T
E
L
KO
M
N
I
KA
T
e
lec
om
m
u
n
icat
ion
,
Com
p
u
t
i
n
g,
E
lec
t
r
on
ics
an
d
Cont
r
ol
Vol.
1
8
,
No.
3
,
J
une
2020
,
pp.
1499
~
1504
I
S
S
N:
1693
-
6930,
a
c
c
r
e
dit
e
d
F
ir
s
t
G
r
a
de
by
Ke
me
nr
is
tekdikti
,
De
c
r
e
e
No:
21/E
/KP
T
/2018
DO
I
:
10.
12928/
T
E
L
KO
M
NI
KA
.
v1
8
i
3
.
14777
1499
Jou
r
n
al
h
omepage
:
ht
tp:
//
jour
nal.
uad
.
ac
.
id/
index
.
php/T
E
L
K
OM
N
I
K
A
Id
e
n
t
ifie
r
of
h
u
m
a
n
e
m
ot
io
n
s
b
ase
d
o
n
c
on
vol
u
t
io
n
al
n
e
u
r
al
n
e
t
w
o
r
k
f
or
a
ssi
st
an
t
r
ob
o
t
F
r
e
d
y
M
ar
t
ín
e
z
,
Cé
s
ar
Her
n
án
d
e
z
,
Angélica
Re
n
d
ón
Facu
l
t
ad
T
ec
n
o
l
ó
g
i
ca,
U
n
i
v
ers
i
d
a
d
D
i
s
t
ri
t
al
Fra
n
ci
s
co
J
o
s
é
d
e
Ca
l
d
a
s
,
Co
l
o
mb
i
a
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
S
e
p
2
,
2019
R
e
vis
e
d
J
a
n
4
,
2020
Ac
c
e
pted
F
e
b
23
,
2020
T
h
i
s
p
ap
er
p
r
o
p
o
s
e
s
a
s
o
l
u
t
i
o
n
f
o
r
t
h
e
p
ro
b
l
em
o
f
c
o
n
t
i
n
u
o
u
s
p
re
d
i
c
t
i
o
n
i
n
real
-
t
i
me
o
f
t
h
e
emo
t
i
o
n
al
s
t
a
t
e
o
f
a
h
u
man
u
s
er
fro
m
t
h
e
i
d
en
t
i
f
i
cat
i
o
n
o
f
ch
aract
er
i
s
t
i
c
s
i
n
faci
a
l
ex
p
res
s
i
o
n
s
.
In
ro
b
o
t
s
w
h
o
s
e
m
ai
n
t
a
s
k
i
s
t
h
e
care
o
f
p
eo
p
l
e
(ch
i
l
d
ren
,
s
i
c
k
o
r
el
d
er
l
y
p
e
o
p
l
e)
i
s
i
m
p
o
r
t
an
t
t
o
mai
n
t
ai
n
a
cl
o
s
e
rel
at
i
o
n
s
h
i
p
man
-
mach
i
n
e,
an
l
d
a
rap
i
d
res
p
o
n
s
e
o
f
t
h
e
ro
b
o
t
t
o
t
h
e
act
i
o
n
s
o
f
t
h
e
p
ers
o
n
u
n
d
er
care.
W
e
p
ro
p
o
s
e
t
o
i
n
crea
s
e
t
h
e
l
ev
el
o
f
i
n
t
i
mac
y
o
f
t
h
e
ro
b
o
t
,
an
d
i
t
s
re
s
p
o
n
s
e
t
o
s
p
ec
i
fi
c
s
i
t
u
at
i
o
n
s
o
f
t
h
e
u
s
er,
i
d
en
t
i
f
y
i
n
g
i
n
real
t
i
me
t
h
e
emo
t
i
o
n
refl
ect
e
d
b
y
t
h
e
p
ers
o
n
'
s
face.
T
h
i
s
s
o
l
u
t
i
o
n
i
s
i
n
t
e
g
ra
t
ed
w
i
t
h
al
g
o
r
i
t
h
ms
o
f
t
h
e
res
earch
g
r
o
u
p
rel
at
e
d
t
o
t
h
e
t
rac
k
i
n
g
o
f
p
e
o
p
l
e
fo
r
u
s
e
o
n
a
n
as
s
i
s
t
a
n
t
r
o
b
o
t
.
T
h
e
s
t
rat
eg
y
u
s
ed
i
n
v
o
l
v
e
s
t
w
o
s
t
a
g
es
o
f
p
r
o
ces
s
i
n
g
,
t
h
e
fi
r
s
t
i
n
v
o
l
v
e
s
t
h
e
d
e
t
ect
i
o
n
o
f
face
s
u
s
i
n
g
H
O
G
an
d
l
i
n
ear
SV
M,
w
h
i
l
e
t
h
e
s
eco
n
d
i
d
en
t
i
f
i
es
t
h
e
emo
t
i
o
n
i
n
t
h
e
face
u
s
i
n
g
a
CN
N
.
T
h
e
s
t
ra
t
eg
y
w
as
c
o
mp
l
et
e
l
y
t
es
t
ed
i
n
t
h
e
l
ab
o
rat
o
ry
o
n
o
u
r
ro
b
o
t
i
c
p
l
at
f
o
rm,
d
emo
n
s
t
rat
i
n
g
h
i
g
h
p
erf
o
rman
ce
w
i
t
h
l
o
w
res
o
u
rce
c
o
n
s
u
m
p
t
i
o
n
.
T
h
r
o
u
g
h
v
ari
o
u
s
co
n
t
ro
l
l
e
d
l
ab
o
rat
o
ry
t
e
s
t
s
w
i
t
h
d
i
fferen
t
p
eo
p
l
e,
w
h
i
c
h
fo
rced
a
cert
ai
n
emo
t
i
o
n
o
n
t
h
ei
r
faces
,
t
h
e
s
ch
eme
w
a
s
ab
l
e
t
o
i
d
e
n
t
i
fy
t
h
e
emo
t
i
o
n
s
w
i
t
h
a
s
u
cces
s
rat
e
o
f
9
2
%
.
K
e
y
w
o
r
d
s
:
A
utonom
ous
r
obot
C
onvolut
ional
ne
ur
ona
l
ne
twor
k
H
uman
e
mot
ions
S
e
r
vice
r
obot
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e
.
C
or
r
e
s
pon
din
g
A
u
th
or
:
F
r
e
dy
M
a
r
tí
ne
z
,
F
a
c
ult
a
d
T
e
c
nológi
c
a
,
Unive
r
s
idad
Dis
tr
it
a
l
F
r
a
nc
is
c
o
J
os
é
de
C
a
ldas
,
B
ogotá
,
C
olom
bia
.
E
mail:
f
hmar
ti
ne
z
s
@udis
tr
it
a
l.
e
du.
c
o
1.
I
NT
RODU
C
T
I
ON
S
e
r
vice
r
obots
a
r
e
de
s
igned
to
de
ve
lop
tas
ks
that
s
uppor
t
the
human
be
ing
[
1
-
3]
.
T
he
s
uc
c
e
s
s
in
the
de
ve
lopm
e
nt
of
thes
e
tas
ks
is
int
im
a
tely
li
nk
e
d
to
their
a
bil
it
y
to
int
e
r
a
c
t
with
the
us
e
r
[
4
-
6]
.
M
os
t
of
the
ke
y
to
human
int
e
r
a
c
ti
on
li
e
s
in
the
a
bil
it
y
of
o
ne
pe
r
s
on
to
f
e
e
l
s
ympathy
f
o
r
the
other
,
that
is
,
th
e
a
bil
it
y
to
identi
f
y
with
a
nother
pe
r
s
on,
know
how
to
li
s
te
n,
unde
r
s
tand
their
pr
oblems
a
nd
e
mot
ions
[
7,
8
]
.
M
uc
h
of
thi
s
e
mpathy
is
vis
ua
li
z
e
d
by
obs
e
r
ving
a
pe
r
s
on’
s
f
a
c
e
.
T
he
human
f
a
c
e
is
de
s
igned
to
e
xp
r
e
s
s
a
lar
ge
number
of
e
mot
ions
a
utom
a
ti
c
a
ll
y
f
r
om
the
e
mot
io
na
l
s
t
a
te
of
the
indi
vidual.
W
hil
e
it
is
tr
ue
that
the
f
a
c
e
is
not
the
only
mea
ns
us
e
d
by
humans
to
c
omm
unica
te
their
e
mot
ional
s
tate
,
the
im
por
tanc
e
of
the
opti
c
a
l
s
e
ns
or
f
o
r
int
e
r
a
c
ti
on
with
the
e
nvir
onment
make
s
thi
s
a
f
unda
me
ntal
f
ield
to
s
tudy
[
9
,
10
]
.
How
e
ve
r
,
ther
e
a
r
e
s
c
he
mes
de
ve
loped
to
identi
f
y
other
mec
ha
nis
ms
f
or
c
omm
unica
ti
ng
e
mot
ions
,
e
ve
n
in
a
nim
a
ls
.
S
ome
of
thes
e
media
include
voc
a
l
c
omm
unica
ti
on
[
11
,
12
]
,
body
e
xpr
e
s
s
ion
[
13]
,
f
or
ms
of
dir
e
c
t
c
ontac
t
[
14]
a
nd
ne
ur
ologi
c
a
l
pa
tt
e
r
ns
[
15,
16]
.
F
a
c
ial
e
xpr
e
s
s
ions
on
the
human
f
a
c
e
c
a
n
be
us
e
d
to
identi
f
y
a
pe
r
s
on’
s
e
mot
ional
s
tate
[
17
]
.
T
ha
t
is
why
many
s
tr
a
tegie
s
ha
ve
be
e
n
de
ve
loped
to
identi
f
y
them
f
r
om
the
p
r
oc
e
s
s
ing
of
the
im
a
ge
of
the
f
a
c
e
[
18
-
20]
.
T
he
s
e
e
xpr
e
s
s
ions
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
1
8
,
No
.
3
,
J
une
2020:
1
499
-
1504
1500
ha
ve
c
omm
on
c
ha
r
a
c
ter
is
ti
c
s
f
or
a
given
e
mot
io
na
l
s
tate
,
c
ha
r
a
c
ter
is
ti
c
s
that
a
r
e
indepe
nde
nt
o
f
r
a
c
e
or
ge
nde
r
.
I
f
it
is
pos
s
ibl
e
to
identi
f
y
thes
e
im
a
ge
s
f
r
om
the
a
na
lys
is
of
a
n
im
a
ge
,
then
it
is
pos
s
ibl
e
to
i
nf
or
m
to
a
n
a
r
ti
f
icia
l
s
ys
tem,
a
s
it
is
the
c
a
s
e
of
a
n
a
s
s
is
tant
r
obot,
how
to
a
djus
t
it
s
be
ha
vior
a
c
c
or
ding
to
the
e
mot
ional
s
tate
of
the
us
e
r
.
T
h
e
m
os
t
r
ob
us
t
s
o
lu
ti
ons
de
ve
lo
pe
d
f
o
r
t
he
a
ut
o
mat
i
c
de
te
c
t
io
n
o
f
f
a
c
ia
l
e
xp
r
e
s
s
i
ons
us
e
la
r
g
e
da
tas
e
ts
f
o
r
t
he
t
r
a
in
in
g
o
f
a
d
j
us
ta
b
le
s
c
he
mes
[
21
-
23
]
.
W
h
il
e
thes
e
s
c
he
mes
p
r
o
vi
de
h
ig
h
-
pe
r
f
o
r
ma
nc
e
va
l
ue
s
,
i
nc
r
e
a
s
in
g
t
he
i
r
r
e
l
ia
bi
l
it
y
w
i
th
lo
w
c
om
pu
ta
ti
on
a
l
c
os
t
ha
s
be
c
o
me
a
c
ha
l
le
ng
e
d
ue
to
t
he
w
i
de
va
r
i
a
b
i
li
ty
o
f
c
ha
r
a
c
te
r
is
ti
c
s
b
e
t
we
e
n
in
di
vi
du
a
ls
,
w
hi
c
h
ma
ke
s
th
e
s
e
pa
r
a
t
io
n
be
t
we
e
n
c
la
s
s
e
s
o
r
c
a
te
go
r
ies
c
o
m
ple
x
,
e
ve
n
w
i
th
ou
t
c
o
ns
i
de
r
in
g
th
e
va
r
iab
les
r
e
la
ted
to
i
ma
ge
c
a
p
tu
r
e
.
T
h
e
wo
r
k
p
r
e
s
e
nt
e
d
i
n
t
his
p
a
p
e
r
a
i
ms
to
c
o
n
tr
i
bu
te
t
o
t
his
a
r
e
a
o
f
r
e
s
e
a
r
c
h
,
pa
r
t
ic
ul
a
r
l
y
e
va
lu
a
t
in
g
th
e
pe
r
f
o
r
ma
nc
e
o
f
a
n
a
ut
o
mat
ic
ide
nt
i
f
i
c
a
ti
on
s
ys
te
m
of
e
mo
t
io
ns
in
r
e
a
l
-
t
im
e
on
a
la
bo
r
a
to
r
y
r
o
bo
t
a
s
s
e
m
b
le
d
to
pe
r
f
o
r
m
tas
ks
o
f
a
s
s
is
tan
c
e
to
h
u
man
be
in
gs
.
I
n
p
r
e
v
io
us
wo
r
ks
,
we
ha
ve
s
h
ow
n
de
ta
il
s
o
f
t
he
r
ob
o
ti
c
p
la
tf
o
r
m
a
nd
a
pe
op
le
tr
a
c
ki
ng
s
c
he
m
e
[
24
]
,
t
his
wo
r
k
is
i
n
te
gr
a
t
e
d
w
i
th
s
uc
h
s
c
h
e
m
e
s
t
o
p
r
od
uc
e
a
n
in
te
gr
a
te
d
m
a
n
-
mac
hi
ne
in
te
r
a
c
t
io
n
s
o
l
ut
io
n
.
Ou
r
r
oboti
c
plat
f
or
m
c
ons
is
ts
of
two
int
e
gr
a
ted
r
o
bots
to
f
or
m
a
s
ingl
e
he
a
lt
hc
a
r
e
s
ys
tem.
W
e
us
e
a
Na
o
r
obot
f
r
om
S
o
f
tB
a
nk
Gr
oup
a
t
the
top
f
o
r
d
ir
e
c
t
int
e
r
a
c
ti
on
with
pe
ople,
a
nd
we
us
e
ou
r
T
ur
tl
e
B
ot
1
r
obot
a
t
the
bott
om
to
s
olve
na
vigation
a
nd
pa
th
plann
ing
pr
oblems
.
As
a
ppli
e
d
r
e
s
e
a
r
c
h
the
r
e
s
e
a
r
c
h
gr
oup
ha
s
s
tudi
e
d
the
pr
oblem
of
s
a
f
e
gr
a
s
ping
of
c
yli
nd
r
ica
l
objec
ts
with
a
nth
r
opomor
phic
ha
nds
[
2
5
]
,
c
o
ns
ider
ing
a
ls
o
the
de
li
c
a
te
int
e
r
a
c
ti
on
with
humans
.
W
e
ha
v
e
a
ls
o
de
ve
loped
many
a
utonom
ous
na
vigation
s
c
he
mes
f
or
loca
ll
y
obs
e
r
va
ble
dy
na
mi
c
e
nvi
r
onments
[
2
6
-
29
]
,
a
nd
a
na
lyze
d
the
p
r
oblem
o
f
dyna
mi
c
bipeda
l
wa
lki
ng
[
3
0
]
.
All
thes
e
a
r
e
c
ur
r
e
nt
uns
olved
pr
obl
e
ms
of
he
a
lt
hc
a
r
e
r
oboti
c
s
,
w
hich
s
e
e
k
to
be
in
tegr
a
ted
int
o
the
de
ve
lopm
e
nt
o
f
mor
e
c
ompl
e
x
tas
ks
s
u
c
h
a
s
t
he
moni
tor
ing
o
f
vit
a
l
s
igns
in
pa
ti
e
nts
,
a
dmi
nis
tr
a
ti
on
of
f
ood
a
nd
medic
ine
o
r
im
media
te
a
tt
e
nti
on
in
c
a
s
e
of
e
mer
ge
nc
ies
.
T
he
r
e
s
t
of
thi
s
pa
pe
r
is
or
ga
nize
d
a
s
f
oll
ows
.
S
e
c
ti
on
2
pr
ovides
s
ome
theor
e
ti
c
a
l
de
tails
of
the
pr
oblem
be
ing
a
ddr
e
s
s
e
d.
S
e
c
ti
on
3
s
hows
th
e
methodologi
c
a
l
s
tr
a
tegy
de
ve
loped
a
s
we
ll
a
s
d
e
tails
of
the
labor
a
tor
y
im
pleme
ntation
.
T
his
is
f
ol
lowe
d
by
the
e
va
luation
of
the
s
ys
tem.
F
inally,
we
c
onc
lude
by
dis
c
us
s
ing
the
les
s
on
s
lea
r
ne
d
a
nd
how
the
s
ys
tem
c
ould
be
f
ur
ther
de
ve
lo
pe
d.
2.
P
ROB
L
E
M
F
ORM
UL
AT
I
ON
T
he
goa
l
of
thi
s
r
e
s
e
a
r
c
h
is
to
de
ve
lop
a
r
obus
t
a
nd
high
-
pe
r
f
or
manc
e
s
of
twa
r
e
tool
that
a
ll
ows
the
de
ve
lopm
e
nt
of
a
utono
mous
tas
ks
of
identif
yin
g
human
e
mot
ions
by
a
s
mall
a
utonom
ous
r
obo
t.
T
he
wor
k
is
s
tr
ongly
mot
ivate
d
by
the
ne
e
d
f
o
r
th
is
f
e
a
tur
e
a
s
pa
r
t
of
the
r
outi
ne
int
e
r
a
c
ti
on
of
a
n
a
s
s
is
ti
ve
r
obot
that
ope
r
a
tes
in
unknown
indoor
e
nvi
r
onments
.
T
his
s
or
t
of
r
obot
mus
t
na
vigate
a
n
unknown
e
nvir
onment,
but
the
main
mot
ivation
of
it
s
a
c
ti
on
is
s
tr
ongly
de
ter
mi
ne
d
by
i
ts
leve
l
o
f
int
e
r
a
c
ti
on
with
the
pe
r
s
on
f
or
whom
it
wor
ks
.
L
e
t
W
⊂
ℝ
2
be
the
c
los
ur
e
o
f
a
c
ontr
a
c
ti
ble
ope
n
s
e
t
in
the
plane
that
ha
s
a
c
onne
c
ted
ope
n
int
e
r
ior
with
obs
tac
les
that
r
e
pr
e
s
e
nt
inac
c
e
s
s
ibl
e
r
e
gions
.
L
e
t
be
a
s
e
t
of
obs
tac
le
s
,
in
whic
h
e
a
c
h
O
⊂
is
c
los
e
d
with
a
c
onne
c
ted
piec
e
wis
e
-
a
na
lyt
ic
bounda
r
y
that
is
f
ini
te
in
length.
T
he
pos
it
ion
of
obs
tac
les
in
the
e
nvir
on
ment
c
ha
nge
s
ove
r
ti
me
in
a
n
unknown
wa
y,
but
they
a
r
e
de
tec
table
by
dis
tanc
e
s
e
ns
or
s
.
I
n
a
ddit
ion,
the
obs
tac
les
in
a
r
e
pa
ir
wis
e
-
dis
joi
nt
a
nd
c
ountably
f
ini
te
in
number
.
L
e
t
E
⊂
W
be
the
f
r
e
e
s
pa
c
e
in
the
e
nvir
onment
,
whic
h
is
the
ope
n
s
ubs
e
t
of
W
w
it
h
the
obs
tac
les
r
e
moved.
T
his
s
pa
c
e
c
a
n
be
f
r
e
e
ly
na
vigate
d
by
the
r
obot,
but
it
c
a
n
a
ls
o
be
oc
c
upied
a
t
a
ny
ti
m
e
by
a
n
obs
tac
le.
T
he
r
obot
knows
the
f
r
e
e
s
pa
c
e
in
the
e
nvir
onment
E
f
r
om
obs
e
r
va
ti
ons
,
us
ing
s
e
ns
or
s
.
T
he
s
e
obs
e
r
va
ti
ons
a
ll
ow
him
to
buil
d
a
n
inf
or
mati
on
s
pa
c
e
I
.
An
inf
o
r
mation
mapping
is
of
the
f
o
r
m:
:
→
(
1)
whe
r
e
S
de
note
a
n
obs
e
r
va
ti
on
s
pa
c
e
,
c
ons
tr
uc
ted
f
r
om
s
e
ns
or
r
e
a
dings
ove
r
t
im
e
,
i.
e
.
,
th
r
ough
a
n
obs
e
r
va
ti
on
his
tor
y
of
the
f
o
r
m:
̃
:
[
0
,
]
→
(
2)
T
he
int
e
r
pr
e
tation
of
thi
s
inf
o
r
mation
s
pa
c
e
,
i
.
e
.
,
I
✕
S
￫
I
,
is
that
whic
h
a
ll
ows
the
r
obot
to
make
de
c
is
ions
.
T
he
pr
oblem
c
a
n
be
de
f
ined
a
s
the
de
f
ini
ti
on
of
a
s
pe
c
if
ic
f
il
ter
to
a
pply
to
the
da
ta
s
e
ns
e
d
by
the
r
obot's
c
a
mer
a
in
r
e
a
l
-
ti
me
to
identif
y
e
mot
ional
s
tate
s
in
the
pe
r
s
on
with
whom
it
int
e
r
a
c
ts
.
T
his
inf
or
mation
mus
t
mod
if
y
the
be
ha
vior
o
f
th
e
r
obot,
both
int
e
r
a
c
ti
on,
a
nd
moveme
nt
in
the
e
nvir
o
nment.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
I
de
nti
fi
e
r
of
human
e
motions
bas
e
d
on
c
onv
olut
ion
al
ne
ur
al
ne
tw
or
k
for
as
s
is
tant
r
obot
(
F
r
e
dy
M
ar
tí
ne
z
)
1501
3.
M
E
T
HO
DOL
OG
Y
Our
s
c
he
me
of
identi
f
ica
ti
on
of
human
e
mot
ions
i
s
c
ompos
e
d
of
two
blocks
:
the
f
ir
s
t
block
of
f
a
c
e
de
tec
ti
on,
a
nd
the
s
e
c
ond
block
of
identif
ica
ti
on
o
f
e
mot
ions
in
the
f
a
c
e
s
e
xtr
a
c
ted
f
r
om
the
video
f
r
a
mes
a
s
s
hown
in
F
igur
e
1.
T
he
f
inal
r
e
s
ult
(
identif
ied
f
a
c
e
a
nd
e
mot
ion)
is
plac
e
d
on
e
a
c
h
f
r
a
me
to
r
e
c
ons
tr
uc
t
the
video
(
thi
s
is
v
is
ua
li
z
e
d
on
a
7
-
inch
touch
s
c
r
e
e
n
on
the
r
obo
t
f
o
r
va
li
da
ti
on
pur
pos
e
s
)
.
F
or
f
a
c
e
de
tec
ti
on
we
us
e
dli
b
.
ge
t_f
r
ontal
_face
_de
tec
tor
(
)
.
T
his
f
a
c
e
de
tec
tor
is
ba
s
e
d
on
the
his
togr
a
m
o
f
o
r
iente
d
gr
a
dients
(
HO
G)
a
nd
li
ne
a
r
S
VM
(
S
uppor
t
-
Ve
c
tor
M
a
c
hine)
.
T
he
r
e
is
a
ls
o
a
f
a
c
e
de
tec
tor
in
dli
b
ba
s
e
d
on
C
NN
(
c
onvolut
ional
ne
ur
a
l
ne
twor
k
)
,
a
nd
a
lt
hough
it
ha
s
a
be
tt
e
r
pe
r
f
or
manc
e
de
tec
ti
ng
f
a
c
e
s
in
man
y
a
ngles
(
dli
b.
ge
t_f
r
ontal
_face
_de
tec
tor
(
)
only
wor
ks
we
ll
with
f
r
ontal
f
a
c
e
s
)
,
we
de
c
ided
not
to
us
e
it
due
to
it
s
high
c
ons
umpt
ion
of
r
e
s
our
c
e
s
,
whic
h
made
it
im
pos
s
ibl
e
to
us
e
in
r
e
a
l
-
ti
me
on
our
a
s
s
is
tant
r
obot
(
the
s
tr
a
tegy
ba
s
e
d
on
C
NN
wa
s
a
li
tt
le
mor
e
than
15
ti
mes
s
lowe
r
in
e
a
c
h
f
r
a
me
than
the
s
tr
a
tegy
ba
s
e
d
on
HO
G
)
.
T
he
identif
ica
ti
on
of
e
mo
ti
ons
is
done
th
r
ough
a
c
onvolut
ional
ne
ur
a
l
ne
two
r
k
that
im
pleme
nts
the
n
e
twor
k
s
tr
uc
tur
e
pr
opos
e
d
in
[
31
]
.
T
h
is
a
r
c
hit
e
c
tur
e
us
e
s
global
a
ve
r
a
ge
pooli
ng
in
it
s
output
laye
r
with
S
of
max
a
c
ti
va
ti
on
f
unc
ti
on
(
number
o
f
c
las
s
e
s
e
q
ua
l
to
the
number
o
f
f
e
a
tur
e
maps
)
to
e
li
mi
na
te
t
he
f
ull
y
c
onne
c
ted
laye
r
s
,
a
nd
thus
c
ons
ider
a
bly
r
e
duc
e
t
he
number
of
a
djus
table
ne
twor
k
pa
r
a
mete
r
s
.
T
h
e
r
e
s
t
of
the
ne
twor
k
is
a
s
tanda
r
d
f
ull
y
-
c
onvolut
ional
ne
ur
a
l
ne
twor
k
with
a
tot
a
l
o
f
nine
c
onvolut
ion
la
ye
r
s
with
R
e
L
U
(
r
e
c
ti
f
ied
li
ne
a
r
unit
)
a
c
ti
va
ti
on
f
unc
ti
on
to
incr
e
a
s
e
s
e
n
s
it
ivi
ty
a
nd
a
void
s
a
tur
a
ti
on,
a
nd
ba
tch
nor
maliza
ti
on
s
o
that
e
a
c
h
laye
r
of
the
ne
two
r
k
lea
r
ns
by
it
s
e
lf
mo
r
e
f
e
a
tur
e
s
indepe
nde
ntl
y
o
f
t
he
other
laye
r
s
.
T
he
f
inal
model
c
ontains
a
li
t
tl
e
les
s
th
a
n
600,
000
a
djus
table
pa
r
a
mete
r
s
a
nd
wa
s
tr
a
in
e
d
with
a
da
t
a
s
e
t
of
7,
000
im
a
ge
s
,
with
1,
000
im
a
ge
s
in
e
a
c
h
of
it
s
s
e
ve
n
c
a
tegor
ies
:
a
ngr
y,
dis
gus
t,
f
e
a
r
,
ha
ppy,
s
a
d,
s
ur
pr
is
e
a
nd
ne
utr
a
l
.
T
he
t
r
a
ini
ng
us
e
d
70%
of
th
e
da
tas
e
t,
a
nd
the
r
e
maining
30%
wa
s
us
e
d
f
or
va
li
da
ti
on.
T
he
a
c
c
ur
a
c
y
a
c
hieve
d
by
the
model
with
th
e
va
li
da
ti
on
da
ta
wa
s
58%
.
T
he
c
ode
wa
s
de
ve
loped
in
P
ython,
a
nd
we
us
e
Ke
r
a
s
a
s
the
de
e
p
lea
r
ning
f
r
a
mew
or
k.
I
n
F
igur
e
2
we
s
how
a
s
a
mpl
e
of
the
im
a
ge
s
us
e
d
f
or
the
a
ngr
y
c
a
tegor
y.
A
de
e
p
lea
r
ning
f
r
a
mew
or
k
is
a
n
int
e
r
f
a
c
e
,
a
li
br
a
r
y
or
a
tool
that
e
na
bl
e
s
to
buil
d
de
e
p
lea
r
ning
models
mor
e
e
a
s
il
y
a
nd
quickly
,
wi
thout
going
i
nto
the
de
tails
o
f
the
unde
r
lyi
ng
a
lgor
it
hms
.
T
he
y
pr
ovide
a
c
lea
r
a
nd
c
onc
is
e
wa
y
to
de
f
ine
models
us
ing
a
c
oll
e
c
ti
on
of
pr
e
-
buil
t
a
nd
opti
mi
z
e
d
c
omponents
.
I
ns
tea
d
of
wr
it
ing
hundr
e
ds
of
li
ne
s
of
c
ode
,
it
is
pos
s
ibl
e
t
o
us
e
a
s
uit
a
ble
f
r
a
mew
or
k
that
a
ll
ows
a
model
of
t
his
type
to
be
buil
t
quickly.
T
he
r
e
a
r
e
many
mac
hine
lea
r
ning
f
r
a
mew
or
ks
a
lr
e
a
dy
in
plac
e
,
a
nd
ne
w
f
r
a
mew
or
ks
a
r
e
a
ppe
a
r
ing
r
e
gular
ly
to
a
ddr
e
s
s
s
pe
c
if
ic
niche
s
.
W
e
a
r
e
us
ing
Ke
r
a
s
a
s
a
f
r
a
mew
or
k
(
ht
tps
:/
/ker
a
s
.
io/
)
with
T
e
ns
or
f
low
ba
c
ke
nd
(
htt
ps
:/
/www
.
tens
or
f
low
.
or
g)
.
T
his
e
mot
ion
identif
ica
ti
on
pr
og
r
a
m
wa
s
im
pleme
nted
on
our
a
s
s
is
tant
r
obot
to
wor
k
in
pa
r
a
ll
e
l
with
other
a
lgor
it
hms
of
s
pe
c
if
ic
pur
pos
e
,
in
pa
r
t
icula
r
with
our
r
obot
-
human
ba
s
ic
int
e
r
a
c
ti
on
s
c
he
me
s
uppor
ted
in
vis
ua
l
tr
a
c
king
f
o
r
a
s
s
is
tant
r
obot
[
24]
.
Our
a
s
s
is
ti
ve
r
obot
c
ons
is
ts
of
two
r
oboti
c
plat
f
or
ms
:
A
h
umanoid
Na
o
r
obot
f
r
om
S
of
tB
a
nk
Gr
oup
f
or
int
e
r
a
c
ti
on
wi
th
humans
a
nd
the
e
nvi
r
onment,
a
nd
a
n
AR
M
OS
T
ur
tl
e
B
ot
1
r
obot
f
r
om
the
AR
M
OS
r
e
s
e
a
r
c
h
g
r
oup
f
or
ind
oor
na
vigation
a
s
s
hown
in
F
igu
r
e
3
.
C
omm
unica
ti
on
with
the
two
p
latf
or
ms
is
via
a
W
i
-
F
i
c
onne
c
ti
on
.
T
he
a
lgor
it
hm
wa
s
de
ve
loped
in
P
ython
a
nd
wa
s
pr
og
r
a
mm
e
d
on
the
AR
M
OS
T
ur
tl
e
B
ot
1
r
obot.
T
he
vide
o
is
c
a
ptur
e
d
by
the
c
a
mer
a
on
the
he
a
d
of
the
Na
o
r
obot,
a
nd
tr
a
ns
mi
tt
e
d
by
W
i
-
F
i.
P
r
oc
e
s
s
ing
is
pe
r
f
or
m
e
d
by
the
AR
M
OS
T
u
r
tl
e
B
ot
1
r
obot
in
r
e
a
l
-
ti
me
in
it
s
c
ontr
ol
unit
(
Dr
a
gonB
oa
r
d
410c
of
Ar
r
ow
E
lec
tr
o
nics
with
AR
M
C
or
tex
-
A53
Qua
d
-
c
or
e
up
to
1.
2
GH
z
pe
r
c
or
e
a
nd
Qua
lcomm
Ad
r
e
no
306
@
400M
Hz
)
.
W
e
us
e
in
our
im
pleme
ntation
Ope
nC
V
4.
1.
0
,
nu
mp
y
1.
16
.
2,
Ke
r
a
s
2.
2.
4,
T
e
ns
or
F
low
1.
14
.
0,
dli
b
19
.
17.
0
,
a
nd
i
mut
il
s
0.
5.
2
.
F
igur
e
1.
P
r
opos
e
d
s
c
he
me
f
o
r
the
identif
ica
ti
on
o
f
human
e
mot
ions
on
our
r
obot
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
1
8
,
No
.
3
,
J
une
2020:
1
499
-
1504
1502
F
igur
e
2.
Da
tas
e
t
us
e
d
f
or
c
onvolut
ional
ne
twor
k
t
r
a
ini
ng
in
the
a
ngr
y
c
a
tegor
y
F
igur
e
3.
E
xpe
r
im
e
ntal
s
e
tup
f
o
r
the
e
mot
ion
ident
if
ica
ti
on
s
ys
tem.
I
t
is
c
ompos
e
d
of
a
humanoid
Na
o
r
obot
f
r
om
S
of
tB
a
nk
Gr
oup
a
t
the
top
a
nd
a
n
AR
M
OS
T
ur
tl
e
B
ot
1
tank
r
o
bot
f
r
om
the
AR
M
OS
r
e
s
e
a
r
c
h
gr
oup
a
t
the
bott
om
T
he
im
a
ge
s
we
r
e
c
a
ptur
e
d
by
the
Na
o
r
obot
a
t
a
s
ize
of
1280
✕
720
pixels
.
E
a
c
h
f
r
a
me
wa
s
s
c
a
led
to
632
✕
480
f
or
pr
oc
e
s
s
ing
in
the
Dr
a
gonB
oa
r
d.
W
e
c
onduc
ted
s
e
v
e
r
a
l
e
xpe
r
im
e
nts
with
dif
f
e
r
e
n
t
li
ghti
ng
c
ondit
ions
,
e
nvir
on
ments
,
pe
ople
a
nd
dis
tanc
e
s
t
o
the
r
obot
.
W
e
e
ve
n
pe
r
f
o
r
med
tes
ts
with
two
or
mo
r
e
pe
ople
a
t
dif
f
e
r
e
nt
dis
tanc
e
s
,
but
a
lwa
ys
withi
n
th
e
r
obot's
f
ield
o
f
vis
ion.
One
o
f
thes
e
tes
ts
c
a
n
be
s
e
e
n
in:
htt
ps
:
//
y
outu.
be
/yQ
ajUZ
T
T
w
mA
a
s
s
hown
in
F
igur
e
4.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
I
de
nti
fi
e
r
of
human
e
motions
bas
e
d
on
c
onv
olut
ion
al
ne
ur
al
ne
tw
or
k
for
as
s
is
tant
r
obot
(
F
r
e
dy
M
ar
tí
ne
z
)
1503
F
igur
e
4.
C
a
ptur
e
o
f
two
e
mot
ive
s
tate
s
identif
ied
dur
ing
one
o
f
the
tes
ts
de
ve
loped
in
the
labor
a
tor
y
to
the
p
r
opos
e
d
s
c
he
me
4.
F
I
ND
I
NGS
I
n
ge
ne
r
a
l
,
the
r
obot
e
a
s
il
y
de
tec
ts
pe
ople's
f
a
c
e
s
a
nd
c
or
r
e
c
tl
y
identi
f
ies
thei
r
e
mot
ions
.
T
he
a
lgor
it
hm
de
mons
tr
a
ted
high
pe
r
f
or
manc
e
in
labor
a
tor
y
tes
ts
a
ll
owing
r
e
a
l
-
ti
me
ope
r
a
ti
on
a
nd
pa
r
a
ll
e
l
e
xe
c
uti
on
with
t
r
a
c
king
a
nd
tr
a
c
ing
s
c
he
mes
.
T
h
e
f
law
s
in
the
identif
ica
ti
on
of
e
mot
ion
in
a
f
r
a
me
we
r
e
s
olved
by
a
ve
r
a
ging
the
r
e
s
ult
of
ne
ighbor
ing
f
r
a
mes
,
whi
c
h
r
e
s
ult
e
d
in
a
s
uc
c
e
s
s
r
a
te
gr
e
a
ter
than
92%
.
As
mentioned,
the
f
a
c
e
de
tec
ti
on
s
c
he
me
ha
s
p
r
oblems
identif
ying
f
a
c
e
s
that
a
r
e
not
c
ompl
e
tel
y
f
a
c
ing
the
r
obot
.
I
n
the
f
ir
s
t
ins
tanc
e
,
thi
s
ha
s
not
b
e
e
n
a
pr
oblem
f
o
r
ou
r
a
ppli
c
a
ti
on
s
ince
it
s
e
e
ks
dir
e
c
t
human
-
mac
hine
int
e
r
a
c
ti
on.
How
e
ve
r
,
the
r
e
s
ult
s
s
how
that
it
is
ne
c
e
s
s
a
r
y
to
e
va
luate
mor
e
r
obus
t
s
tr
a
tegie
s
.
T
he
e
mot
ion
identif
ica
ti
on
s
c
he
me
pr
e
s
e
nts
s
ome
leve
l
of
bias
c
onc
e
r
ning
the
da
taba
s
e
us
e
d
f
or
it
s
tr
a
ini
ng.
M
os
t
of
the
f
a
c
e
s
us
e
d
f
o
r
tr
a
ini
ng
we
r
e
we
s
ter
n
pe
ople,
a
nd
thi
s
is
r
e
f
lec
ted
in
the
r
e
s
ult
s
f
or
s
ome
f
r
a
mes
.
T
his
pr
oblem
wa
s
s
olved
by
a
ve
r
a
ging
the
be
ha
vior
a
long
with
s
e
ve
r
a
l
f
r
a
mes
unde
r
the
a
s
s
umpt
ion
that
the
pe
r
s
on
r
e
mains
with
the
s
a
me
e
mot
ion
a
f
e
w
mi
ll
is
e
c
onds
,
a
nd
ther
e
f
or
e
is
obs
e
r
ve
d
in
s
e
ve
r
a
l
ne
ighbor
ing
f
r
a
mes
.
I
t
wa
s
a
ls
o
obs
e
r
ve
d
in
s
ome
c
a
s
e
s
pr
oblems
to
identif
y
the
e
mot
ions
whe
n
th
e
pe
r
s
on
ha
s
glas
s
e
s
a
nd/or
c
igar
e
tt
e
s
,
the
pr
oblem
is
mi
no
r
,
but
a
ls
o
r
a
is
e
s
the
ne
e
d
to
im
pr
ove
the
tr
a
ini
ng
da
taba
s
e
of
the
ne
twor
k
.
5.
CONC
L
USI
ON
I
n
thi
s
pa
pe
r
,
we
p
r
e
s
e
nt
the
a
ppli
c
a
ti
on
of
a
s
c
he
me
f
or
the
identif
ica
ti
on
of
human
e
mot
ions
f
o
r
a
s
s
is
t
a
nc
e
r
obots
.
T
he
objec
ti
ve
of
thi
s
a
ppli
c
a
ti
on
is
to
im
pr
ove
the
human
-
mac
hine
int
e
r
a
c
ti
on
of
our
r
oboti
c
platf
or
m.
T
he
s
c
he
me
is
s
tr
uc
tur
e
d
a
r
oun
d
two
f
unc
ti
ona
l
blocks
,
a
f
a
c
e
de
tec
tor
us
ing
H
OG
a
nd
li
ne
a
r
S
VM
,
a
nd
a
n
e
mot
ion
ident
if
ica
ti
on
block
u
s
ing
C
NN
.
T
he
output
inf
o
r
mation
o
f
the
s
c
he
me
(
loca
ti
on
of
the
f
a
c
e
a
nd
e
mot
ion
o
f
the
pe
r
s
on)
is
us
e
d
to
c
oor
dinate
the
moveme
nt
a
nd
r
e
a
c
ti
ons
o
f
th
e
r
obot
.
T
he
im
pleme
nted
s
c
he
me
de
mons
tr
a
ted
a
low
c
om
putational
c
os
t,
a
ll
owing
it
s
us
e
in
r
e
a
l
-
ti
me
a
nd
in
pa
r
a
ll
e
l
with
other
r
outi
ne
s
of
a
c
ti
ve
tr
a
c
king
of
pe
ople
,
t
his
is
the
mos
t
outs
tanding
r
e
s
ult
o
f
ou
r
im
pleme
ntation.
T
he
pe
r
f
o
r
manc
e
e
va
luation
a
ll
owe
d
t
o
identif
y
pos
s
ibl
e
im
pr
ove
ments
to
the
s
tr
a
tegy,
in
p
a
r
ti
c
ular
,
to
im
pr
ove
the
tr
a
ini
ng
of
the
c
onvolut
ional
ne
twor
k
by
i
mpr
oving
the
ini
ti
a
l
da
tas
e
t.
Af
ter
a
ve
r
a
ging
the
identif
ica
ti
on
c
a
pa
bil
it
y
ove
r
s
e
ve
r
a
l
ne
ighb
or
ing
f
r
a
mes
to
s
olve
the
nois
e
ident
i
f
ica
ti
on
p
r
oblems
in
the
im
a
ge
s
,
the
s
c
he
me
a
c
hieve
d
a
92%
s
uc
c
e
s
s
r
a
te
ope
r
a
ti
ng
in
r
e
a
l
ti
me
on
the
r
obot
c
ontr
ol
unit
.
AC
KNOWL
E
DGM
E
N
T
S
T
his
wor
k
wa
s
s
uppor
ted
by
the
Unive
r
s
idad
Di
s
tr
it
a
l
F
r
a
nc
is
c
o
J
os
é
de
C
a
ldas
,
in
pa
r
t
th
r
ough
C
I
DC
,
a
nd
pa
r
tl
y
by
the
F
a
c
ult
a
d
T
e
c
nológi
c
a
.
T
he
view
s
e
xpr
e
s
s
e
d
in
thi
s
pa
pe
r
a
r
e
not
ne
c
e
s
s
a
r
il
y
e
ndor
s
e
d
by
Dis
tr
ict
Unive
r
s
it
y.
T
he
a
uthor
s
thank
the
r
e
s
e
a
r
c
h
gr
oup
AR
M
OS
f
or
the
e
va
luation
c
a
r
r
ied
out
on
pr
otot
ype
s
of
idea
s
a
nd
s
tr
a
tegie
s
.
RE
F
E
RE
NC
E
S
[1
]
M.
H
ao
,
et
al
.
,
“
Pro
p
o
s
a
l
o
f
i
n
i
t
i
at
i
v
e
s
er
v
i
ce
mo
d
e
l
fo
r
s
erv
i
ce
ro
b
o
t
,”
CA
A
I
Tr
a
n
s
a
c
t
i
o
n
s
o
n
In
t
el
l
i
g
e
n
ce
Tech
n
o
l
o
g
y
,
v
o
l
.
2
,
n
o
.
4
,
p
p
.
1
4
8
-
1
5
3
,
2
0
1
7
.
[2
]
C.
Si
ri
t
h
u
n
g
e,
et
a
l
.
,
“
Pro
act
i
v
e
r
o
b
o
t
s
w
i
t
h
t
h
e
p
erce
p
t
i
o
n
o
f
n
o
n
v
er
b
al
h
u
ma
n
b
e
h
av
i
o
r
:
A
re
v
i
e
w
,”
IE
E
E
A
cc
es
s
,
v
o
l
.
7
,
n
o
.
1
,
p
p
.
7
7
3
0
8
-
7
7
3
2
7
,
2
0
1
9
.
[3
]
M.
A
.
V
.
J
.
Mu
t
h
u
g
al
a
an
d
A
.
G
.
B
.
P.
J
ay
as
ek
ara
,
“
A
rev
i
e
w
o
f
s
erv
i
ce
ro
b
o
t
s
co
p
i
n
g
w
i
t
h
u
n
cer
t
ai
n
i
n
f
o
rmat
i
o
n
i
n
n
a
t
u
ra
l
l
a
n
g
u
ag
e
i
n
s
t
ru
c
t
i
o
n
s
,”
IE
E
E
A
cce
s
s
,
v
o
l
.
6
,
n
o
.
1
,
p
p
.
1
2
9
1
3
-
1
2
9
2
8
,
2
0
1
8
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
1
8
,
No
.
3
,
J
une
2020:
1
499
-
1504
1504
[4
]
D
.
P.
L
o
s
ey
an
d
M.
K
.
O
âMal
l
ey
,
“
E
n
ab
l
i
n
g
ro
b
o
t
s
t
o
i
n
fer
h
o
w
en
d
-
u
s
er
s
t
each
a
n
d
l
ear
n
t
h
r
o
u
g
h
h
u
man
-
r
o
b
o
t
i
n
t
eract
i
o
n
,”
IE
E
E
R
o
b
o
t
i
c
s
a
n
d
A
u
t
o
m
a
t
i
o
n
Let
t
er
s
,
v
o
l
.
4
,
n
o
.
2
,
p
p
.
1
9
5
6
-
1
9
6
3
,
2
0
1
9
.
[
5
]
S
.
S
a
u
n
d
e
r
s
o
n
a
n
d
G
.
N
e
j
a
t
,
“
I
t
w
o
u
l
d
m
a
k
e
m
e
h
a
p
p
y
i
f
y
o
u
u
s
e
d
m
y
g
u
e
s
s
:
C
o
m
p
a
r
i
n
g
r
o
b
o
t
p
e
r
s
u
a
s
i
v
e
s
t
r
a
t
e
g
i
e
s
i
n
s
o
c
i
a
l
h
u
m
a
n
-
r
o
b
o
t
i
n
t
e
r
a
c
t
i
o
n
,”
I
E
E
E
R
o
b
o
t
i
c
s
a
n
d
A
u
t
o
m
a
t
i
o
n
L
e
t
t
e
r
s
,
v
o
l
.
4
,
n
o
.
2
,
p
p
.
1
7
0
7
-
1
7
1
4
,
2
0
1
9
.
[6
]
L
.
Ch
en
,
et
al
.
,
“
In
fo
rmat
i
o
n
-
d
ri
v
en
mu
l
t
i
ro
b
o
t
b
e
h
av
i
o
r
ad
ap
t
at
i
o
n
t
o
em
o
t
i
o
n
al
i
n
t
e
n
t
i
o
n
i
n
h
u
ma
n
-
ro
b
o
t
i
n
t
erac
t
i
o
n
,”
IE
E
E
Tr
a
n
s
a
ct
i
o
n
s
o
n
Co
g
n
i
t
i
ve
a
n
d
D
eve
l
o
p
m
en
t
a
l
S
ys
t
em
s
,
v
o
l
.
10
,
n
o
.
3
,
p
p
.
6
4
7
-
6
5
8
,
2
0
1
8
.
[7
]
C.
Cl
av
el
a
n
d
Z
.
Cal
l
ej
a
s
,
“
Sen
t
i
men
t
an
a
l
y
s
i
s
:
Fro
m
o
p
i
n
i
o
n
mi
n
i
n
g
t
o
h
u
ma
n
-
ag
en
t
i
n
t
erac
t
i
o
n
,”
IE
E
E
Tr
a
n
s
a
ct
i
o
n
s
o
n
A
f
f
ec
t
i
ve
C
o
m
p
u
t
i
n
g
,
v
o
l
.
7
,
n
o
.
1
,
p
p
.
74
-
9
3
,
2
0
1
6
.
[8
]
U
.
J
ai
n
,
et
a
l
.
,
“
Cu
b
i
c
s
v
m
cl
a
s
s
i
fi
er
-
b
as
e
d
feat
u
re
ex
t
ract
i
o
n
a
n
d
em
o
t
i
o
n
d
e
t
ect
i
o
n
fro
m
s
p
eec
h
s
i
g
n
al
s
,”
2
0
1
8
In
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
er
e
n
ce
o
n
S
en
s
o
r
Net
w
o
r
k
s
a
n
d
S
i
g
n
a
l
P
r
o
ces
s
i
n
g
(
S
NS
P
)
, p
p.
3
8
6
-
3
9
1
,
2
0
1
8
.
[9
]
J
.
S.
Cas
t
añ
e
d
a
B.
an
d
Y
.
A
.
Sal
g
u
er
o
L.
,
“
A
d
j
u
s
t
me
n
t
o
f
v
i
s
u
a
l
i
d
e
n
t
i
fi
c
at
i
o
n
al
g
o
ri
t
h
m
f
o
r
u
s
e
i
n
s
t
a
n
d
-
a
l
o
n
e
ro
b
o
t
n
a
v
i
g
at
i
o
n
ap
p
l
i
cat
i
o
n
s
,”
Tekh
n
ê
,
v
o
l
.
14
,
n
o
.
1
,
p
p
.
73
-
8
6
,
2
0
1
7
.
[1
0
]
L
.
Cao
,
et
al
.
,
“
Ro
b
u
s
t
p
ca
fo
r
face
reco
g
n
i
t
i
o
n
w
i
t
h
o
ccl
u
s
i
o
n
u
s
i
n
g
s
y
mme
t
ry
i
n
fo
rma
t
i
o
n
,”
IE
E
E
1
6
t
h
In
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
er
e
n
ce
o
n
Net
wo
r
ki
n
g
,
S
e
n
s
i
n
g
a
n
d
C
o
n
t
r
o
l
(ICNS
C
2
0
1
9
)
, p
p.
3
2
3
-
3
2
8
,
2
0
1
9
.
[1
1
]
S.
A.
Red
d
y
,
et
al
.
,
“
T
h
e
d
eci
s
i
v
e
emo
t
i
o
n
i
d
e
n
t
i
fi
er?
”
2
0
1
1
3
rd
In
t
er
n
a
t
i
o
n
a
l
Co
n
f
e
r
en
ce
o
n
E
l
ect
r
o
n
i
c
s
Co
m
p
u
t
er
Tech
n
o
l
o
g
y
,
v
o
l
.
2
,
p
p.
28
-
3
2
,
2
0
1
1
.
[1
2
]
M.
Si
d
o
ro
v
,
et
al
.
,
“
E
mo
t
i
o
n
s
are
a
p
ers
o
n
al
t
h
i
n
g
:
T
o
w
ar
d
s
s
p
eak
er
-
a
d
ap
t
i
v
e
emo
t
i
o
n
reco
g
n
i
t
i
o
n
,”
IE
E
E
In
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
e
r
en
ce
o
n
A
c
o
u
s
t
i
cs
,
S
p
eech
a
n
d
S
i
g
n
a
l
P
r
o
ces
s
i
n
g
(ICA
S
S
P
2
0
1
4
)
, p
p.
4
8
0
3
-
4
8
0
7
,
2
0
1
4
.
[1
3
]
A
.
K
l
ei
n
s
m
i
t
h
an
d
N
.
Bi
an
c
h
i
-
Ber
t
h
o
u
ze
,
“
A
ffec
t
i
v
e
b
o
d
y
ex
p
re
s
s
i
o
n
p
ercep
t
i
o
n
an
d
reco
g
n
i
t
i
o
n
:
A
s
u
r
v
e
y
,”
IE
E
E
Tr
a
n
s
a
ct
i
o
n
s
o
n
A
f
f
ec
t
i
ve
C
o
m
p
u
t
i
n
g
,
v
o
l
.
4
,
n
o
.
1
,
p
p
.
15
-
3
3
,
2
0
1
3
.
[1
4
]
G
ao
Y.
,
et
al
.
,
“
W
h
at
d
o
es
t
o
u
ch
t
e
l
l
u
s
ab
o
u
t
em
o
t
i
o
n
s
i
n
t
o
u
c
h
s
cree
n
-
b
a
s
ed
g
ame
p
l
a
y
?
”
Tr
a
n
s
a
c
t
i
o
n
s
o
n
Co
m
p
u
t
er
-
H
u
m
a
n
In
t
e
r
a
c
t
i
o
n
,
v
o
l
.
19
,
n
o
.
4
,
p
p
.
1
-
1
9
,
D
ec
2
0
1
2
.
[
1
5
]
C
.
Q
i
n
g
,
e
t
a
l
.
,
“
I
n
t
e
r
p
r
e
t
a
b
l
e
e
m
o
t
i
o
n
r
e
c
o
g
n
i
t
i
o
n
u
s
i
n
g
e
e
g
s
i
g
n
a
l
s
,”
I
E
E
E
A
c
c
e
s
s
,
v
o
l
.
7
,
n
o
.
1
,
p
p
.
9
4
1
6
0
-
9
4
1
7
0
,
2
0
1
9
.
[1
6
]
R.
M
.
Meh
mo
o
d
,
et
a
l
.
,
“
O
p
t
i
mal
fea
t
u
re
s
e
l
ect
i
o
n
an
d
d
eep
l
earn
i
n
g
e
n
s
em
b
l
e
s
met
h
o
d
f
o
r
emo
t
i
o
n
rec
o
g
n
i
t
i
o
n
fro
m
h
u
ma
n
b
ra
i
n
eeg
s
en
s
o
r
s
,”
IE
E
E
A
cce
s
s
,
v
o
l
.
5
,
n
o
.
1
,
p
p
.
1
4
7
9
7
-
1
4
8
0
6
,
2
0
1
7
.
[1
7
]
H
.
W
an
g
an
d
J
.
G
u
,
“
T
h
e
ap
p
l
i
cat
i
o
n
s
o
f
faci
a
l
ex
p
re
s
s
i
o
n
reco
g
n
i
t
i
o
n
i
n
h
u
man
-
c
o
mp
u
t
e
r
i
n
t
er
act
i
o
n
,”
IE
E
E
In
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
e
r
en
ce
o
n
A
d
va
n
ced
M
a
n
u
f
a
c
t
u
r
i
n
g
(ICA
M
2
0
1
8
)
, p
p.
2
8
8
-
2
9
1
,
2
0
1
8
.
[1
8
]
J
.
D
en
g
,
et
a
l
.
,
“C
g
a
n
b
a
s
ed
fac
i
al
ex
p
res
s
i
o
n
rec
o
g
n
i
t
i
o
n
fo
r
h
u
ma
n
-
ro
b
o
t
i
n
t
er
act
i
o
n
,”
IE
E
E
A
cces
s
,
v
o
l
.
7
,
n
o
.
1
,
p
p
.
9
8
4
8
-
9
8
5
9
,
2
0
1
9
.
[1
9
]
P.
T
zi
ra
k
i
s
,
et
a
l
.
,
“
E
n
d
-
to
-
e
n
d
m
u
l
t
i
m
o
d
a
l
em
o
t
i
o
n
r
eco
g
n
i
t
i
o
n
u
s
i
n
g
d
eep
n
eu
ra
l
n
et
w
o
r
k
s
,”
IE
E
E
Jo
u
r
n
a
l
o
f
S
el
ec
t
ed
To
p
i
c
s
i
n
S
i
g
n
a
l
P
r
o
ces
s
i
n
g
,
v
o
l
.
11
,
n
o
.
8
,
p
p
.
1
3
0
1
-
1
3
0
9
,
2
0
1
7
.
[2
0
]
R.
Y
an
g
,
et
al
.
,
“
In
t
el
l
i
g
e
n
t
mi
rr
o
r
s
y
s
t
em
b
as
e
d
o
n
faci
al
ex
p
res
s
i
o
n
reco
g
n
i
t
i
o
n
an
d
co
l
o
r
emo
t
i
o
n
ad
a
p
t
a
t
i
o
n
â
â
i
mi
rr
o
r
,”
3
7
t
h
Ch
i
n
e
s
e
Co
n
t
r
o
l
Co
n
f
er
e
n
ce
(CCC
2
0
1
8
)
,
p
p.
3
2
2
7
-
3
2
3
2
,
2
0
1
8
.
[2
1
]
L
.
Ch
en
,
et
a
l
.
,
“
T
h
ree
-
l
a
y
er
w
e
i
g
h
t
e
d
f
u
zzy
s
u
p
p
o
rt
v
ect
o
r
reg
re
s
s
i
o
n
f
o
r
emo
t
i
o
n
a
l
i
n
t
en
t
i
o
n
u
n
d
ers
t
an
d
i
n
g
i
n
h
u
ma
n
-
ro
b
o
t
i
n
t
eract
i
o
n
,”
IE
E
E
Tr
a
n
s
a
c
t
i
o
n
s
o
n
F
u
z
z
y
S
ys
t
em
s
,
v
o
l
.
26
,
n
o
.
5
,
p
p
.
2
5
2
4
-
2
5
3
8
,
2
0
1
8
.
[2
2
]
Z
.
L
i
u
,
et
al
.
,
“
A
faci
al
ex
p
res
s
i
o
n
em
o
t
i
o
n
reco
g
n
i
t
i
o
n
-
b
as
e
d
h
u
ma
n
-
ro
b
o
t
i
n
t
erac
t
i
o
n
s
y
s
t
em
,”
IE
E
E
/
C
A
A
Jo
u
r
n
a
l
o
f
A
u
t
o
m
a
t
i
ca
S
i
n
i
c
a
,
v
o
l
.
4
,
n
o
.
4
,
p
p
.
6
6
8
-
6
7
6
,
2
0
1
7
.
[2
3
]
H
.
Men
g
,
et
al
.
,
“
T
i
me
-
d
el
a
y
n
eu
ra
l
n
et
w
o
r
k
fo
r
co
n
t
i
n
u
o
u
s
emo
t
i
o
n
a
l
d
i
me
n
s
i
o
n
p
red
i
ct
i
o
n
f
ro
m
faci
al
ex
p
res
s
i
o
n
s
eq
u
en
ce
s
,”
IE
E
E
Tr
a
n
s
a
ct
i
o
n
s
o
n
Cyb
e
r
n
e
t
i
c
s
,
v
o
l
.
46
,
n
o
.
4
,
p
p
.
9
1
6
-
9
2
9
,
2
0
1
6
.
[2
4
]
F.
Mart
í
n
ez,
et
al
.
,
“
Ro
b
o
t
-
h
u
man
b
a
s
i
c
i
n
t
eract
i
o
n
s
c
h
eme
s
u
p
p
o
r
t
ed
i
n
v
i
s
u
al
t
rack
i
n
g
fo
r
as
s
i
s
t
a
n
ce
ro
b
o
t
u
s
i
n
g
d
i
ffere
n
t
i
al
fi
l
t
er
a
n
d
k
-
me
an
s
c
l
u
s
t
er
i
n
g
,”
Tecn
u
r
a
,
v
o
l
.
24
,
n
o
.
6
3
,
p
p
.
1
-
1
9
,
2
0
2
0
.
[2
5
]
E
.
Ro
d
rí
g
u
ez,
et
al
.
,
“
Fu
zzy
Co
n
t
r
o
l
fo
r
Cy
l
i
n
d
ri
ca
l
G
ras
p
i
n
A
n
t
h
ro
p
o
m
o
rp
h
i
c
H
a
n
d
,”
Co
n
t
em
p
o
r
a
r
y
E
n
g
i
n
ee
r
i
n
g
S
ci
e
n
ces
,
v
o
l
.
10
,
n
o
.
30
,
p
p
.
1
4
8
5
-
1
4
9
2
,
2
0
1
7
.
[2
6
]
F.
Mart
í
n
ez,
et
a
l
.
,
“
A
S
t
u
d
y
o
n
Mach
i
n
e
L
earn
i
n
g
Mo
d
e
l
s
fo
r
C
o
n
v
erg
e
n
ce
T
i
me
Pre
d
i
c
t
i
o
n
s
i
n
Reac
t
i
v
e
N
av
i
g
a
t
i
o
n
St
ra
t
eg
i
es
,”
Co
n
t
e
m
p
o
r
a
r
y
E
n
g
i
n
eer
i
n
g
S
ci
e
n
ces
,
v
o
l
.
10
,
n
o
.
25
,
p
p
.
1
2
2
3
-
1
2
3
2
,
2
0
1
7
.
[2
7
]
F.
Mart
í
n
ez,
et
al
.
,
“
V
i
s
u
al
i
d
e
n
t
i
fi
ca
t
i
o
n
an
d
s
i
mi
l
ar
i
t
y
meas
u
re
s
u
s
ed
fo
r
o
n
-
l
i
n
e
mo
t
i
o
n
p
l
an
n
i
n
g
o
f
au
t
o
n
o
m
o
u
s
ro
b
o
t
s
i
n
u
n
k
n
o
w
n
en
v
i
r
o
n
me
n
t
s
,”
E
i
g
h
t
h
In
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
e
r
en
ce
o
n
G
r
a
p
h
i
c
a
n
d
Im
a
g
e
P
r
o
ce
s
s
i
n
g
(ICG
IP
2
0
1
6
)
,
v
ol
.
1
0
2
2
5
,
p
p.
1
-
6
,
2
0
1
7
.
[2
8
]
F.
Mart
í
n
ez,
et
a
l
.
,
“
A
D
at
a
-
D
ri
v
en
Pa
t
h
P
l
an
n
er
fo
r
S
mal
l
A
u
t
o
n
o
mo
u
s
R
o
b
o
t
s
U
s
i
n
g
D
ee
p
Reg
re
s
s
i
o
n
Mo
d
el
s
,”
In
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
er
e
n
ce
o
n
D
a
t
a
M
i
n
i
n
g
a
n
d
B
i
g
D
a
t
a
(
D
M
B
D
2
0
1
8
)
,
p
p.
5
9
6
-
6
0
3
,
2
0
1
8
.
[2
9
]
F.
Mart
í
n
ez,
et
a
l
.
,
“
A
n
A
l
g
o
r
i
t
h
m
Bas
e
d
o
n
t
h
e
Bact
er
i
al
Sw
arm
a
n
d
I
t
s
A
p
p
l
i
cat
i
o
n
i
n
A
u
t
o
n
o
mo
u
s
N
av
i
g
a
t
i
o
n
Pro
b
l
ems
,”
In
t
er
n
a
t
i
o
n
a
l
Co
n
f
e
r
en
ce
o
n
S
w
a
r
m
I
n
t
el
l
i
g
e
n
ce
(ICS
I
2
0
1
8
)
,
p
p
.
3
0
4
-
3
1
3
,
2
0
1
8
.
[3
0
]
J.
G
o
rd
i
l
l
o
an
d
F.
Mart
í
n
ez
,
“
Fu
zz
y
Co
n
t
ro
l
fo
r
Bi
p
ed
al
R
o
b
o
t
Co
n
s
i
d
er
i
n
g
E
n
er
g
y
Ba
l
an
ce
,”
Co
n
t
e
m
p
o
r
a
r
y
E
n
g
i
n
eer
i
n
g
S
ci
e
n
ces
,
v
o
l
.
11
,
n
o
.
39
,
p
p
.
1
9
4
5
-
1
9
5
2
,
J
a
n
2
0
1
8
.
[3
1
]
O
.
A
rri
a
g
a,
et
al
.
,
“
Real
-
t
i
m
e
co
n
v
o
l
u
t
i
o
n
al
n
e
u
r
al
n
et
w
o
rk
s
fo
r
em
o
t
i
o
n
a
n
d
g
en
d
er
cl
a
s
s
i
fi
ca
t
i
o
n
,”
arX
i
v
:
1
7
1
0
.
0
7
5
5
7
,
arX
i
v
.
o
r
g
,
2
0
1
7
.
Evaluation Warning : The document was created with Spire.PDF for Python.