Int
ern
at
i
onal
Journ
al of
R
obot
ic
s
and
Autom
ati
on (I
JRA)
Vo
l.
8
,
No.
3
,
S
eptem
ber
201
9,
pp.
19
4
~
20
4
IS
S
N:
20
89
-
4856,
DOI: 10
.11
591/
i
jra
.
v
8
i
3
.
pp19
4
-
20
4
194
Journ
al h
om
e
page
:
http:
//
ia
escore.c
om/j
ourn
als/i
ndex.
ph
p/IJRA
Fuzzy ne
uro
-
gene
tic app
roach f
or fe
ature sele
ction
and ima
ge
classific
ation in
augmente
d realit
y systems
Rajendr
a
T
hil
ahar
C
.
,
Siv
ar
am
ak
ri
shn
an
R
.
Depa
rtment
o
f
P
roduc
ti
on
Techn
olog
y
,
MIT
,
An
na
Univer
si
t
y
,
C
henna
i
,
Ind
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ma
y
5
, 2
01
9
Re
vised
Ju
l
9
,
201
9
Accepte
d
J
ul
21
, 2
01
9
In
thi
s
pap
er,
a
new
appr
o
ac
h
for
implement
i
ng
an
Augm
ented
R
ea
l
ity
s
y
stem
b
y
applying
fuz
z
y
gen
eti
c
neur
al
n
et
wor
ks
is
proposed.
I
t
consists
of
two
component
s
namel
y
featu
re
select
ion
an
d
cl
assifi
cation
m
odule
s.
For
fea
tur
e
d
etec
t
ion,
ext
r
ac
t
io
n
and
se
le
ct
ion
,
the
proposed
m
odel
uses
a
fu
zzy
log
ic
ba
sed
in
cre
m
ental
fea
tur
e
select
ion
a
lgori
thm
whi
c
h
has
be
en
proposed
in
thi
s
work
in
orde
r
t
o
rec
ogni
ze
the
important
fe
at
ur
es
from
3D
images.
Moreov
er,
th
is
pap
er
ex
pla
ins
the
imple
m
ent
at
ion
and
r
esult
s
of
th
e
proposed
a
lgorithm
s
for
an
Augm
ent
ed
Re
al
ity
s
y
s
te
m
u
sing
image
rec
ogni
ti
on,
fe
a
ture
ext
ra
ct
ion
,
feature
se
lection
and
cl
assifi
ca
t
ion
b
y
conside
ring
the
globa
l
and
local
fea
tur
es
of
the
i
m
age
s.
For
thi
s
purpose,
we
propose
a
three
l
a
y
er
fu
zzy
neur
a
l
ne
tw
ork
that
h
as
bee
n
impleme
nte
d
b
ase
d
on
weight
ad
just
m
ent
s
using
fuz
z
y
rul
es
in
th
e
co
nvolut
ional
neur
al
ne
tworks
with
gen
etic
al
g
orit
hm
for
eff
e
ctive
op
ti
m
izati
on
of
ru
le
s.
Th
e
class
ifi
catio
n
al
gorit
hm
is
al
s
o
base
d
on
fuz
z
y
neur
o
-
g
ene
t
ic
appr
oac
h
which
co
nsists
of
two
phase
s
namel
y
Tr
ai
ning
ph
ase
and
te
sting
phase
.
During
t
he
training
phase
,
rul
es
ar
e
form
ed
base
d
o
n
objects
and
th
ese
ru
le
s
are
ap
pli
ed
during
the
t
esti
ng
ph
ase
for
r
ec
ogni
zi
n
g
the
ob
jects
which
c
an
b
e
use
d
in
robot
ics
for
eff
ective
ob
je
c
t
r
ec
ogn
it
ion
.
From
the
experim
ent
s
condu
cted
in
thi
s
work,
i
t
is
pro
ved
tha
t
the
p
roposed
m
odel
is
m
ore
accu
rat
e
in
3D
obje
c
t
re
cognitio
n.
Ke
yw
or
d
s
:
Augm
ented
rea
li
ty
Cl
assifi
cat
ion
Feat
ur
e
sel
ect
ion
Fu
zzy
lo
gic
Gen
et
ic
al
gorithm
Neural
netw
orks
Tem
po
ral
r
ules
Copyright
©
201
9
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Ra
j
en
dr
a
T
hilahar
C
.
,
Dep
a
rtm
ent o
f Prod
uctio
n
Tec
hnology,
MIT Cam
pu
s, An
na U
niv
e
rsity
,
Chen
nai
600044, I
nd
ia
.
Em
a
il
:
thil
ahar
74@
gm
ail.co
m
1.
INTROD
U
CTION
In
the
rece
nt
ye
ars,
the
resea
rch
e
rs
in
the
a
rea
of
im
age
pr
oces
sin
g,
com
pu
te
r
visio
n
a
nd
rob
otics
hav
e
s
hifte
d
from
2
Dim
ension
al
te
ch
nolo
gi
es
to
3
Dim
e
ns
io
nal
te
c
hnol
og
ie
s
f
or
ca
pt
ur
i
ng
a
nd
a
nal
yz
ing
i
m
ages
due
to
the r
ecent
a
dva
nce
in
c
om
pu
ti
ng
with
res
pect
to
m
e
m
or
y
an
d
processi
ng
powe
r.
A
n
a
ug
m
ented
reali
ty
syst
e
m
needs
to
pe
rform
m
any
op
erat
ion
s
i
n
real
ti
m
e
including
i
m
age
captu
rin
g
us
in
g
in
put
s
ens
or
s,
processi
ng
the
im
ages
and
pro
vid
i
ng
s
uitable
re
spo
ns
e
t
o
t
he
us
er
.
Re
search
in
r
obot
ic
s
and
a
ug
m
ented
reali
ty
syst
e
m
s
us
e
the
in
for
m
at
ion
te
chnolog
y
f
or
pro
vid
i
ng
e
ff
ic
ie
nt
s
olu
ti
on
i
n
m
edic
al
diagnosis,
m
il
it
ary
app
li
cat
io
ns
, o
nline e
du
c
at
ion, r
o
bo
ti
cs a
nd c
om
pu
te
r
ai
de
d m
anu
fact
ur
i
ng.
Augm
ented
re
al
it
y
helps
to
enh
a
nce
the
view
of
real
world
obj
ect
s
pr
ese
nt
in
t
he
physi
cal
env
i
ronm
ent
in
te
rm
s
of
vir
tual
el
e
m
ents
so
that
t
he
view
is
no
t
hing
but
the
blen
d
of
real
a
nd
virtu
al
el
e
m
ents
[
1
]
.
I
t
is
inte
racti
ve
in
nat
ur
e
a
nd
he
nce
it
li
ves
bet
ween
t
he
r
eal
it
y
and
the
virtu
al
w
or
l
d.
Using
virtu
al
reali
ty
the
si
m
ulati
on
of
real
world
is
po
ssi
ble
an
d
it
al
so
pro
vid
es
ad
diti
onal
featur
e
s
f
or
m
otion,
siz
ing
,
c
olorin
g
a
nd
e
ff
e
ct
iv
e
recog
niti
on
of
obj
ect
s
.
I
n
com
pu
te
r
visi
on
based
r
obotics,
tw
o
dim
en
sion
a
l
i
m
ages
we
re
use
d
f
or
f
oreca
s
ti
ng
t
he
ne
xt
m
ov
e
of
t
he
r
obot
by
a
naly
zi
ng
the
past
a
nd
t
he
cu
rr
e
nt
m
ov
es
of
the
r
obot
al
on
g
with
the
r
ules
gove
r
ning
change
of
the
en
vironm
ent
by
ap
plyi
ng
c
on
st
raint
sat
isf
act
ion
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
R
ob
&
A
uto
m
IS
S
N:
20
89
-
4856
Fuzz
y
n
e
uro
-
ge
netic
appro
ac
h
fo
r feat
ur
e
se
le
ct
ion
an
d
…
(
Rajen
dra T
hila
ha
r
C
.
)
195
te
chn
iq
ues
.
O
n
the
oth
er
ha
nd
i
n
au
gm
ented
reali
ty
,
the
visio
n
rec
ogniti
on
syst
em
of
a
r
obot
use
s
a
th
ree
dim
ension
al
r
e
nd
e
rin
g
e
ngine
for
blen
ding t
he virtual
obj
e
ct
s w
it
h
t
he
r
ea
l objects i
n
the
world
[
2
]
.
Learn
i
ng
is
a
n
im
po
rtant
phen
om
eno
n
in
the
a
rea
of
r
obotics
an
d
a
ug
m
ented
real
it
ie
s
du
e
to
the
nat
ur
e
of
t
he
im
ages
to
be
proces
sed
in
the
syst
em
s
fo
r
m
aking
the
rob
ot
m
ov
em
e
nts
m
or
e
ef
fici
ently
.
The
m
achine
l
earn
i
ng
resea
rc
her
s
in
the
past
pro
posed
fee
d
f
orward
a
nd
ba
ck
pro
pag
at
io
n
neural
netw
ork
f
or
pro
vid
in
g
ef
fe
ct
ive
trai
ning
for
m
aking
th
e
syst
e
m
s
intelli
gen
t.
H
ow
e
ver,
the
trai
ni
ng
ti
m
e
is
m
o
re
an
d
the
accu
racy
achieve
d
by
s
uch
syst
e
m
s
i
s
no
t
s
uffici
ent
to
m
ake
eff
ic
ie
nt
decisi
ons
in
m
any
real
li
f
e
app
li
cat
io
ns
in
cl
ud
in
g
c
om
pu
te
r
visi
on.
I
n
su
ch
a
sce
na
rio,
m
any
extensio
ns
wer
e
pro
po
se
d
by
va
rio
us
researc
hers to
the e
xisti
ng m
a
chine
le
a
rn
i
ng
al
gorithm
s.
These
al
gorith
m
s
increase
the
acc
ur
acy
a
nd
re
duce
the
co
nv
e
r
gen
ce
tim
e.
A
m
on
g
these,
dee
p
le
arn
in
g
is
cu
r
ren
tl
y
the
area
of
inte
rest
f
or
resea
rch
e
rs
i
n
the
a
reas
of
a
rtific
ia
l
intel
lig
ence
a
nd
m
a
chine
le
arn
in
g.
Mo
re
ov
e
r,
t
he
dee
p
le
arn
in
g
al
gorithm
s
m
ake
us
e
of
a
la
r
ge
-
scal
e
an
d
hierar
c
hi
cal
neural
network
wh
ic
h
pr
ov
i
de
s
m
or
e
eff
ic
ie
ncy
thr
ough
st
ron
g
co
nnect
io
ns
.
O
ne
of
the
m
ai
n
adv
a
nta
ges
of
d
ee
p
le
arn
i
ng
al
gorithm
is
t
hat,
suc
h
al
go
rithm
s
pr
ovide
gr
eat
pre
dicta
bili
ty
with
hig
h
acc
uracy
and
he
nce
sur
pa
sses
the
oth
er
tra
di
ti
on
al
m
achine
le
arn
i
ng
al
go
rithm
s
wh
ic
h
are
us
e
d
in
va
rio
us
a
ppli
cat
i
on
s
incl
ud
i
ng
i
m
age
processi
ng
a
nd
sp
eec
h
re
co
gn
it
ion
,
natu
ral
la
ngua
ge
proce
ssing
an
d
intel
li
gen
t
quest
io
n
ans
wer
i
ng
syst
e
m
s,
cro
ss
li
ngual s
yst
e
m
s an
d
m
a
chine
t
ran
sla
ti
on
[
3
,
4
]
.
In
the
past,
th
e
Ale
xNet
sys
tem
was
pro
pose
d
a
nd
it
us
ed
a
c
onvoluti
on
al
neural
ne
twork
that
com
pr
ise
d
of
60
m
i
ll
ion
pa
r
a
m
et
ers
as
at
tr
ibu
te
s
a
nd
hence
was
able
t
o
com
e
first
in
the
Im
ageN
et
Larg
e
Scal
e
Visu
al
Re
cogniti
on
C
halle
ng
e
by
prov
i
ding
hi
gher
accuracy
of
c
la
ssific
at
ion
a
nd
re
du
ce
d
e
rror
rate
wh
e
n
it
was
c
om
par
ed
with
oth
e
r
m
achine
le
ar
ning
al
gor
it
h
m
s.
A
fter
th
at
,
m
any
exte
nded
form
s
of
neural
netw
orks
incl
udin
g
ra
dial
ba
s
is
f
un
ct
i
on
bas
ed
ne
ur
al
net
w
orks
we
re
intr
oduce
d
a
nd
hence
they
we
re
a
ble
t
o
handle
c
om
ple
x
sit
uatio
ns
le
a
ding
to
t
he
us
e
of
dee
p
le
ar
ni
ng
al
gorithm
s
in
m
any
areas
of
im
age
pr
oc
es
sing.
On
e
am
on
g
th
e
i
m
age
proc
e
ssing
ap
plica
ti
on
nam
el
y
face
rec
ogniti
on
s
yst
e
m
s
us
ed
a
dataset
of
fac
es
an
d
cl
assifi
ed
them
u
sin
g
ne
ural
net
work
s a
nd pr
ov
i
ded
a
n
acc
uracy
r
at
e of 97.
45% in th
e e
ffec
ti
ve
recog
niti
on
of
faces
w
hic
h
ar
e
avail
able
in
a
database
of
L
abell
ed
Faces
in
th
e
Wild
a
nd
it
wo
r
ke
d
well
by
the
op
ti
m
izati
on
of tran
sfe
r
le
ar
ning alg
ori
thm
s [
5
].
In
im
age
pro
cessi
ng
syst
e
m
s,
featur
e
e
xtracti
on
pro
vi
des
la
r
ge
nu
m
ber
of
feat
ures.
Eve
n
in
the
be
nch
m
ark
dataset
s,
the
nu
m
ber
of
fea
tures
is
s
o
hi
gh
in
s
uch
a
w
ay
that
they
i
ncr
ease
t
he
ti
m
e
fo
r
cl
assifi
cat
ion
.
More
ov
e
r,
m
any
featu
res
obta
ined
t
hro
ugh
feat
ur
e
e
xtr
act
ion
ei
ther
do
not
co
ntribute
in
the
cl
assifi
cat
ion
proce
ss
or
t
hey
pro
vid
e
only
a
ne
gligibl
e
am
ou
nt
of
c
on
t
rib
ution
f
or
decisi
on
m
aking
but
increase
t
he
ti
m
e
ex
pone
ntial
ly
.
In
s
uch
a
scenari
o,
t
he
m
os
t
con
trib
uting
a
nd
im
po
r
ta
nt
featu
res
m
us
t
be
identifie
d.
M
ost
of
the
featu
r
e
sel
ect
ion
al
gorithm
s
perfor
m
featur
e
re
duct
ion
base
d
on
the
e
ntr
op
y
va
lues.
Su
c
h
te
c
hn
i
ques
filt
er
the
fea
tures
from
the
whole
set
of
fe
at
ur
es.
O
n
the
oth
e
r
hand,
the
increm
ental
fe
at
ure
sel
ect
ion
te
ch
ni
qu
es
sel
ect
a
su
bse
t
of
featu
res
an
d
a
naly
ze
them
first.
They
ad
d
ad
diti
on
al
fe
at
ur
e
s
after
perform
ing
cl
assifi
cat
ion
with
init
ia
l
set
of
featur
e
s
a
nd
by
analy
zi
ng
a
nd
a
dd
i
ng
ot
her
i
m
po
rta
nt
feat
ur
es
i
n
order
t
o
pe
rform
featur
e
opti
m
iz
at
ion
.
Ma
ny
op
ti
m
iz
at
ion
te
chn
iq
ues
a
re
avail
able
in
the
li
te
ratur
e
for
f
eat
ur
e
reducti
on
i
nclu
ding
ge
netic
al
gorithm
s,
r
ule o
ptim
iz
ation
te
chn
i
qu
e
s
a
nd
li
near pro
gr
am
m
ing
m
od
el
s.
A
m
on
g
them
,
swar
m
intel
li
gen
ce
t
echn
i
qu
e
s
a
nd
na
ture
ins
pire
d
te
chn
iq
ues
pro
vid
e
m
or
e
sm
art
decisi
ons
th
rou
gh
intel
li
gen
t deci
sion m
aking
[
6
].
Gen
et
ic
Algo
r
it
h
m
(GA)
is
one
su
c
h
opt
i
m
iz
ation
m
eth
od
that
has
been
r
esearc
he
d
by
m
any
researc
hers
in
the
areas
of
a
r
ti
fici
al
intel
l
igence,
m
ach
ine
le
arn
in
g,
i
ntr
us
io
n
detect
ion
syst
em
s
and
i
m
age
processi
ng
are
as.
T
he
G
As
f
ollow
the
sur
vi
val
of
t
he
fitt
est
m
od
el
by
s
el
ect
ing
tw
o
pa
ren
ts
at
a
ti
m
e
an
d
pro
du
ci
ng
chil
dr
e
n.
A
n
act
ivati
on
f
unct
io
n
is
us
e
d
to
fin
d
the
fitt
est
child
ren
w
hi
ch
will
be
us
ed
as
the
pa
ren
ts
for
the
nex
t
generati
on.
It
pro
vid
es
operat
ors
li
ke
cr
os
sov
er
an
d
m
utatio
n
t
o
ge
ne
rate
m
ore
nu
m
ber
of
off
sprin
gs
.
The
GA
proc
ess
use
s
a
heurist
ic
ap
proac
h
to
s
olv
e
t
he
op
ti
m
iz
at
ion
pro
ble
m
by
reducin
g
the
se
arch
s
pace.
It
can
ei
ther
wor
k
ind
ivi
du
al
ly
in
the
decisi
on
m
aking
proces
s
or
it
can
w
or
k
with
fu
zzy
r
ules
a
nd
m
achine
le
a
rn
i
ng
al
gorith
m
s
to
so
lve
a
n
optim
iz
ation
pro
blem
m
or
e
ef
fecti
vely
.
It
use
s
a
num
ber
of
ge
ner
at
io
n
te
ch
niq
ue
s
for
ge
nerat
ing
i
nd
i
vidua
ls
f
ro
m
the
give
n
popula
ti
on.
Ther
e
f
or
e,
it
is
m
or
e
su
it
able f
or
perform
ing
f
eat
ur
e sele
ct
ion
a
s
well
as classi
fi
cat
ion
[
7
].
Fu
zzy
set
the
or
y
wa
s
pro
pose
d
by
Za
de
h
(
1965)
[
8
]
and
it
ha
s
be
en
use
d
m
or
e
fr
e
qu
e
ntly
in
the
de
velo
pm
ent
of
intel
li
ge
nt
decisi
on
m
akin
g
syst
em
s
beca
us
e
of
it
s
sim
plicity
th
rou
gh
r
ules
a
nd
it
s
si
m
il
arity
to
hum
an
reasonin
g
based
on
qua
li
ta
ti
ve
reaso
ni
ng.
Ma
ny
f
uzz
y
m
achine
le
ar
ning
al
go
rithm
s
we
re
us
e
d
in
the
li
te
ratur
e
f
or
ge
ner
at
in
g
a
nd
ap
plyi
ng
r
ules
f
ro
m
a
dat
aset
.
S
uch
sy
stem
s
us
ed
st
and
a
r
d
m
e
m
ber
sh
ip
f
unct
i
ons
for
m
akin
g
e
ff
ic
ie
nt
decisi
ons.
Hi
gher
or
der
lo
gic
s
li
ke
f
uzzy
lo
gic
a
nd
te
m
po
r
al
log
ic
hav
e
t
he
capa
bili
ty
of
perf
or
m
ing
reas
on
ing
unde
r
unc
ertai
nty
throu
gh
t
he
gra
dation
of
t
ru
t
h
va
lues.
Tem
po
ral
lo
gic
is
use
f
ul
to
a
naly
ze
the
past
data
a
nd
is
al
s
o
u
sef
ul
t
o
pr
e
dict
the
fu
t
ur
e
us
in
g
c
urre
nt
a
nd
pa
st
data.
The
com
bin
at
io
n
of
fu
z
zy
log
ic
an
d
tem
po
ral
lo
gic
help
to
de
rive
infer
e
nce
r
ules
w
hich
ca
n
pe
rfor
m
exp
la
natio
n
ba
sed
reas
onin
g
and
a
nalo
gy
ba
sed
rea
sonin
g.
In
im
age
pro
cessi
ng
a
ppli
cat
ion
s,
fu
zzy
te
m
po
ral
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2089
-
4856
I
nt
J
R
ob
&
A
uto
m
,
Vo
l.
8
, No
.
3
,
Septem
ber 2
019
:
19
4
–
20
4
196
ru
le
s
can
be
co
m
bin
ed
with
ge
netic
al
gorith
m
s
fo
r
pe
rfor
m
ing
opti
m
iz
a
ti
on
i
n
t
he
proce
s
s
of
feat
ur
e
sel
ect
ion
and th
us
reduci
ng the classi
ficat
ion
ti
m
e lea
din
g
t
o
inc
rease
in classi
ficat
io
n
acc
ur
acy
.
In
t
his
work,
a
ne
uro
fu
zzy
cl
assifi
er
de
ve
lop
e
d
u
sin
g
c
onvoluti
onal
ne
ur
al
netw
orks
with
a
bias
functi
on
for
er
r
or
re
duct
ion
a
nd
fuzzy
r
ules
f
or
wei
ght
a
djustm
ents
is
al
so
pro
po
se
d.
Fi
na
ll
y
the
r
ules
de
rive
d
from
trai
nin
g
us
in
g
f
uzzy
ne
ur
al
net
works
are
us
e
d
t
o
cl
assify
the
sim
i
la
r
obj
ect
s
f
or
pro
vi
ding
e
ffec
ti
ve
recog
niti
on
.
T
he
pro
po
se
d
f
uzzy
ne
ural
ne
twork
us
e
s
sig
m
oid
al
fu
nctio
n
as
act
ivate
d
functi
on.
For
f
eat
ure
sel
ect
ion
t
rian
gu
la
r
m
e
m
ber
sh
ip
f
unct
ion
is
us
e
d
t
o
form
f
uzzy
ru
le
s.
On
the
ot
her
hand
du
rin
g
cl
assi
ficat
ion
trapez
oid
al
m
em
ber
sh
ip
f
un
ct
io
n
is
us
e
d
in
order
to
im
pr
ove
t
he
cl
assifi
cat
ion
accu
rac
y.
T
he
sim
ulati
on
is
carried
out
us
i
ng
m
at
la
b
si
m
ulati
on
t
oo
l
a
nd
the
res
ults
obta
ined
f
ro
m
t
his
w
ork
ha
ve
been
com
par
e
d
with
oth
e
r
existi
ng
syst
e
m
s
in
this
area.
F
r
om
th
e
exp
e
rim
ents
cond
ucted
us
in
g
a
r
obotic
a
ppli
cat
ion
sce
na
rio
in
wh
ic
h
the
r
obots
are
pro
gra
m
m
ed
fo
r
pic
k
a
nd
place
ta
sk
s
wh
ic
h
a
re
captu
red
by
a
sens
or
an
d
hav
i
ng
the
im
age
capt
ur
i
ng
syst
em
with
a
cam
era,
it
is
f
ound
t
ha
t
the
pro
pose
d
m
od
el
pro
vid
e
s
m
or
e
reali
sti
c
vie
ws
of r
eal
world
3 Dim
ension
al
obj
ect
s
.
The
im
age
based
rec
ogniti
on
syst
em
pr
opos
e
d
in
this
w
ork
co
ns
ist
s
of
sub
com
pone
nts
nam
el
y
recog
niti
on
of
visio
n
in
f
or
m
a
ti
on
base
d
on
i
m
age
proce
ssing
te
ch
niques
[9]
.
For
t
his
pur
pose
a
re
relaed
a
n
input
vid
e
o
is
div
ide
d
int
o
a
nu
m
ber
of
f
ra
m
es.
Am
on
g
al
l
these
fr
am
es
the
m
os
t
relevan
t
fr
am
es
to
the
r
obot
obj
ect
rec
ogniti
on
syst
em
pr
opos
e
d
i
n
t
his
w
ork.
It
co
nsi
der
s
the
key
f
ram
es
and
ana
ly
ses
them
based
o
n
colo
r,
te
x
tu
re,
m
ov
em
ent
an
d
de
pth
detai
ls
based
on
t
he
dim
ensions
of
the
im
age.
It
pe
rfo
rm
s
featur
e
identific
at
ion
us
in
g
f
uzzy
ne
ur
al
netw
ork
a
nd
e
xtracts
t
he
releva
nt
featu
res
by
a
pp
ly
in
g
f
uzzy
r
ules.
Fr
om
the
extracte
d
t
he
feat
ur
es
,
m
os
t
rele
van
t
fe
at
ur
es
a
re
se
le
ct
ed
by
a
pp
ly
ing
f
uzzy
r
ules
an
d
the
sel
ec
te
d
featur
e
s
are
use
d
as
in
put
fe
at
ur
e
vecto
r
to
the
cl
assifi
er.
The
m
ai
n
ad
van
ta
ge
of
t
he
pro
po
se
d
syst
e
m
is
the inc
rease i
n cl
assifi
cat
ion
a
ccur
acy
.
The
rest
of
th
is
pa
per
is
or
gan
iz
e
d
as
f
ol
low’s
sect
io
n
2
pr
ovides
a
s
urvey
of
relat
ed
wor
ks
in
the
areas
of
a
ug
m
ented
real
it
y,
neu
ral
ne
twork
,
f
uzzy
s
yst
e
m
s
and
cl
assifi
cat
ion
al
gorithm
s.
Sect
ion
3
exp
la
in
s
the
al
gorithm
s
propose
d
in
this
pa
per
al
ong
with
the
m
od
el
pro
pose
d
in
this
w
ork
.
Sect
io
n
4
de
picts
the
res
ults
obt
ai
ned
f
ro
m
thi
s
w
ork
a
nd
al
so
pro
vid
e
s
s
ui
ta
ble
disc
us
si
on
s
on
t
he
res
ult.
Sect
i
on
5
giv
es
con
cl
us
io
ns
on
this
work an
d al
so
prov
i
des
s
om
e su
ggest
io
ns
for feat
ure e
nh
a
ncem
ents.
2.
RELATE
D
W
ORKS
Ther
e
a
re
m
any
wo
r
ks
t
hat
ha
ve
bee
n
dev
el
op
e
d
by
var
i
ous
resea
rc
her
s
in
the
di
recti
on
of
feat
ur
e
sel
ect
ion
,
cl
as
sific
at
ion
an
d
aug
m
ented
rea
li
ty
in
the
pas
t
[
7,
8,
10
-
14
]
.
Am
on
g
them
,
M
oh
am
ma
d
et
al
(20
14)
de
velo
ped
a
nove
l
al
gorithm
cal
l
ed
e
voluti
onar
y
or
ie
nted
in
crem
ental
al
go
rithm
for
s
el
ect
ing
the
us
e
fu
l feat
ur
es b
y
usi
ng
their n
ew
fr
am
ework.
Mo
re
over,
they
h
ave
te
ste
d
their f
ram
ewor
k
on
a
n
or
din
a
ry
gen
et
ic
al
gorithm
and
al
s
o
de
velo
ped
ne
w
m
et
hods
.
I
n
a
dd
i
ti
on
,
t
heir
w
ork
em
plo
ye
d
with
tw
o
new
op
erators
for
perform
ing
the
ad
diti
on
op
e
rati
on
a
nd
dele
ti
on
operat
ion
that
are
us
e
d
for
cha
ngin
g
t
he
le
ng
th
of
the
m
et
ho
ds
ra
ndom
ly
.
The
m
ajo
r
a
dvanta
ge
of
thei
r
al
gorithm
is
that
i
t
increases
the
cl
assifi
cat
ion
a
ccur
ac
y
and re
duces th
e cla
ssific
at
ion t
i
m
e.
Sannasi
et
al
(
2016)
[1
5
]
introd
uced
an
intel
li
gen
t
C
onditi
on
al
Ra
ndom
Fiel
d
ba
sed
Feat
ur
e
Sele
ct
ion
al
gor
it
h
m
to
de
vel
op
a
n
e
ff
ic
ie
nt
intr
us
io
n
detect
ion
syst
em
.
M
or
e
over,
they
hav
e
us
ed
intel
li
gen
t
al
gorithm
s
for
ge
ne
rati
ng
ru
l
es
that
are
helpful
for
m
aking
e
ff
ect
i
ve
de
ci
sion
s
ove
r
t
he
K
D
D
C
up
da
ta
set
wh
ic
h
is
us
e
d
for
e
valuati
ng
their
int
ru
si
on
detect
ion
syst
em
.
They
hav
e
cond
ucted
nu
m
ber
of
ex
per
i
m
ents
and
pro
ve
d
tha
t
their
featu
re
sel
ect
ion
al
gor
it
h
m
ta
kes
le
ss
tim
e
and
le
ss
nu
m
ber
of
featur
es
t
hat
are
use
d
to
enh
a
nce
the
cl
assifi
cat
ion
a
c
cur
acy
.
Ga
na
p
at
hy
et
al
.
(20
13)
[
3
]
disc
usse
d
i
n
detai
l
about
t
he
intel
l
igent
te
chn
iq
ues
f
or
perf
or
m
ing
f
eat
ur
e
sel
ect
io
n
a
nd
cl
assi
ficat
ion
by
ap
pl
yi
ng
the
sel
ect
ed
an
d
c
on
t
ribu
ti
n
g
featur
e
s.
Mo
r
eov
e
r,
they
ha
ve
f
ocu
se
d
ov
e
r
the
sel
ect
ion
of
c
ontr
ibu
te
d
feature
s
that
are
us
efu
l
f
or
perform
ing
classi
ficat
ion
proc
ess ef
fecti
vely
for
int
ru
si
on det
ect
ion
syst
e
m
.
Ulrich
ne
um
ann
a
nd
Su
ya
(1999
)
[
16
]
dem
on
st
rated
t
hat
the
a
ug
m
ented
reali
ty
app
li
cat
ion
s
w
hic
h
direct
the
scen
e
an
no
ta
ti
on,
t
he
po
se
sta
bili
zat
ion
a
nd
the
exten
dib
le
trac
king
r
an
ge
t
hat
are
us
e
d
f
or
tr
acki
n
g
the
natu
ral
fe
at
ur
es
in
un
-
prepa
red
e
nv
i
ronm
ents.
Y
uan
et
al
(20
06)
[
5
]
de
vel
op
e
d
a
ne
w
re
gis
trat
io
n
al
gorithm
that
is
us
e
d
f
or
es
tim
a
ti
ng
t
he
m
et
ho
d
r
obus
t
ly
,
tracki
ng
th
e
feat
ur
e
a
nd
for
t
he
rec
onstructio
n
m
et
ho
d
w
hich
is
al
so
use
d
for
trac
king
the
nat
ur
al
fe
at
ur
es.
Mor
eo
ver,
they
ha
ve
super
im
po
s
ed
a
nd
the
virtu
al
ob
je
ct
s
over
the
a
rb
it
ra
ry
re
gion
by
the
us
e
r
pr
efer
red.
Niti
n
et
al
(
2018)
[1
7
]
de
vel
op
e
d
a
no
vel
m
et
ho
d
that
is
to
fin
d
t
he
obj
ect
e
dg
es
f
r
om
i
m
age.
Mor
eov
e
r,
they
ha
ve
m
od
ifie
d
th
e
f
uzzy
m
e
m
b
ersh
i
p
functi
on
that
use
d
t
o
c
on
c
entr
at
e
the
m
ulti
-
fo
cu
s
f
us
i
on
ov
er
the
f
uzzy
e
dge
detect
ion.
I
n
a
dd
it
io
n,
the
y
have
pro
ved
thei
r
m
et
hod
by
c
on
duct
ing
t
he
e
xp
erim
ents
with
t
he
sat
el
li
te
and
col
or
im
ages
that
perform
the
f
us
e
edg
e
detect
io
n
an
d
c
om
par
e
d
with
ot
her
e
xi
sti
ng
syst
em
s.
Ak
as
h
Ba
pat
e
t
al
(20
16)
[
18
]
dev
el
op
e
d
a
novel
schem
e
wh
ic
h
is
us
e
d
a
r
olli
ng
shutt
er
cam
e
ra
t
hat
us
e
d
for
trac
kin
g
the
6
D
OF
hea
d
pose
an
d
al
so
use
d
t
o
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
R
ob
&
A
uto
m
IS
S
N:
20
89
-
4856
Fuzz
y
n
e
uro
-
ge
netic
appro
ac
h
fo
r feat
ur
e
se
le
ct
ion
an
d
…
(
Rajen
dra T
hila
ha
r
C
.
)
197
validat
e
their
new
m
et
hod
with
the
help
of
G
oP
r
o
ca
m
eras
thr
ough
the
hel
p
of
sel
f
-
cal
ibr
at
ion
and
the au
t
om
at
ed
synch
ronizat
io
n.
Gholam
et
al
(20
15)
[
19
]
de
velo
ped
a
new
m
od
el
that
is
a
sim
ple
and
li
near
it
erati
ve
cl
us
te
rin
g
m
od
el
to
co
ns
t
ru
ct
t
he
s
uper
pix
el
s
with
a
novel
sc
hem
e
and
it
is
al
s
o
gen
e
rated
the
su
pe
r
pi
xels
by
us
in
g
cl
us
te
rin
g
m
eth
od
ba
sed
on
t
he
pro
xim
i
ty
and
the
c
olor
sim
il
arity
from
the
im
age
plane
.
Io
a
nnis
et
al
(
2012)
[
20
]
pro
posed
a
ne
w
tra
ns
m
issi
o
n
syst
em
,
desig
n,
co
ntr
ol
an
d
al
s
o
th
e
perf
or
m
ance
of
t
he
te
le
-
operati
on
force
fee
db
ac
k.
T
hey
ha
ve
a
chieve
d
bette
r
pe
rfo
rm
ance
than
the
oth
er
existi
ng
syst
em
s.
Kan
im
ozh
i
et
a
l
(20
18)
[
21
]
propose
d
a
new
f
uzzy
log
ic
bas
ed
pre
dicti
on
m
od
el
cal
le
d
In
te
ll
igent
Ri
sk
Pr
e
dicti
on
Sys
tem
fo
r
pr
e
dicti
ng
t
he
Breast
Ca
ncer
disease
by
usi
ng
t
he
f
uzzy
te
m
po
ral
ru
le
s.
They
ha
ve
ap
pl
ie
d
intel
li
gen
t
fu
zz
y
ru
le
s
f
or
perf
orm
ing
the
feat
ur
e
sel
ect
io
n
a
nd
cl
assifi
cat
io
n
proces
ses
ov
er
t
he
sta
ndar
d
be
nc
hm
ark
da
ta
set
cal
le
d
Breas
t C
ancer
d
at
aset
wh
ic
h
is a
vaila
ble in UCI
Ma
chine L
ear
ning
d
at
aset
.
Selvi
et
al
(2016
)
[
22
]
pr
opose
d
f
uzzy
l
og
ic
an
d
ti
m
e
co
ns
trai
nt
ba
sed
m
et
ho
d
f
or
perform
ing
eff
ect
ive
r
outi
ng
process
ove
r
the
wireless
s
ens
or
net
works
.
T
hey
hav
e
ac
hieve
d
bette
r
pe
rfor
m
ance
in
te
rm
s
of
net
wor
k
li
f
et
i
m
e
and
e
ff
i
ci
ency
over
th
e
com
m
un
ic
at
i
on
pr
ocess.
T
hi
s
is
due
t
o
t
he
fact
t
hat
the
us
e
of
fu
zzy
l
og
ic
a
nd
the
intel
li
gence
r
ules.
Set
hukk
a
rasi
et
al
(2014
)
[
23
]
intr
oduce
d
a
n
i
ntell
igent
neural
ne
twork
base
d
f
uzzy
lo
gic
w
hich
inc
orp
or
at
es
the
ti
m
e
con
strai
ns
for
r
e
pr
ese
ntin
g
the k
no
wled
ge
m
od
el
wh
ic
h
i
s
us
e
d
for
m
ining
the
te
m
po
ral
patte
rn
s
.
They
ha
ve
pr
ov
e
d
that
their
predict
io
n
syst
em
is
bette
r
i
n
t
he
direct
ion
o
f
diabeti
c
a
nd
he
art
disease
i
den
ti
ficat
io
n
a
nd
detect
ion
wh
e
n
it
is
c
om
par
ed
w
it
h
t
he
e
xisti
ng
sy
stem
s.
Jai
sank
a
r
et
al
(20
12)
[
24
]
de
velo
ped
an
i
nt
el
li
gen
t
age
nt
base
d
syst
em
by
ap
plyi
ng
t
he
fu
zzy
rou
gh
set
and
the
ou
tl
ie
r
dete
ct
ion
te
c
hn
i
ques
f
or
detect
in
g
t
he
i
ntr
ud
e
rs
.
T
hey
have
ac
hieve
d
bette
r
pe
rfor
m
ance
in
te
r
m
s
of
i
ntr
us
io
n
de
te
ct
ion
accu
r
acy
and
false
al
arm
rate.
Sairam
esh
et
al
(2015
)
[
25
]
pr
op
os
e
d
a
new
pr
e
dicti
on
syst
e
m
that
is
us
e
d
for
pre
dicti
ng
t
he
us
e
r
i
nterests
t
o
pro
vide
the
rele
van
t
inf
or
m
at
ion
by
us
i
ng
the
releva
nt
f
eedb
ac
k
a
nd
the
re
-
ra
nkin
g
pro
cess
.
Th
e
auth
or
s
achie
ve
d
bette
r
pred
ic
ti
on
accu
rac
y
than
the ex
ist
in
g
sy
stem
s.
A
new
m
od
el
c
al
le
d
he
ur
ist
ic
m
at
he
m
at
ic
a
l
m
od
el
that
has
been
pro
pose
d
by
Tah
riri
et
al
(20
15)
[
6
]
that
m
od
el
use
s
a
te
c
hniq
ue
f
or
opti
m
izing
t
he
r
obot
a
rm
m
ov
e
m
ent
tim
e
that
m
ini
m
izes
m
ake
-
sp
a
n
and
al
s
o
m
axi
m
iz
ed
the
total
nu
m
ber
of
unit
s
pro
duced
eac
h
day
i
n
rob
ot
cel
ls.
Their
m
od
el
use
s
the
use
rs
ke
ys
i
n
the
sta
ti
on,
unit
sect
ion
an
d
r
obot
i
nto
t
he
s
i
m
ulati
on
soft
war
e
after
crea
ti
ng
3D
file
s
a
nd
al
so
se
nd
i
ng
t
hem
to
the
virt
ual
r
eal
it
y
s
yst
e
m
.
More
ov
e
r,
the d
at
a
is
conve
rted
into
a
te
xt f
il
e
and
al
so
it
tr
ansf
e
rr
e
d
to
t
he
robo
t
arm
m
ov
e
m
ent
with
the
hel
p
of
ti
m
e
op
ti
m
iz
at
ion
al
gor
it
h
m
fo
r
deter
m
ining
the
optim
u
m
ta
sk
seq
uen
ce
.
WeiJi
e
W
a
ng. Hua
Ge
n Wan
[
26
]
desc
ribe
d t
he
cha
racteri
st
ic
s an
d t
he
ess
entia
l serv
ic
es
of
a
ugm
ented
r
eal
it
y
about
t
he
r
eal
world
ap
plica
ti
on
s
that
a
re
use
d
a
ug
m
ented
reali
ty
.
They
ha
ve
giv
e
n
br
ie
f
histor
y
a
bout
the
virtu
al
reali
ty
and
t
he
au
gme
nted
reali
ty
a
nd
t
he
w
ork
f
low
of
th
e
au
gm
ented
reali
ty
is
al
so
exp
l
ai
ned
.
More
ov
e
r,
t
he
y
hav
e
e
xpla
ined
t
he
diff
e
r
ent
sta
ges
of
the
va
rio
us
a
ugm
ented
reali
ty
app
li
cat
ion
s
su
c
h
as
Im
age
Acquisi
ti
on
process
,
t
he
Feat
ur
e
E
xt
racti
on
proce
s
s,
t
he
Feat
ur
e
Ma
tc
hing
pro
cess,
t
he
Geom
et
ric
Ver
ific
at
io
n
proces
s
an
d
th
e
associat
ed
inf
or
m
at
ion
r
et
rieval
proce
ss.
I
n
ad
diti
on,
they
highli
gh
te
d
the
esse
ntial
ne
eds
a
nd
the
e
ssence
of
the
aug
m
ented
rea
li
ty
in
their
w
ork.
H
ow
e
ve
r,
m
os
t
of
the
works
pr
ese
nt
in
t
he
li
te
ratur
e
are
not
able
to
pro
vid
e
s
uffici
ent
cl
assifi
cat
ion
accuracy
du
e
t
o
the
la
c
k
of
bette
r
featur
e
sel
ect
ion.
Hen
ce
a
ne
w
featu
re
sel
ect
ion
al
gorith
m
is
pro
po
se
d
this
pap
e
r
t
o
handle
t
he
acc
ur
acy
of cla
ssific
at
io
n.
3.
SY
STE
M AR
CHI
TE
CT
U
R
E
The
ov
e
rall
ar
chite
ct
ur
e
of
t
he
pr
opos
e
d
cl
assifi
cat
ion
sy
stem
is
sh
own
in
Fig
ur
e
1.
I
t
con
sist
s
of
seve
n
m
ajo
r
c
om
po
ne
nts
na
m
el
y
Au
gm
ented
Re
al
it
y
Im
age
Dataset
,
Ca
m
era
Ca
pture
Im
age
Dataset
,
Use
r
In
te
r
face
M
odule,
Object
Re
cogniti
on
Syst
e
m
,
Decisi
on
Ma
nag
e
r,
T
em
poral
I
nfo
rm
ation
Ma
nag
e
r,
Fu
zzy
Rule B
ase a
nd
Kno
wled
ge ba
se
.
Augme
nted
R
ealit
y
Im
ag
e
Da
t
as
et
:
A
sta
nd
a
r
d
ben
c
h
m
ark
data
set
cal
le
d
AR
im
a
ge
dataset
is
use
d
f
or
evaluati
ng t
he pr
opos
e
d
syst
e
m
w
hich
co
nta
ins m
illi
on
s of
i
m
ages.
Camer
a
C
ap
tu
re Im
ag
e
Data
set
:
It c
onsist
s
of 10
00 im
age
s that are
take
n by cam
era that are c
ollec
te
d.
User
I
nterf
ace
Modul
e:
It
co
ll
ect
s
the
nece
ssary
data
fro
m
these
two
da
ta
set
and
it
al
so
decides
w
ha
t
data
hav
e
to
be
col
le
ct
ed
f
or
e
val
uating
t
he
pro
po
s
ed
syst
em
.
T
he
c
ollec
te
d
d
at
a
are
tra
nsfer
red
to
the
obj
ect
recog
niti
on
sy
stem
for
f
ur
th
er
processe
s
over
the
im
age
data
a
nd
rece
ived
the
r
eco
m
m
end
ed
res
ul
t
from
decisi
on m
anag
er
.
Ob
je
ct
Re
co
gni
ti
on
Syste
m:
I
t
con
sist
s
of
four
s
ub
-
syst
em
s
nam
ely
the
featur
e
e
xtracti
on,
featur
e
sel
e
ct
io
n
and
ob
j
ect
rec
ogniti
on
as
a
n
i
m
age
pr
e
-
proc
essing
subsyst
e
m
and
th
e
cl
assifi
cat
ion
s
ub
syst
e
m
wh
ic
h
con
ta
in
s
an
e
xisti
ng
cl
assifi
cat
ion
al
gorithm
cal
le
d
Co
nvol
ution
a
l
Ne
ur
al
Net
w
ork
(CN
N
).
I
n
t
he
pre
-
proc
essing
su
bsy
ste
m
,
i
t
perform
s
the
pre
-
processi
n
g
t
ask
by
us
in
g
a
new
ly
pr
opose
d
pre
-
proces
s
ing
al
go
rithm
cal
le
d
In
te
ll
igent
A
ge
nt
base
d
I
ncre
m
ental
Feat
ure
Sele
ct
io
n
Al
gorithm
(IAI
F
SA
)
for
ef
fecti
ve
pre
-
process
ing
o
f
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2089
-
4856
I
nt
J
R
ob
&
A
uto
m
,
Vo
l.
8
, No
.
3
,
Septem
ber 2
019
:
19
4
–
20
4
198
the
i
m
age
dat
aset
s.
I
n
ad
diti
on,
an
e
xisti
ng
cl
assifi
cat
io
n
al
gorithm
cal
le
d
Conv
olu
ti
on
al
Neural
N
et
work
al
gorithm
that
is
us
e
d
in
th
is
pap
e
r
for
pe
rfor
m
ing
ef
fe
ct
ive
cl
assifi
cat
ion
of
t
he
im
age
dataset
.
I
n
thi
s
cl
assifi
cat
ion
a
lgorit
hm
,
there
is
a
ge
netic
al
gorithm
pr
oces
s
w
hich
us
es
t
e
m
po
ral
c
onstr
ai
nts
an
d
te
m
po
ral
fu
zzy
ru
le
s
.
Figure
1. Syst
em
arch
it
ect
ur
e
Temp
oral
In
f
or
m
ation
Ma
nager:
It
is
use
fu
l
f
or
cl
ass
ific
at
ion
i
n
t
he
ob
j
ect
recog
niti
on
pr
ocess
of
t
his
pro
po
se
d
syst
e
m
. I
t al
so
provi
des
t
he necessa
ry d
et
ai
ls ab
ou
t t
i
m
e con
strai
nts in
the
pro
pose
d
syst
em
.
Fuzzy
Rul
e
B
as
e:
It
c
onta
in
s
the
necessa
ry
f
uzzy
r
ules
f
or
decisi
on
m
aking
ov
e
r
the
da
ta
set
us
i
ng
a
pa
rall
el
appr
oach f
or a
pp
ly
in
g gen
et
i
c operati
ons a
nd to
for
m
f
uzz
y t
e
m
po
ral r
ule
s.
Decisi
on
Man
ag
e
r:
This
de
ci
sion
m
anag
e
r
hel
ps
t
o
ta
ke
final
decisi
on
ov
e
r
th
e
im
age
data
w
hic
h
ar
e
trans
ferred
fro
m
ob
j
ect
rec
ogniti
on
syst
em
wit
h
t
he
hel
p
of
Tem
po
ra
l
I
nfor
m
at
ion
m
anag
er
a
nd
fu
zzy
ru
le
base.
4.
PROP
OSE
D WOR
K
This
sect
ion
e
xp
la
in
s
i
n
det
ai
l
ab
ou
t
the
pro
po
se
d
feat
ure
s
el
ect
ion
al
gorithm
cal
le
d
intel
li
gen
t
increm
ental
featur
e
s
el
ect
io
n
al
gorithm
fo
r
sel
ect
ing
the
m
os
t
con
trib
ut
ed
featur
e
s
tha
t
are
us
ed
to
c
la
ssify
the
im
age
data
set
s.
In
this
propose
d
feat
ur
e
sel
ect
ion
al
gorithm
,
it
extrac
ts
the
rele
van
t
featur
e
from
the
in
pu
t
i
m
ages,
re
co
gn
iz
e
the
im
ages
m
or
e
releva
nc
y
an
d
fina
li
ze
t
he
featu
res
acc
ordin
g
to
the
f
eat
ur
e
ext
racti
on
a
nd
recog
niti
on
res
ults.
Ne
xt,
an
existi
ng
cl
assif
ic
at
ion
al
gorit
hm
cal
le
d
Convo
l
ution
al
Ne
ural
Net
work
is
us
e
d
for
cl
assify
in
g t
he
im
ages w
it
h
the
h
el
p of f
uz
zy
r
ules t
hat a
re
gen
e
rated
by
the pr
opos
e
d
syst
e
m
.
4
.
1.
Fe
at
ure
s
el
ection
In
this
w
ork,
a
n
In
te
ll
ige
nt
A
gen
t
ba
sed
O
pt
i
m
a
l
Feat
ur
e
Sele
ct
ion
Algo
rithm
(I
A
OF
S
A)
ha
s
bee
n
pro
po
se
d
in
or
der
to
im
pr
ove
the
decisi
on
accu
racy
bas
ed
on
cl
assifi
c
at
ion
of
the
i
m
ages
obta
ine
d
from
the
virt
ual
real
it
y
env
iro
nm
ent.
This
pr
opos
e
d
feat
ur
e
sel
ec
ti
on
al
gorithm
has
be
en
de
velop
e
d
by
com
bin
in
g
the
opti
m
iz
at
i
on
r
ules
base
d
on
te
m
po
ral
const
raints
with
intel
li
gen
t
a
gen
ts
i
n
or
der
to
sel
ect
the
featur
e
s
wh
ic
h
can
re
duce
the
cl
assifi
cat
ion
ti
m
e
thr
ough
te
m
po
ral
an
al
ysi
s
of
t
he
dataset
.
T
his
al
gorithm
us
es
tim
e
const
raine
d
an
al
ysi
s
on
feat
ur
e
from
the
i
m
age
wh
ic
h
can
va
ry
over
tim
e
to
form
a
vid
e
o
seq
ue
nc
e
.
Ther
e
f
or
e,
at
each
tim
e
instants
t
he
inf
or
m
at
ion
gain
ra
ti
o
values
are
com
pu
te
d
us
i
ng
t
he
f
reque
nc
ie
s
of
occur
re
nce
of
the
featu
res
in
the
i
m
age
and
the
rate
at
whic
h
it
changes
with
res
pect
to
tim
e.
Her
e,
second
and
m
i
ll
ise
con
d
le
vel
te
m
po
r
al
analy
sis
of
i
m
age
featu
res
are
ca
rr
ie
d
out
in
orde
r
t
o
pro
vid
e
t
he
optim
a
l
pr
e
dicti
on
for
r
obotic
arm
m
o
vem
ent.
The
m
ai
n
a
dvanta
ge
of
the
pro
pose
d
featu
re
sel
ect
ion
al
go
rithm
i
s
that
it
red
uce
s
the
cl
assifi
cat
ion
ti
m
e
and
inc
reas
es
the
acc
ur
ac
y
of
nex
t
m
ove
m
ent
pr
e
dicti
on
t
hro
ugh
th
e
resu
lt
s
of
the
cl
assi
fiers.
This
al
gorithm
find
s
t
he
i
nfor
m
at
ion
ga
in
for
the
se
gme
nts
s
1,
s2
,
s
3,
…,
sn
obta
ine
d
from
the
im
age
at
t
i
m
es
t1,
t2,
.
.
.,
tn
on
the
im
age
I
.
T
he
e
quat
ion
us
e
d
t
o
fin
d
the
i
nform
at
ion
gain
va
lue
of
the im
age I
is
giv
e
n
in
(1
-
3
).
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
R
ob
&
A
uto
m
IS
S
N:
20
89
-
4856
Fuzz
y
n
e
uro
-
ge
netic
appro
ac
h
fo
r feat
ur
e
se
le
ct
ion
an
d
…
(
Rajen
dra T
hila
ha
r
C
.
)
199
(
)
=
−
∑
[
(
,
)
|
|
]
=
1
2
[
(
,
)
|
|
]
(1)
(
)
=
−
∑
[
|
|
|
|
]
=
1
×
(
)
(2)
(
)
=
[
(
)
−
(
)
(
)
+
(
)
]
×
(
)
(3)
Her
e
,
I
is
th
e
dataset
obta
ined
t
he
c
ollec
ti
on
of
fr
am
es
cal
le
d
V
1,
V
2,
…,
V
n.
Si’
s
represe
nt
the
se
gm
ents
obta
ined
f
r
om
t
he
im
age
I,
FR
is
the
set
feat
ures
ob
ta
i
ned
f
r
om
the
i
m
age
segm
ents
an
d
I
nfo
is
us
e
d
to
re
prese
nt
the
in
form
ation
gain
value
s
and
I
GR
re
presents
the
in
f
orm
ation
gai
n
r
at
io
ob
ta
i
ned
f
or
t
he
it
h
fr
am
e
of
th
e
vid
e
o
V.
I
n
t
his
m
od
el
,
a
se
t
of
f
ram
es
colle
ct
ed
from
a
vid
e
o
a
re
c
on
s
idere
d
as
t
he
i
m
ages
and
they
are
s
egm
ented
usi
ng
a
distance
ba
sed
a
ppr
oac
h
that
div
i
des
t
he
im
age
into
ei
gh
t
e
qual
se
gm
ents.
The det
ai
le
d
st
eps of t
he p
rop
os
e
d
I
A
OF
S
A are as
f
ollow
s:
Int
el
li
gent
Agent
b
as
e
d
O
pti
mal Feature
Sele
ction
A
l
gori
th
m
.
Inpu
t:
A vide
o
se
quence
V
Out
p
ut:
A
set
of ‘
m
’
featu
res
FS
c
ollec
te
d from
a set o
f ‘
n’
i
m
ages f
orm
ed
by the
fram
es o
f
V
.
St
ep
1:
Re
a
d
t
he
c
on
te
nts
of
vid
e
o V.
St
ep
2:
I
niti
al
iz
e the set
of f
e
at
ur
es
sel
ect
ed a
s NULL
b
y
se
tt
ing
FS = { }
St
ep
3:
I
niti
al
iz
e i =1
St
ep
4
:
Ca
ll
th
e
f
ram
e
form
ation
a
gen
t
to
d
ivide
the
v
ide
o i
nto
‘
n’ f
ram
es
a
nd
to
m
ake
s
egm
ents
of
e
qu
al
siz
e
on each
fram
e n
am
ed
as S
1,
S
2,
…,
Sn.
St
ep
5:
Ca
ll
th
e
com
pu
ta
ti
on
agen
t
t
o
c
om
pu
te
the
value
of
I
nfo
(I,
t)
fro
m
the
i
m
ages
usi
ng
t
he
form
ulae
for
the tim
e instan
ces t1 a
nd t2
r
e
sp
ect
ively
.
(
,
1
)
=
−
∑
[
(
,
1
,
)
|
|
]
=
1
2
[
(
,
1
,
)
|
|
]
and
(
,
2
)
=
−
∑
[
(
,
2
,
)
|
|
]
=
1
2
[
(
,
2
,
)
|
|
]
St
ep
6:
Ca
ll
the
com
pu
ta
ti
on
age
nt
for
cal
culat
ing
t
he
value
s
of
I
nfo
(F
R)
f
rom
the
i
m
ages
us
in
g
the form
ulae f
or the
par
ti
c
ular
ti
m
e instances t1 a
nd t2
r
es
pe
ct
ively
.
(
,
1
)
=
−
∑
[
|
,
1
|
|
|
]
=
1
×
(
,
1
)
and
(
,
2
)
=
−
∑
[
|
,
2
|
|
|
]
=
1
×
(
,
2
)
St
ep
7:
Ma
ke
the
com
pu
t
at
ion
a
gen
t
t
o
com
pu
te
the
value
of
I
nfor
m
at
ion
Gai
n
Ra
ti
o
value
us
in
g
the form
ulae.
(
,
1
)
=
[
(
,
1
)
−
(
,
1
)
(
,
1
)
+
(
,
1
)
]
×
(
,
1
)
and
(
,
2
)
=
[
(
,
2
)
−
(
,
2
)
(
,
2
)
+
(
,
2
)
]
×
(
,
2
)
St
ep
8:
Com
pu
te
the
valu
e
of
I
GRD
IFF
=
(
IG
R
(V
i,
t2
)
-
IG
R(
Vi,
t
1))
a
nd
com
par
e
it
with
th
res
ho
l
d
values
of f
eat
ures
FRi
call
ed
T
H1, T
H2 and
TH
3 us
ing
ste
p 7.
St
ep
9:
I
f (T
H
1
> FRi
) The
n
FS
=
FS U
(F
R
i, t1) U
(
FRi
,
t2)
Else
I
f
(IGR
(Vi, t1
)
>
TH
2
)
Th
e
n
FS =
FS U
(F
R
i, t1)
Else
I
f
(IGR
(Vi, t2
)
>
TH
3
)
Th
e
n
FS
=
FS U
(F
R
i, t2)
En
d
I
f
St
ep
10:
Set i
= i + 1
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2089
-
4856
I
nt
J
R
ob
&
A
uto
m
,
Vo
l.
8
, No
.
3
,
Septem
ber 2
019
:
19
4
–
20
4
200
St
ep
11:
I
f (i >
n) T
hen
Re
turn FS
Else
Go
t
o
ste
p 5.
4
.
2.
Ne
uro
-
G
enetic w
ith
Fu
zz
y
Ru
le
b
as
e
d C
l
as
sifie
r
(
NGF
RC)
In
this
wor
k,
a
new
al
go
rith
m
cal
le
d
Neu
r
o
-
Ge
netic
Cl
assifi
cat
ion
al
go
rithm
with
fu
z
zy
ru
le
s
has
been
pro
posed
.
In
this
m
od
el
,
deep
ne
ur
al
ne
tworks
a
re
use
d
with
one
input
la
ye
r,
two
hi
dd
e
n
la
ye
rs
a
nd
one
ou
t
pu
t
la
ye
r.
The
sigm
oid
al
f
unct
ion
is
use
d
as
t
he
act
iv
at
ion
f
un
ct
io
n
for
t
he
ne
ural
netw
ork.
Mo
re
ov
e
r,
gen
et
ic
al
gorithm
s
are
us
ed
f
or
w
ei
ght
a
dju
stm
ent
in
wh
ic
h
fu
z
zy
log
ic
is
al
s
o
us
e
d
with
t
riangula
r
m
e
m
ber
sh
ip
f
un
ct
io
n
f
or
f
orm
ing
r
ules
w
hich
are
use
d
in
the
decisi
on
m
aking
pr
ocess.
T
he
ste
ps
of
th
e pr
opos
e
d
cl
assifi
cat
ion
al
gorithm
are
as
f
ollows:
Algori
th
m
:
F
uzzy
Neuro
-
Gen
et
ic
base
d cl
assifi
cat
ion
a
lgorit
hm
.
Inpu
t
:
Ro
bot
m
ov
em
ent search
s
pa
ce data
Out
p
ut
:
Set
of
f
uzzy
Rules an
d
t
he
c
la
ssifie
d
re
su
lt
s.
1.
In
it
ia
li
ze the se
arch space
to fi
rst lo
cat
io
n of t
he
a
rea.
2.
Set t
he
c
onstra
ints f
or roboti
c arm
m
ov
em
en
t usin
g fu
zzy
r
ules.
3.
Re
ad
the
searc
h
s
pace
detai
ls
for op
ti
m
al
m
o
vem
ent.
4.
Fo
rm
the
neura
l netw
ork by
prov
i
ding the
w
e
igh
ts a
nd the
varia
ble
values
.
5.
Apply t
he
act
i
vation f
unct
ion by cal
li
ng the
s
igm
oid
al
ev
al
uation f
unct
ion.
6.
Apply bias
func
ti
on
on the
ou
tpu
t
of n
e
ur
al
net
work.
7.
Com
pu
te
r
the
error rate
a
nd perf
or
m
d
eep
learni
ng
.
8.
Apply f
uzzy r
ul
es an
d per
for
m
the w
ei
ght a
dju
stm
ent
.
9.
Ca
ll
the g
e
netic
searc
h
al
go
rithm
f
or
fin
ding
the opti
m
al
p
ath
us
in
g
t
he Ne
ur
al
netw
ork o
utput.
10.
IF
pat
h
is
bette
r
tha
n pr
e
viou
s p
at
h
T
HE
N
Set B
est
_Move
= Cu
rr
e
nt_
M
ove
Else
Chan
ge direct
i
on of m
ov
em
e
nt
11.
IF
Best
_Move
>= Opti
m
a
l va
lue T
HEN
Re
turn
Else
Go
t
o
ste
p 8.
This
al
gorithm
co
ns
ide
rs
the
m
ov
e
in
ei
gh
t
directi
ons
nam
el
y
fr
ont,
back,
le
ft,
ri
gh
t
a
nd
the
f
ou
r
diag
on
al
direct
ion
s
f
or
e
ff
ect
i
ve
m
ov
em
ent
of
the
rob
otic
a
rm
.
The
intel
li
gen
t
cl
assifi
cat
ion
al
gorithm
a
pp
li
es
fu
zzy
ru
le
s
a
nd
gen
et
ic
base
d
op
ti
m
iz
a
ti
on
te
chn
iq
ues
in
t
he
neural
ne
tw
ork
in
or
der
t
o
fin
d
the
m
os
t
optim
al
m
ov
e
m
ent in th
e
prom
isi
ng
di
recti
on
s.
5.
RESU
LT
S
AND DI
SCUS
S
ION
The
propose
d
m
od
el
has
be
en
pr
opos
e
d
and
im
ple
m
ented
f
or
pr
e
dicti
ng
the
au
gm
e
nted
reali
ty
i
m
ages
accu
rat
el
y
and
it
is
pr
ov
i
ded
s
uffici
ent
data.
Here,
we
ha
ve
m
od
el
ed
a
new
Ro
bo
t
an
d
al
s
o
ha
ve
bee
n
colle
ct
ed
50
i
m
ages
by
c
onduct
ing
va
rio
us
ex
pe
rim
ents.
The
c
ollec
te
d
im
ages
a
re
c
on
si
de
red
our
own
da
ta
set
in this
w
ork
as
input dataset
.
5
.
1
.
Ex
peri
ment
al results
F
igure
2
s
how
s
ex
pe
rim
ental
set
up
of
t
he
ne
w
m
od
el
le
d
r
obot
by
c
ondu
ct
ing
va
rio
us
e
xp
e
rim
ents
.
Table
1
sho
ws
the
nu
m
ber
of
featur
e
s
e
xtrac
te
d
us
in
g
existi
ng
al
go
rithm
and
the
pro
po
se
d
al
gorithm
.
T
able
2
sh
ows
five
im
portant
feat
ur
e
s
that
are
sel
e
ct
ed
by
t
he
pr
opos
e
d
I
ntell
igent
Ag
e
nt
ba
sed
O
ptim
a
l
F
eat
ur
e
Sele
ct
ion
Al
gorithm
(I
AO
F
S
A)
.
From
T
able
2,
it
li
ste
d
out
the
i
m
po
rtant
featur
e
s
su
c
h
as
ang
le
,
tim
e,
sh
a
pe,
sta
rt
posit
io
n(x,
y)
an
d
the
end
posit
ion
(
x,
y).
T
hese
a
ll
featu
res
are
use
d
to
im
pr
ov
e
t
he
cl
assi
f
ic
at
ion
accuracy
w
hen test
the ro
bo
t
im
ages.
Figure
3
s
how
s
the
cl
assi
ficat
ion
acc
ur
acy
of
the
pro
pose
d
al
go
rithm
and
the
e
xisti
ng
al
gorithm
s
wh
e
n
us
es
al
l
the
feat
ur
es
of
the
rob
ot
im
a
ges.
He
re,
we
hav
e
c
on
si
der
e
d
12
feat
ur
es
f
or
m
aking
dec
isi
on
ov
e
r
the
cl
assif
ie
rs
su
c
h
as
S
VM,
D
T,
Naïv
e
Ba
ye
s,
Ra
ndom
Fo
rest
an
d
the
pro
posed
NGFRC
.
From
F
igur
e
3,
it
can
be
see
n
that
the
pe
rfor
m
ance
of
t
he
propose
d
al
go
rithm
is
per
f
orm
ed
well
with
the
us
es
of
al
l
the
1
2
featur
e
s
of
r
ob
ot
i
m
ages
w
he
n
it
is
com
par
ed
with
t
he
ex
ist
ing
cl
assifi
c
at
ion
al
gorith
m
s
su
ch
as
S
V
M,
DT,
Naïve
Ba
ye
s,
Ra
ndom
Fo
rest
and the
pr
opose
d NGFR
C.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
R
ob
&
A
uto
m
IS
S
N:
20
89
-
4856
Fuzz
y
n
e
uro
-
ge
netic
appro
ac
h
fo
r feat
ur
e
se
le
ct
ion
an
d
…
(
Rajen
dra T
hila
ha
r
C
.
)
201
Figure
2.
Ex
pe
rim
ental
setup
Table
1.
List
of f
eat
ur
es
ex
t
ra
ct
ed
S.No.
Na
m
e
of
the f
eatu
re
1
Distan
ce
2
An
g
le
3
Mar
k
ers
4
Heig
h
t
5
W
id
th
6
Co
n
trast
7
Co
lo
r
8
Textu
re
9
Sh
ap
e
10
p
ix
el in th
e
d
ep
th
im
a
g
e
11
p
ix
el in th
e ce
n
ter
i
m
ag
e
Table
2.
List
of sele
ct
ed feat
ures
S.No.
Na
m
e
of
the f
eatu
re
1
An
g
le
2
Mar
k
ers
4
Sh
ap
e
5
Start
p
o
sitio
n
(x,y
)
6
End
po
sitio
n
(
x
,y
)
Figure
3. Cl
assifi
cat
ion
acc
uracy
w
it
h fu
ll
fe
at
ur
es
Figure
4
s
how
s
the
cl
assi
ficat
ion
acc
ur
acy
of
the
pro
pose
d
al
go
rithm
and
the
e
xisti
ng
al
gorithm
s
wh
e
n
us
es
al
l
t
he
feat
ur
es
of
t
he
r
obot
im
ages.
He
re,
we
ha
ve
c
on
si
der
e
d
on
ly
6
feat
ur
es
that
are
sel
ect
ed
by
the
pr
opos
e
d
f
eat
ur
e
sel
ect
io
n
al
go
rithm
fo
r
m
aking
decis
ion
ove
r
t
he
cl
assifi
ers
s
uc
h
as
SV
M
,
D
T,
Naïv
e
Ba
ye
s,
Ra
nd
om
Fo
rest
a
nd
t
he
pro
pose
d
N
GF
RC
.
F
ro
m
F
igure
4,
it
ca
n
be
see
n
that
the
perf
or
m
ance
of
the
pro
po
se
d
al
go
rithm
per
f
or
m
s
well
with
the
us
e
of
the
sel
ect
ed
feat
ur
es
f
r
om
ro
bot
im
ages
w
he
n
it
is
com
par
ed
with
the
e
xisti
ng
cl
assifi
cat
ion
al
gorithm
s
su
c
h
a
s
S
VM,
DT,
N
aï
ve
Ba
ye
s,
Ra
ndom
Forest
a
nd
t
he
pro
po
se
d N
GFR
C.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2089
-
4856
I
nt
J
R
ob
&
A
uto
m
,
Vo
l.
8
, No
.
3
,
Septem
ber 2
019
:
19
4
–
20
4
202
Figure
4. Cl
assifi
cat
ion
acc
uracy
w
it
h
sel
ect
ed feat
ur
e
s
Figure
5
s
how
s
the
c
om
par
at
ive
a
naly
sis
ac
cordin
g
to
the
cl
assifi
cat
io
n
accuracy
o
f
th
e
pr
opos
e
d
al
gorithm
by
cond
ucting
var
i
ou
s
e
xp
e
rim
ents
su
c
h
a
s
E
1,
E2,
E
3,
E4
a
nd
E
5.
Her
e
,
we
ha
ve
co
ns
ide
r
ed
fu
ll
featur
e
s
[15]
a
nd
the
sel
ect
ed
[5]
featu
res
for
perf
or
m
ing
cl
assifi
cat
ion
process
.
F
ro
m
F
ig
ur
e
5,
it
c
an
be
ob
s
er
ved
that
the
pe
rfor
m
anc
e
of
the
pro
po
sed
al
gorithm
in
al
l
t
he
e
xper
i
m
ents
perf
orm
bette
r
well
wh
e
n
it
us
es
the
sel
ect
ed
featu
res
th
an
w
he
n
it
use
s
the
f
ull
fea
tures
of
t
he
r
obot
im
ages
t
hat
are
us
es
i
n
th
e
exp
e
rim
ents.
This
is
du
e
to
t
he
us
e
of
ef
fecti
ve
a
nd
m
or
e
con
t
rib
uted
fea
tures
w
hile
m
a
king
decisi
on
i
n
the
cl
assifi
cat
ion
proces
s.
M
or
e
over
,
the
pro
po
s
ed
cl
assi
fier
use
s
the
ne
uro
-
ge
netic
al
gorith
m
and
f
uzzy
r
ul
es
f
or
the d
eci
si
on m
akin
g process
in
the
pr
opos
e
d cl
assifi
er.
Figure
5. Com
par
at
ive
Analy
sis betwee
n Fu
ll
featur
es
and
Sele
ct
ed
Feat
ures
6.
CONCL
US
I
O
N AND F
UT
U
RE W
ORKS
In
t
his
pap
e
r,
a
ne
w
a
ppro
a
c
h
f
or
im
ple
m
e
nting
an
A
ug
m
ented
Re
al
it
y
syst
e
m
by
app
ly
in
g
fu
zzy
gen
et
ic
ne
ur
al
netw
orks
ha
s
be
en
pro
posed
a
nd
im
ple
m
ente
d
f
or
ef
fecti
ve
i
m
age
cl
assifi
cat
ion
a
nd
predi
ct
ing
the
arm
m
ov
em
ent.
It
c
onsi
sts
of
t
wo
c
om
po
nen
ts
nam
el
y
featur
e
sel
ect
ion
a
nd
cl
a
ssific
at
ion
m
od
ules
.
The
pr
opos
e
d
m
od
el
us
es
a
fu
zzy
lo
gic
based
i
ncr
em
ental
featur
e
s
el
ect
ion
f
or
e
xtracti
ng
the
r
el
evan
t
featur
e
s
that
a
re
us
e
d
t
o
rec
ognize
the
im
port
ant
f
eat
ur
e
s
from
3D
im
ages.
M
or
e
ove
r,
this
pa
per
e
xp
la
in
s
the
im
ple
m
ent
at
ion
a
nd
resul
ts
of
the
pro
pose
d
al
go
rithm
s
f
or
an
Aug
m
ented
Re
al
it
y
syst
e
m
us
in
g
im
age
recog
niti
on
,
f
e
at
ur
e
extr
act
io
n,
feat
ur
e
sel
e
ct
ion
a
nd
cl
as
sific
at
ion
by
consi
der
i
ng
th
e
gl
ob
a
l
an
d
local
featur
e
s
of
th
e
im
ages.
For
this
pur
pose,
we
pr
opos
e
a
thre
e
la
ye
r
fu
zzy
ne
ur
al
netw
ork
t
hat
has
bee
n
i
m
ple
m
ented
base
d
on
wei
gh
t
a
dju
stm
ents
us
in
g
f
uzzy
ru
le
s
in
the
conv
olu
ti
onal
neural
net
wor
ks
with
gen
et
ic
al
gorithm
fo
r
ef
fecti
ve
op
ti
m
iz
ati
on
of
r
ules.
Th
e
cl
assifi
cat
io
n
al
gorithm
is
al
so
base
d
on
fu
zz
y
neur
o
-
gen
et
ic
appr
oach
wh
ic
h
c
onsist
s
of
two
phases
na
m
el
y
Trainin
g
phase
a
nd
te
sti
ng
phase.
D
ur
i
ng
the
trai
ni
ng
ph
ase,
r
ules
a
re
f
or
m
ed
base
d
on
obj
ect
s
a
nd
these
ru
le
s
a
re
app
li
ed
du
rin
g
t
he
te
sti
ng
ph
ase
f
or
recog
nizing
th
e
obje
ct
s
w
hic
h
ca
n
be
us
e
d
i
n
rob
otics
f
or
eff
ect
ive
ob
j
ec
t
recog
niti
on
.
Fr
om
the
e
xpe
rim
ent
s
cond
ucted
i
n
t
his wo
rk, it i
s
pro
ved that t
he
prop
os
ed
m
od
el
is m
or
e accu
rate i
n 3
D obje
ct
r
eco
gnit
ion
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
R
ob
&
A
uto
m
IS
S
N:
20
89
-
4856
Fuzz
y
n
e
uro
-
ge
netic
appro
ac
h
fo
r feat
ur
e
se
le
ct
ion
an
d
…
(
Rajen
dra T
hila
ha
r
C
.
)
203
REFERE
NCE
S
[1]
Sannasi
Gan
apat
h
y
,
K
ana
gasa
b
ai
Kulothung
an,
S
anna
s
y
Muthur
ajkum
ar,
Muthusa
m
y
Vi
jay
a
la
ksh
m
i,
Pal
anicha
m
y
Yogesh
And
Arputhar
a
j
Kann
a
n
“
Inte
lligent
Feat
ur
e
Sel
ection
And
Cla
ss
ifica
ti
on
T
ec
hn
ique
s
For
Intrusion
Dete
c
ti
on
In
Networks:
A
Surve
y
,
”
Journ
al
On
Wireless
Co
mm
unic
ations
And
Net
w
orking
,
Vol.
2
71,
pp.
1
-
11,
2013.
[2]
Azum
a,
R.
,
Ba
illot,
Y
.
,
B
ehr
ing
er,
R
.
,
Fe
ine
r
,
S
.
,
Jul
ie
r
,
S.
And
Mac
inty
r
e,
B.
:
“
Rec
ent
Advan
c
es
in
Augm
ent
e
d
Rea
lit
y
,
”
I
EE
E
Computer
Gr
aphic
s and Applica
ti
ons
,
Vol
.
21,
No.6, pp. 34
–
47,
2
001.
[3]
Mark
Bil
l
inghur
st,
Adria
n
C
la
rk
,
And
Gun
Lee
“
A
Survey
of
A
ugm
ent
ed
Re
al
i
t
y
,
”
Foundat
ions
and
Tr
ends
in
Hum
an
-
Comput
er
Inte
rac
ti
on
,
V
ol.
8
,
No
.
2
-
3
,
P
P.
73
–
272,
2014.
[4]
Rafa
e
l
Bastos
and
Migue
l
Sa
le
s
Dias,
“
Autom
at
ic
C
amera
Pos
e
Initialization,
Us
ing
Sc
al
e
,
Rot
at
ion
a
nd
Luminanc
e
Inva
ria
nt
Natur
al
Fe
a
ture
Tracki
ng
,
”
J
ournal
of
Wscg
,
PP
.
97
-
104.
[5]
Ulric
h
Neum
an
n
and
Su
y
a
Yo
u,
“
Natur
al
Fea
ture
Tr
ac
k
ing
f
or
Augm
ent
ed
Rea
lit
y
,
”
IE
EE
Tr
ansacti
ons
on
Mult
imedi
a
,
Vol
.
1
,
No.
1,
PP
.
5
3
-
64,
1999
.
[6]
M.
L.
Yu
an,
S.
K.
Ong,
a
nd
A.
Y
.
C.
Ne
e,
“
Reg
istra
ti
on
using
N
at
ura
l
Fea
ture
s
f
or
Augm
ent
ed
R
ea
l
ity
S
y
s
te
m
s
,
”
IEE
E
Tr
ansacti
o
ns on
Vi
sual
izati
on
and
Compute
r Gr
aphic
s
,
Vol.
12,
No.
4
,
Jul
y
/
Augus
t
2006.
[7]
Akash
Bapa
t,
En
rique
Dunn,
Jan
-
Micha
e
l
Frahm
,
“
Towa
rds
Kilo
-
Hert
z
6
-
Dof
Visual
Tr
ac
king
usi
ng
an
Egocent
ri
c
Cluste
r
of
Rol
li
n
g
Shutte
r
Camer
as
,
”
IE
EE
Tr
ansacti
ons on
V
isuali
zation
and
Co
mputer
Gr
aphic
s
,
Vol.
22
,
No.
11
,
Novem
ber
2016.
[8]
Gholam
Alimontazer,
Dav
arg
iv
eki
,
"A
n
Im
prove
d
Radial
Bas
is
Functi
on
Ne
ura
l
Network
f
or
Objec
t
Im
ag
e
Ret
ri
eva
l
,
"
Neur
ocomputi
ng
,
Vol
.
168
,
pp
.
221
–
2
33,
2015
.
[9]
Moham
ed
Boume
hra
z
,
Z
ine
b
Ha
bba,
Ra
fia
H
assani,
“
Vision
Base
d
Tra
ck
ing
and
I
nte
rc
ept
ion
o
f
Moving
Ta
rg
et
b
y
Mobile
Robot
using
Fuzz
y
C
ontrol
,
”
Journal
of
App
li
ed
En
gine
ering
S
cien
ce
&
Technol
ogy
,
Vol
.
4,
No.
2
,
pp.
159
-
165
,
20
18.
[10]
Niti
n
Sharm
aa,
Om
Praka
sh
Verm
a,
“
A
Fuzz
y
Edge
b
ase
d
M
ult
ifoc
us
Fus
ion
,
”
Proce
dia
Co
mputer
Scienc
e
,
Vol.
132
,
pp
.
12
98
–
1307,
2018
.
[11]
Ioa
nnis
Sara
kog
lou,
Nikos
G.
T
saga
rak
is
"A
High
Perform
anc
e
Ta
c
ti
l
e
Feedb
a
ck
Displa
y
and
it
s
Inte
gr
at
ion
i
n
Te
l
eope
ra
ti
on
,
"
I
EE
E
Tr
ansacti
o
ns on
Haptic
s
,
V
ol.
5
,
No
.
3
,
Jul
y
-
Septe
m
ber
201
2.
[12]
S
Gana
pat
h
y
,
P
Vijay
akumar
,
P
Yogesh,
A
Kanna
n,
“
An
Inte
l
l
ige
nt
CRF
Base
d
Feat
ure
Se
lect
ion
for
Eff
ec
t
iv
e
Intrusion
Det
ection
,
”
Int
ernational
Arab
Jou
rnal
of
I
n
formation
Te
chnol
o
gy
(
IAJ
IT)
,
Vol.
13
,
No
.
1
,
pp.
64
-
74
,
2016
.
[13]
U.
Kanimozhi
,
S.
Gana
p
at
h
y
,
D
.
Man
jul
a
,
A
.
K
anna
n,
"A
n
In
te
l
li
gent
Risk
Pred
ic
ti
on
S
y
s
te
m
fo
r
Bre
ast
C
anc
er
Us
ing
Fuzz
y
Te
m
pora
l
Rule
s
,
"
Nat
io
nal
Aca
d
em
y
Scie
nc
e
L
et
t
ers,
Avai
lable
in
Onl
i
ne:
htt
ps://
do
i.
org/1
0.
1007/s40009
-
0
18
-
0732
-
0,
2018
.
[14]
M
Selvi
,
R
Log
ambigai
,
S
Gan
apa
th
y
,
L
S
Ra
m
esh,
HK
Nehe
m
ia
h,
A.
Kanna
n,
"
Fuzzy
Tempor
al
Approach
f
or
Ene
rgy
E
ff
i
cient
Rout
ing
in
WS
N
,
"
Proce
ed
ings
of
The
In
te
rn
ational
Conf
ere
n
c
e
on
Inform
at
i
c
s
and
Anal
y
tic
s,
Vol.
13
,
pp
.
117
-
123,
2016
.
[15]
R
Sethukka
r
asi,
S Gana
pat
h
y
,
P
Yogesh,
A Kann
an,
"A
n
int
e
ll
ig
e
nt
neur
o
fuz
z
y
t
e
m
pora
l
knowled
ge
rep
r
ese
nt
at
io
n
m
odel
for
m
ini
n
g
te
m
pora
l
pa
tt
e
rns
,
"
Journal
o
f Inte
lligen
t
&
Fu
zz
y
Syste
ms
,
Vol
.
26
,
No.
3,
pp.
1
167
-
1178,
2014
.
[16]
N
Jaisanka
r
,
S
Gana
pat
h
y
,
P
Y
ogesh,
A
Kann
a
n,
K
Anand
,
"A
n
in
te
l
li
gen
t
agent
b
ase
d
int
rusio
n
de
tecti
on
s
y
s
tem
using f
uzzy
roug
h
set
b
ase
d
ou
tl
i
er
de
tecti
on
,
"
So
ft
Computing
Te
chni
ques
in Vi
si
on
Scienc
e
,
pp.
1
47
-
153,
2012
.
[17]
L
S
Ramesh,
S
Gana
pat
h
y
,
R
B
huvan
eshwari
,
K
Kulothunga
n,
V.
Pandi
y
a
raj
u
,
A.
Kanna
n
,
“
Prediction
of
use
r
int
er
ests
for
pro
vidi
ng
r
el
ev
ant
informati
on
usin
g
releva
nc
e
f
eedbac
k
and
r
e
-
ra
nking
,
"
Int
ernat
ional
Journal
of
Inte
lligen
t
Infor
mation
Techno
lo
gie
s (
IJI
IT)
,
Vol.
11,
No.4, pp. 2,
2015.
[18]
W
ei
Jie
W
ang.
Hua
Gen
W
an,
"Rea
l
-
t
ime
ca
m
era
tr
ac
k
ing
usi
ng
h
y
br
id
fe
at
ur
es
in
m
obil
e
au
gm
ent
ed
re
al
i
t
y
,
"
Information
Sc
ience
,
Vol.
58,
No
.
11
,
pp
.
1
-
13
,
2
015.
[19]
Ta
hrir
i,
F
.
Mous
avi
,
M.
Yap
H.
J,
Siti
Za
wi
ah,
M.
D.
&
T
aha,
Z.
,
"O
pti
m
iz
ing
the
Robot
Arm
Movem
ent
Ti
m
e
using
Virtua
l
R
ea
l
ity
Robot
ic
Te
a
chi
ng
S
y
s
tem
,
”
Inte
rnat
ion
al
Journal
of
S
imulat
ion
Mod
e
l
,
Vol
.
14,
No
.
1,
pp.
28
-
38
,
2015
.
[20]
Surende
r
Kum
ar,
Kavi
ta
R
ani
,
V.
K.
Bang
a,
“
Roboti
c
Arm
Movem
ent
Optimi
za
t
ion
Us
ing
Soft
Com
puti
ng
,
”
Inte
rnational
Jo
u
rnal
of Robot
i
c
s and
Aut
omat
io
n
(
IJR
A)
,
Vol.
16
,
No.1
,
pp
.
1
-
14
,
2017.
[21]
Du
sko
Kati
c
,
Miom
ir
Vukobrat
o
vic
,
“
Gene
t
ic
Al
gorit
hm
s
in
Rob
oti
cs
,
”
Inte
rnat
i
onal
S
erie
s
on
Mic
ropr
oce
ss
or
-
Based
and
I
ntell
ige
nt
Syste
ms
En
gine
ering
,
Volu
m
e
25,
2003
.
[22]
V.
K.
Bang
a,
Ja
sjit
K
aur
,
R.
Ku
m
ar,
Y.
Singh
,
“
Modeli
ng
and
Si
m
ula
ti
on
of
Rob
oti
cs
Arm
Movem
ent
using
Soft
Com
puti
ng”,
W
orld
Aca
d
em
y
of
Scie
n
ce,
75
,
pp
.
614
-
620,
2011.
[23]
L.
A.
Z
ade
h
,
“
Fuzz
y
se
ts
,”
Infor
mation
and
Con
t
rol
,
8,
pp.
338
-
3
53,
1965
.
[24]
Ta
das
Ba
lt
ruša
itis
,
Chaitan
y
a
Ahuja
,
Loui
s
-
Phil
ippe
Moren
c
y
“
Multi
m
odal
Ma
chi
ne
Learni
ng:
A
Surve
y
an
d
Ta
xonom
y
,
”
IE
EE
Tr
ansacti
o
ns
on
Pattern
Analysis
and
Mac
hin
e
In
telli
gen
ce
,
Volum
e:
41
,
Iss
ue:
2
,
pp.
423
–
443
,
2
019.
[25]
Shuiwang
Ji
,
W
ei
Xu
,
Ming
Ya
ng
,
Ka
i
Yu
,
“
3D
Convolut
iona
l
Neura
l
N
et
work
s
for
Hum
an
Ac
ti
on
R
ec
ogni
ti
on
,
”
IEE
E
Tr
ansacti
o
ns on
Pattern
An
aly
sis and
Ma
ch
ine
In
te
l
li
gen
ce
,
Volum
e:
35
,
Iss
ue:
1
,
pp
.
221
–
231,
2013
.
[26]
Yos
hua
Bengi
o
,
Aaron
Courvill
e
,
Pasc
al
Vin
ce
n
t
,
“
Repre
sent
ati
on
L
ea
rn
ing:
A
Rev
ie
w
and
N
ew
Perspe
ct
iv
es
,
”
IEE
E
Tr
ansactions
on
Pat
te
rn
Anal
ysis
and
Mac
hine
In
telligence
,
Volum
e:
35
,
Iss
ue:
8
,
pp.
1798
–
182
8
,
Aug.
2013.
Evaluation Warning : The document was created with Spire.PDF for Python.