa
m
ples
of
t
he
dataset
to
create
sup
port
vecto
rs.
T
he
RA
E
rati
o
is al
s
o hig
h becau
se it
h
a
nd
le
s
the att
ri
bute
s
of
t
he data
set
ind
e
pe
nd
e
nt
ly
.
Table
1.
Me
as
ur
em
ent o
f
the
perform
ance o
f
the classi
fie
rs
Clas
sıf
ıers
Accurac
y
(
%)
Mean Abs
o
lu
te E
rr
o
r
Relativ
e Abs
o
lu
te E
rr
o
r
(
%)
Seq
u
en
tial M
in
i
m
a
l Opti
m
iz
atio
n
(
S
MO)
7
2
.39
0
.03
9
8
.01
Naïv
e Bay
es
5
1
.06
5
0
.02
5
0
.61
Decisıo
n
Tr
ee
(J4
8
)
7
5
.27
0
.01
2
7
.73
Ran
d
o
m
Tr
ee
7
2
.77
0
.01
2
7
.80
Ru
le (
PART
)
7
6
.75
0
.01
2
5
.20
Bayes N
et
6
8
.15
0
.03
3
4
.20
5
.
1.
Discre
tized filter
Fo
r
tra
nsfo
rm
i
ng
nu
m
erıcal
values
of
al
l
at
trıbu
te
s
into
cat
e
gorical
counter
par
t
a
discreti
z
e
filt
er
has
been
us
e
d,
where,
the
values
hav
e
usual
ly
discreti
zed
in
m
od
el
in
g
m
et
ho
d
base
d
n
fr
e
quencies
ta
bles
,
wh
ic
h
m
ay
i
m
pr
ov
e
t
he
perform
ance
of
the
syst
e
m
an
d
accu
rac
y
of
the
predi
ct
ion
m
od
el
.
T
his
filt
er
is
a
t
oo
l
t
o
reduce
no
n
-
li
ne
arit
y
and
noise
as
well
,
therefore
it
con
s
idere
d
to
ident
ify
the
ou
tl
ie
rs
and
m
issi
ng
valu
e
of
at
trib
utes.
T
he
pe
rfor
m
ance
of
the
syst
e
m
is
increased
th
rou
gh
usi
ng
th
e
discreti
ze
filter,
w
her
e
,
the
value
s
of
the
at
trib
utes
conver
te
d
f
ro
m
con
ti
nu
es
to
discrete
va
lues,
this
is
cl
early
visible
and
dem
on
strat
ed
by
the
a
ccu
racy
ra
ti
o
m
entioned
in
T
a
ble
2.
Th
e
refor
e
,
this
filt
er
can
be
c
onsidere
d
as
a
pre
-
processi
ng
sta
ge
f
or
the
ra
w
da
ta
,
t
his
data
pas
se
d
to
the
cl
assi
f
ie
rs
to
be
te
st
ed
,
t
he
resu
lt
s
s
howe
d
this
filt
er
is
s
uitable
t
o
s
uc
h
data
an
d
ca
pa
ble
to
i
ncr
ease
the
perform
ance
of
the
syst
em
that
us
in
g
S
MO,
Naive
Ba
ye
s,
an
d
Naiv
e
ne
t
cl
assifi
ers.
on
the
c
on
t
rar
y,
t
his
filt
er
ef
fect
ne
gativel
y
on
th
e
syst
e
m
that
usi
ng
decisi
on
t
ree
an
d
ra
ndom
tree
cl
assifi
er,
w
he
re
it
is
qu
it
e
diff
ic
ult
to
those
cl
assifi
ers
to
disti
nguis
h
a
m
on
g
cl
asses
,
because
,
al
l
values
discreti
zed
sim
ultaneo
us
ly
,
a
nd
the
case
s
of
diff
e
re
nt
cl
asses
gro
up
e
d
i
nto
the
sam
e
insta
nce,
wh
e
re
it
usual
ly
there
w
ou
l
d be
m
ixtur
e
of da
ta
f
r
om
sev
eral
cl
asses in eac
h i
nterv
al
.
Fr
om
the
T
abl
e
2
,
t
he
re
su
lt
s
ind
ic
at
e
that
t
he
SM
O
cl
assifi
er
ac
hieve
d
hi
gh
est
recog
niti
on
acc
ur
acy
rate,
this
is du
e
to
t
he
m
echani
s
m
of
t
his
filt
er,
w
her
e
it
m
ax
i
m
iz
es
the
interde
pe
nd
e
nce bet
ween
the v
a
riable
s
value
an
d
the c
la
ss labels,
m
i
nim
iz
e inf
or
m
at
ion
loss,
a
nd
r
edu
ce
the
nu
m
ber
of
v
al
ues
a
s co
ntin
uous
va
riable
assum
es
by
groupin
g
them
i
nto
a
nu
m
ber
of
i
nterv
al
s
or
bin
s
,
w
hich
c
orres
ponds
to
the
natu
re
of
SM
O
cl
assifi
er for de
al
ing
with
dat
a whic
h
m
ini
m
iz
e the opti
m
izati
on
of all
d
at
a.
Table
2
Me
as
urem
ent o
f
th
e
pe
rfor
m
ance of
the cla
ssifie
rs
us
in
g discreti
z
e filt
er
Clas
sıf
ıers
Accurac
y
(
%)
Mean Abs
o
lu
te E
rr
o
r
Relativ
e Abs
o
lu
te E
rr
o
r
(
%)
Seq
u
en
tial M
in
i
m
a
l Opti
m
iz
atio
n
(
S
MO)
9
1
.43
0
.03
9
8
.2
Naïv
e Bay
es
6
8
.64
0
.01
3
3
.42
Decisıo
n
Tr
ee
(J4
8
)
6
9
.72
0
.01
3
4
.12
Ran
d
o
m
Tr
ee
6
3
.86
0
.01
3
7
.25
Ru
le (
PART
)
7
6
.75
0
.01
2
5
.20
Bayes N
et
6
9
.53
0
.01
3
2
.39
5
.
2.
Res
am
ple fil
ter
The
dataset
is
con
sist
in
g
of
39
cl
asses,
so
m
e
of
these
cl
asses
hav
e
ver
y
fe
wer
cl
asses,
an
d
accor
dingly
,
th
ey
are
con
si
dered
as
an
unbal
anced
dataset
.
Ther
e
f
or
e,
us
i
ng
the
Re
sam
ple
filt
er
m
ay
in
crease
the
nu
m
ber
of
instances
of
th
e
cl
asses
has
f
ew
instance
s,
the
pro
duced
da
ta
set
is
stron
gl
y
biased
in
te
rm
s
of
cl
ass
fo
r
w
hich
only
a
few
s
a
m
ples
are
ava
il
able.
Table
3
sh
ows
that
ap
ply
this
filt
er
on
the
dataset
le
ads
t
o
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
Hum
an gait
re
cogniti
on usi
ng
pr
e
pr
oce
ssin
g and cl
as
sif
ic
ation t
ech
niqu
es
(
Samer
K
ais
Jameel
)
29
83
i
m
pr
ove
the
pe
rfor
m
ance
of
the
cl
assifi
ers
wh
e
re
this
filt
er
gi
ves
them
the
powe
r
to
r
ecognize
the
c
la
sse
s
eff
ic
ie
ntly
by
gen
e
rati
ng
instances
f
or
m
ino
rity
cl
asses
tha
t
m
ake
the
cl
a
sses
alm
os
t
balanced
.
The
De
ci
sion
tree,
Ra
ndom
tree
[
43
,
44
]
,
and
P
ART
cl
assifi
ers
are
res
ults
of
m
or
e
accuracy
by
us
i
ng
this
filt
er,
wh
e
re,
increasin
g
t
he
nu
m
ber
s
of
ins
ta
nces
giv
e
t
he
cl
assifi
es
wi
de
area
to
cal
cul
at
ing
the
gai
n
f
or
buil
ding
the
tree.
On
the
oth
e
r
side,
this
filt
er
a
dv
e
rsely
aff
ect
s
the
perform
a
nce
of
a
syst
em
that
us
ed
N
aï
ve
Ba
ye
s
and
Ba
ye
s
net
cl
assifi
ers
because
these
cl
assifi
ers
treat
the
data
ind
ep
end
e
ntly
.
Furtherm
or
e
,
util
iz
ing
this
filt
er
he
lps
t
o
decr
ease
the
MAE a
nd RA
E for
all
classi
fier
s.
Table
3
.
Me
as
ur
em
ent all
classi
fiers
perf
orm
ance u
si
ng
re
sam
ple
filt
er
Clas
sıf
ıers
Accurac
y
(
%)
Mean Abs
o
lu
te E
rr
o
r
Relativ
e Abs
o
lu
te E
rr
o
r
(
%)
(SM
O)
7
3
.21
0
.03
98
.01
Naïv
e Bay
es
5
1
.82
0
.02
4
9
.97
Decisıo
n
Tr
ee
(J4
8
)
8
4
.14
0
.00
1
17
.62
Ran
d
o
m
Tr
ee
8
5
.55
0
.00
9
1
4
.74
Ru
le (
PART
)
8
5
.73
0
.00
6
1
5
.88
Bayes N
et
7
0
.55
0
.01
3
1
.04
5.3.
Spre
ad s
ub
-
s
amp
le
F
ro
m
T
able
4
it
can
obser
ve
that
the
accu
ra
ci
es
we
re
decre
ased.
Wh
e
re,
thr
ough
us
in
g
the
s
pr
e
a
d
su
b
-
sam
ple
filter,
ra
ndom
ly
un
de
rsam
pling
the
m
ajo
rity
of
cl
asses,
there
f
or
e
,
the
nu
m
ber
of
instan
ces
of
one
cl
ass
is
bec
omi
ng
e
qual
to
th
e
num
ber
of
in
sta
nces
of
a
no
ther
.
T
he
data
we
us
e
are
un
balance
d
in
sta
nces
of
cl
asses;
there
f
or
e
,
em
plo
yi
ng
this
filt
er
go
i
ng
to
unde
rsam
pling
th
e
m
ajo
rity
cl
asses,
a
nd
re
du
c
e
the
ove
rsam
pl
ing
cl
asse
s,
s
o,
num
ber
of
instances
be
the
sam
e
as
the
fe
we
r
in
sta
nce
cl
asses
causi
ng
decr
eas
i
ng
the
pe
rfor
m
ance
of
the
syst
em
due
t
o
l
os
in
g
a
lot
of
i
ns
ta
nces
from
different
cl
asse
s
t
o
be
balance
d.
C
on
seq
uen
tl
y,
balancin
g
a
datas
et
has
m
any
cl
asses
with
a
big
ga
p
be
tw
een
the
num
ber
s
of
instances
a
f
fec
t
neg
at
ively
on
the
eff
ic
ie
ncy
of
the
syst
em
p
er
form
ance.
Obviously
,
us
i
ng
under
-
sam
pling
on
the d
at
aset
i
nclud
e
m
any cla
sses,
need
e
d
t
o
r
edu
ce
the
val
ue
s of MA
E a
nd RAE a
s il
lustr
at
ed
in
T
able
4.
Table
4
.
Me
as
ur
em
ent o
f
the
perform
ance o
f
the classi
fie
rs
us
in
g
S
pread
s
ub
-
sam
ple
filt
er
Clas
sıf
ıers
Accurac
y
(
%)
Mean Abs
o
lu
te E
rr
o
r
Relativ
e Abs
o
lu
te E
rr
o
r
(
%)
(SM
O
)
6
6
.36
0
.03
0
.13
Naïv
e Bay
es
5
1
.38
0
.02
0
.13
Decisıo
n
Tr
ee
(J4
8
)
6
8
.85
0
.01
0
.10
Ran
d
o
m
Tr
ee
6
6
.29
0
.01
0
.11
Ru
le (
PART
)
7
0
.06
0
.01
0
.10
Bayes N
et
6
3
.47
0
.01
0
.11
6.
CONCL
US
I
O
N
Thro
ugh
the
c
on
st
ru
ct
io
n
of
an
integ
rated
s
yst
e
m
witho
ut
us
in
g
de
script
or
s
,
to
disti
ng
uish
hu
m
an
s
base
d
on
a
set
of
points
obta
ined
by
usi
ng
a
Kinect
se
nso
r,
the
syst
em
c
an
ide
ntify
the
m
through
t
heir
gait
des
pite
the
diff
ere
nce
of
sp
e
eds,
di
recti
ons
,
ages
,
an
d
ge
nd
e
rs.
T
he
res
ults
ind
ic
at
e
the
syst
e
m
that
us
in
g
the
disc
reti
zed
filt
er
with
SM
O
cl
assifi
er
gi
ves
e
xcell
ent
pe
rfor
m
ance
m
or
e
tha
n
t
he
rest,
w
he
re
giv
es
91.3%
as r
ec
ogniti
on
rate f
or all
the
giv
e
n data
.
REFERE
NCE
S
[1]
Ahm
ed,
M.H.,
“
Kinec
t
-
B
ase
d
Hum
an
Gait
Rec
o
gnit
ion
Us
ing
Stat
i
c
and
D
y
n
amic
Feat
ure
s
,
”
Int
e
rnational
Journal
of
Computer
S
cienc
e
and
In
formation
S
ec
urit
y
,
v
ol.
14
,
no
.
12
,
p
p.
425
,
2016
.
[2]
Chowdhur
y
,
S.
M.
A.
,
Pa
twa
r
y
,
T
.
W
.
U.
,
and
R
ahman,
R.
,
“
Hum
an
rec
ogni
ti
on
usin
g
wire
l
ess
rout
er
signal
,”
Doc
tora
l
dissert
at
ion
,
Br
ac
Univ
ersity
,
2
019
.
[3]
W
ang,
K.,
Li
u
,
L.
,
Le
e
,
Y.,
Din
g,
X.,
and
L
in,
J
.
,
“
Nons
ta
ndar
d
Periodi
c
Gai
t
En
erg
y
Im
age
for
Gait
Re
cogn
it
io
n
and
Dat
a
Augm
ent
a
ti
on,
”
In
Chi
nese
Conf
ere
nc
e
on
Pa
tt
ern
Rec
ognit
ion
and
Co
mputer
Vi
sion
(
PR
CV)
,
Springe
r,
Cham
,
pp.
197
-
2
08,
Nov
.
2019
.
[4]
Choi,
S.,
Kim
,
J
.
,
Kim
,
W
.
,
and
Kim
,
C.
,
“
Skele
ton
-
base
d
Ga
it
Rec
ognition
via
Robust
Fram
e
-
le
vel
Ma
tc
hing
,
”
IEE
E
Tr
ansacti
o
ns on
Informatio
n
Forensic
s and
Sec
urit
y
,
vol. 14
,
no.
10
,
pp.
2577
-
2592
,
2019
.
[5]
W
ang,
Jin,
She,
Mar
y
,
Nah
ava
nd
i,
Saei
d
and
Kou
za
ni
,
Abbas
,
“
A
rev
ie
w
of
vision
-
base
d
gai
t
re
co
gnit
ion
m
et
hods
for
hum
an
ide
nti
ficat
ion
,
”
in
DICT
A
2010:
Proce
ed
ings
of
the
Digit
al
Im
age
Computing:
Techni
ques
and
Appl
ic
a
ti
on
,
I
EEE
,
Pis
c
atawa
y
,
N.J.,
pp
.
320
-
32
7
,
2010
.
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.
10
, No
.
3
,
J
une
2020
:
29
7
8
-
298
5
2984
[6]
Jam
ee
l,
S.K.
an
d
Manz
a,
R.
R
.
,
“
Color
i
m
age
segm
ent
at
ion
us
ing
wave
let,
”
In
t.
J
.
Appl
.
In
form.
Syst.
(
IJA
IS)
,
vol.
1
,
no
.
6
,
pp
.
1
-
4,
2012
.
[7]
Jam
ee
l,
S.,
“
Face
re
cogni
t
ion
s
ystem
using
PC
A
and
DCT
in
HM
M,”
Int
.
J
.
Ad
v.
Re
s.
Comput.
Comm
un.
Eng
,
vol.
4
,
no
.
1
,
pp
.
13
-
18
,
2015
.
[8]
Bashir
K.,
Xian
g
T.
and
Gong
S.,
“
Gait
rec
ogn
it
ion
without
su
bje
c
t
coope
r
at
io
n,
”
Pat
t
ern
Rec
ognit
ion
Letters
,
vol.
31
,
no
.
13
,
p
p.
2052
-
2060
,
2
010
.
[9]
W
ang,
L.,
Ta
n
,
T.
,
Ning
,
H.
an
d
Hu,
W
.
,
“
Silhoue
tte
anal
y
sis
-
base
d
gai
t
r
ec
og
nit
ion
for
hum
a
n
ide
nti
f
ic
a
ti
on
,
”
IEE
E
transacti
o
ns on
pattern an
aly
sis and
mach
i
ne
in
te
l
li
gen
ce
,
vol.
25
,
no
.
12
,
p
p.
1505
-
15[023]
,
2003
.
[10]
Kozlowski,
L.T
.
and
J.E
.
,
“
Cutt
ing,
R
ec
ogni
zing
the
sex
of
a
walke
r
fro
m
a
d
y
n
amic
poin
t
-
li
ght
displ
a
y
,
”
Pe
rce
p
ti
on
&
P
s
yc
hophysi
cs
,
vo
l
.
21
,
no
.
6
,
p
p
.
5
75
-
580
,
1977
.
[11]
Cutt
ing,
J
.
E.
an
d
L.
T
.
Kozlows
ki,
“
Rec
ogni
zi
n
g
frie
nds
b
y
their
walk:
Ga
it
pe
rce
pt
ion
without
famili
ar
ity
cu
es,
”
Bul
letin
of the
psyc
honomic
soci
e
ty
,
vol
.
9
,
no
.
5
,
p
p
.
353
-
356
,
19
77
.
[12]
J.
Perr
y
and
Other
s,
“
Gait
an
aly
s
is:
n
orm
al
and
pat
ho
logica
l
func
t
ion,”
Jo
urnal
of
P
edi
at
ric
Or
thopae
di
c
s
,
vol.
12
,
no
.
6
,
pp
.
815
,
1992
.
[13]
M.
P.
Murra
y
,
“
Gait
as
a
total
pat
t
ern
of
m
ovement,
”
Ame
ric
an
Journal
of
P
hysic
al
M
edi
c
in
e
,
vol
.
46
,
no
.
1
,
pp.
290
,
1967
.
[14]
D.
A.
W
inter
,
“
Biom
ec
hanics
a
nd
m
otor
control
of
hum
an
g
ai
t
,
”
Norm
al
El
derl
y
and
P
atho
logi
ca
l
,
pp
.
143
,
1991
.
[15]
Z.
Li
u
and
S.
Sarka
r,
“
Eff
e
ct
of
silhoue
t
te
quali
t
y
on
h
a
rd
prob
le
m
s
in
Gai
t
re
c
ognit
ion,”
I
EE
E
transacti
onson
systems,
man,
a
nd
cy
berne
t
ic
s
,
Part
B
,
C
y
ber
n
e
ti
cs
:
a
publicati
on
of
the
IEE
E
Sy
stems
,
Man,
and
Cy
b
ern
e
ti
c
s
Socie
t
y
,
vol
.
35
,
no.
2
,
pp
.
170
-
1
83,
Apr.
2005.
[16]
M.
Ni
xon,
“
Ne
w
Advanc
e
s
in
Autom
at
ic
Gait
Rec
ognition,
”
Information
Sec
uri
ty
Techni
ca
l
Repor
t
,
vol.
7,
no.
4
,
pp.
23
-
35
,
De
c.
2002.
[17]
J.
Cutt
ing
,
D.
Proffitt,
and
L
.
Ko
zl
ows
ki,
“
A
bio
m
ec
hani
c
a
l
invariant
for
ga
it
p
er
ce
pt
ion,
”
Journa
l
of
Ex
p
erime
nta
l
Psyc
holog
y:
Hu
man P
erc
ep
ti
on
and
Pe
rform
anc
e
,
vo
l.
4
,
no.
3
,
p
p.
357
–
372
,
197
8.
[18]
J.
Cutt
ing
and
L
.
Kozlowski,
“
Rec
ogni
zi
ng
fri
en
ds
by
the
ir
wa
lk
:
gai
t
per
c
ep
ti
on
without
famili
a
rity
cue
s,
”
Bull.
Psyc
honometri
c Soc.
,
vol
.
9
,
no
.
5
,
pp
.
353
–
356
,
1977.
[19]
L.
Kozlowski
and
J.
Cutt
ing,
“
Rec
ognizing
the
sex
of
a
walke
r
from
a
d
y
namic
poi
nt
-
li
ght
displ
a
y
,
”
Pe
rce
pti
on
and
Psyc
hophysi
cs,
vol.
21
,
no
.
6
,
pp
.
575
–
580
,
1977
.
[20]
A.
F.
Bobi
ck
a
nd
A.
Y.
Johns
on,
"G
ai
t
re
cog
nit
ion
using
static,
activit
y
-
spec
i
fic
p
ara
m
eters,"
in
Pro
ce
ed
ings
of
th
e
2001
I
EE
E
Computer
Society
Conf
e
renc
e
on
Computer
V
ision
a
nd
Pat
t
ern
R
e
cogni
ti
on
,
vo
l.
1
,
pp.
I
-
423
-
I
-
430
,
2001
.
[21]
N.
Goddar
d,
“
The
Pe
rc
ept
ion
o
f
Arti
cul
a
te
d
Mot
ion:
Recogni
zin
g
Mov
ing
Light
Displays
,”
PhD
the
sis,
Univer
si
t
y
of
Roche
st
er,
19
92.
[22]
C.
Bregler
and
J.
Mali
k
,
“
Tr
ac
king
p
eople
w
it
h
twists
and
expone
nt
ial
m
aps,
”
In
Proc.
of
IEEE
CV
PR
,
pp.
8
-
15
,
1998
.
[23]
D.
DiFranc
o,
T.
Cham,
and
J.
Rehg
,
“
Rec
ov
er
y
of
3D
a
rticulated
m
oti
on
from
2d
cor
responde
nce
s,”
Technica
l
report,
Compaq
Cambr
idge
Re
se
arch
Lab
,
p
p
.
71
-
73,
1999
.
[24]
H.
Sidenbladh,
M.
Black,
and
D.
Fleet
,
“
Stoch
asti
c
tracki
ng
of
3d
hum
an
figu
res
u
sing
2D
i
m
age
m
oti
on,
”
In
European
Conf
e
renc
e
on
Compu
te
r V
ision
(
ECCV
)
,
vol
.
2,
p
p.
70
2
–
718
,
2000
.
[25]
R.
Cutl
er
and
L.
Davis
,
“
R
obust
real
-
ti
m
e
per
iodic
m
otion
detec
t
ion,
a
naly
s
is,
and
ap
pli
c
at
ions,
”
IE
E
E
Tr
ansacti
ons on Patt
ern
Analysis and
Mac
h
ine In
te
lligen
ce
,
vol
.
2
2,
no
.
8
,
pp
.
781
-
796,
2000
.
[26]
R.
Polan
a and
R
.
Nelson,
“
Dete
c
t
ing
a
ct
iv
it
i
es,
”
I
n
CVPR
,
1993.
[27]
J.
Littl
e and
J.
B
o
y
d
,
“
Rec
ogn
iz
i
ng
people
b
y
their
ga
it
:
Th
e
shap
e
of
m
oti
on
,”
V
i
dere
,
vo
l.
1
,
no.
2
,
1998
.
[28]
M.
Nixon,
J.
C
a
rte
r,
J.
Nash,
P.
Huang,
D.
Cun
ado,
and
S.
Ste
vena
ge
,
“
Autom
at
i
c
gait
re
cognition,”
In
Mot
ion
Anal
ysis
and
Tr
ack
ing
(
Re
f. No.
1999/103)
,
IEE
Coll
oquium
o
n
,
p
p.
3/1
–
3
/6, 199
9.
[29]
J.
S
hutl
er,
M.
N
i
xon,
and
C.
H
arr
is,
“
Stat
isti
ca
l
g
a
it
r
e
cogni
t
ion
vi
a
vel
o
ci
t
y
m
om
ent
s,”
In
V
isual
B
iomet
rics,
I
EE
E
,
p
p.
10/1
–
10
/5, 2
000.
[30]
J.
Littl
e
and
J.
Bo
y
d
,
“
Rec
ogn
iz
i
ng
people
b
y
their
ga
it
:
The
shap
e
of
m
oti
on
,
”
Videre:
Journal
of
Computer
Vi
sio
n
Re
search
,
vol
.
1
,
no.
2
,
1998.
[31]
Ca
y
ci,
A.
,
Ei
b
e,
S.,
Mena
sa
lva
s,
E.
and
Sa
y
gin
,
Y.,
“
Ba
y
esi
an
n
et
works
to
pre
d
i
ct
da
ta
m
ini
ng
a
lgori
thm
beha
v
i
or
in
ubiqui
tous
c
om
puti
ng
envi
r
onm
ent
s,”
In
An
aly
sis
of
Soci
al
Me
dia
and
Ubiquit
ous
Data
,
Springer,
Berlin
,
Heide
lb
erg
,
pp.
119
-
141,
2010
.
[32]
Sharm
a,
H.,
Aa
kash,
A.,
and
Aggarwal
,
A
.
G.,
“
A
Supervised
Guest
Sati
sfac
tion
Cla
ss
ifi
cation
with
Revi
ew
T
ex
t
and
Ra
ti
ngs,
”
In
Data
Scienc
e
,
C
RC Pre
ss
,
pp.
26
7
-
284
,
2019
.
[33]
Murali
dhar
an,
V
.
and
Sugum
aran,
V.,
“
A
compa
rat
iv
e
stud
y
of
Naïve
Ba
y
es
cla
ss
ifi
er
and
B
a
y
e
s
net
class
ifi
er
f
or
fau
lt
dia
gnosis
of
m
onobloc
k
c
ent
rifug
al
pum
p
using
wave
le
t
ana
l
y
s
is,
”
Appl
i
e
d
Soft
Comput
in
g
,
vol
.
12
,
no
.
8
,
pp.
2023
-
2029
,
2012.
[34]
Maji
dpour,
J
.
an
d
Jam
ee
l,
S.K
.
,
“
Autom
at
ic
ima
ge
annotati
on
ba
se
on
Naïve
B
a
y
es
and
Dec
isio
n
Tre
e
class
ifi
er
s
using
MP
EG
-
7,
”
In
2019
5th
Confe
renc
e
on
Knowle
dge
Bas
ed
Engi
ne
ering
and
Innov
a
ti
o
n
(
KBE
I)
,
IEE
E,
pp.
007
-
012
,
20
19
.
[35]
Pati
l,
T.
R
.
and
Shere
kar
,
S.S.
,
“
Perform
anc
e
ana
l
y
s
is
of
Naive
Ba
y
es
and
J48
cl
assificat
ion
al
gorit
hm
for
data
cl
assifi
ca
t
ion,”
I
nte
rnational
jou
rnal
of compute
r
scie
n
ce and
app
li
cations
,
vo
l.
6,
no.
2
,
pp
.
256
-
2
61,
2013
.
[36]
Kale
em,
A.
,
an
d
Prabha
,
I.
S.,
“
Boosting
of
Urb
an
Sound
Cla
ss
i
fic
a
ti
on
using
Preproc
essing
Tec
hnique
s,”
Journ
al
of
Ad
vanc
ed
R
ese
arch
in
Dynami
c
and
Con
trol
S
y
stems
,
vol
.
11(0
5
-
Speci
a
l
Iss
ue)
,
pp.
2293
-
2304
,
2019
.
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
Hum
an gait
re
cogniti
on usi
ng
pr
e
pr
oce
ssin
g and cl
as
sif
ic
ation t
ech
niqu
es
(
Samer
K
ais
Jameel
)
2985
[37]
Parsania
,
V.
,
Bh
al
odi
y
a
,
N.
and
Jani
,
N.N.
,
“
Ap
pl
y
ing
Naïve
ba
y
es
,
Ba
y
esNet
,
PA
RT,
JRip
and
OneR
Algorit
hm
s
on
H
y
poth
y
roid
Data
base
fo
r
C
om
par
at
ive
Ana
l
y
sis,
I
nte
rnat
io
nal
Journal
of
Dar
shan
Instit
ute
on
Engi
n
ee
rin
g
Re
search
&
E
merging
Technol
og
ie
s,
vo
l. 3, no
.
1,
2014.
[38]
[Online
]
.
Avai
lable:
htt
p
:/
/www
.
stat
.
b
erk
ele
y
.
edu
/users/bre
iman/
RandomF
ore
sts/
[39]
W
endl
,
C
.
,
“
Novelty
D
et
e
ction
in
Convo
lut
ion
al
Neur
al
Netw
orks
Us
ing
Densit
y
Forests
(No
.
STUD
ENT)
,
”
SIE
Master
Proj
ec
t
,
EP
FL
,
2018.
[40]
Zha
o,
Y.
and
Z
hang,
Y.
,
“
Com
par
ison
of
d
ec
is
ion
tree
m
et
hod
s
for
findi
ng
act
ive
obj
ects,”
Ad
vanc
es
in
Spac
e
Re
search,
vol
.
4
1,
no
.
12
,
pp
.
19
55
-
1959,
2008
.
[41]
Plat
t, J.
,
“
Seque
nti
al m
ini
m
al
op
ti
m
iz
ation:
A fa
s
t
a
lgori
thm for
tr
ai
ning
support
v
ec
tor
m
ac
hin
es,
”
1998.
[42]
Fata
m
a,
A.
Z
.
,
Haque
,
A.
,
and
Khan,
M.
A.
“
A
Multi
Feat
ure
Based
Island
ing
Cla
ss
ifi
c
at
io
n
Te
chni
qu
e
fo
r
Distribut
ed
Gen
era
t
ion
S
y
s
te
m
s,”
In
2019
Int
ernati
onal
Con
fe
re
nce
on
Mac
h
ine
Learning,
Bi
g
Data,
Cloud
an
d
Parall
el Comput
ing
(
COMIT
Con
)
,
IEE
E
,
pp
.
160
-
166,
Feb
.
2019
.
[43]
Toppi
red
d
y
,
H.
K.
R.
,
Saini,
B.
,
and
Hada
,
P.
S.
,
“
Aca
demic
Enhance
m
ent
S
y
st
em
using
EDM
Ap
proa
ch,”
In
2019
Sec
ond
In
te
rnat
ional
Conf
ere
n
ce
on
Adv
an
ced
Computati
on
al
and
Comm
unic
ati
on
Paradi
gms
(
ICACCP
),
IEE
E
,
pp
. 1
-
10
,
Feb
.
2019
.
[44]
Ho,
T.
K.
,
“
Random
dec
ision
fore
sts,”
In
Procee
dings
of
3rd
int
ernati
onal
con
fer
enc
e
on
docum
ent
analy
sis
an
d
rec
ognit
ion
,
IEEE,
vo
l
.
1,
pp.
27
8
-
282,
Aug
.
199
5.
Evaluation Warning : The document was created with Spire.PDF for Python.