I
AE
S
I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
Ar
t
if
icial
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
Vol.
14
,
No.
1
,
F
e
br
ua
r
y
2025
,
pp.
1
93
~
203
I
S
S
N:
2252
-
8938
,
DO
I
:
10
.
11591/i
jai
.
v
14
.i
1
.
pp
19
3
-
203
193
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
ai
.
iaes
c
or
e
.
c
om
Im
itat
io
n
of
t
h
e
h
u
m
a
n
u
p
p
e
r
l
imb
b
y c
o
n
vol
u
t
io
n
al
n
e
u
r
al
n
e
t
w
or
k
s
P
au
la
Us
e
c
h
e
M
u
r
il
lo
1
,
Robi
n
s
on
Ji
m
é
n
e
z
-
M
or
e
n
o
1
,
Javier
E
d
u
ar
d
o
M
ar
t
ín
e
z
B
aq
u
e
r
o
2
1
D
ep
ar
t
men
t
o
f
Mech
a
t
ro
n
i
c
E
n
g
i
n
e
eri
n
g
,
U
n
i
v
ers
i
d
a
d
Mi
l
i
t
ar
N
u
ev
a
G
ran
a
d
a,
Bo
g
o
t
á
,
Co
l
o
m
b
i
a
2
Facu
l
t
y
o
f
Ba
s
i
c
Sci
e
n
ces
a
n
d
E
n
g
i
n
eer
i
n
g
,
U
n
i
v
ers
i
d
a
d
d
e
l
o
s
L
l
an
o
s
,
V
i
l
l
av
i
cen
c
i
o
,
Co
l
o
m
b
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
Apr
3
,
2024
R
e
vis
e
d
J
un
28
,
2024
Ac
c
e
pted
J
ul
26
,
2024
T
h
e
p
ap
er
o
u
t
l
i
n
es
t
h
e
d
e
v
el
o
p
me
n
t
o
f
an
al
g
o
ri
t
h
m
fo
cu
s
ed
o
n
i
m
i
t
a
t
i
n
g
mo
v
eme
n
t
s
o
f
a
h
u
man
arm
an
d
rep
l
i
cat
i
n
g
s
t
ro
k
es
g
e
n
erat
e
d
b
y
t
h
e
u
s
er'
s
h
an
d
w
i
t
h
i
n
a
w
o
r
k
i
n
g
e
n
v
i
ro
n
men
t
.
T
h
e
a
l
g
o
ri
t
h
m
w
as
craft
ed
t
o
d
i
s
cer
n
t
h
e
p
o
s
i
t
i
o
n
o
f
ei
t
h
er
t
h
e
u
s
er'
s
l
eft
o
r
ri
g
h
t
arm,
t
rack
i
n
g
each
s
ect
i
o
n
(fi
n
g
er
s
,
w
ri
s
t
,
el
b
o
w
,
an
d
s
h
o
u
l
d
er)
t
h
r
o
u
g
h
a
d
et
ec
t
i
o
n
an
d
t
ra
ck
i
n
g
s
y
s
t
em.
T
h
es
e
mo
v
eme
n
t
s
are
t
h
en
re
p
l
i
cat
e
d
o
n
t
o
a
v
i
rt
u
al
arm,
s
i
m
u
l
at
i
n
g
t
h
e
ac
t
i
o
n
s
o
f
a
cu
t
t
i
n
g
t
o
o
l
,
g
e
n
erat
i
n
g
s
t
r
o
k
e
s
as
i
t
mo
v
es
.
Co
n
v
o
l
u
t
i
o
n
a
l
n
eu
ra
l
n
et
w
o
r
k
s
(CN
N
s
)
w
ere
emp
l
o
y
ed
t
o
d
et
ec
t
an
d
cl
a
s
s
i
fy
each
arm
s
ect
i
o
n
,
w
h
i
l
e
g
eo
me
t
ri
c
a
n
al
y
s
i
s
d
et
erm
i
n
e
d
t
h
e
ro
t
at
i
o
n
an
g
l
e
s
o
f
e
ach
j
o
i
n
t
,
faci
l
i
t
at
i
n
g
t
h
e
v
i
rt
u
al
ro
b
o
t
's
mo
t
i
o
n
.
T
h
e
s
t
r
o
k
e
rep
l
i
ca
t
i
o
n
p
r
o
g
ra
m
ach
i
ev
e
d
an
8
4
.
2
%
accu
racy
i
n
s
t
ro
k
e
ex
ecu
t
i
o
n
,
g
au
g
ed
b
y
t
h
e
cl
o
s
u
re
o
f
t
h
e
p
o
l
y
g
o
n
,
d
i
s
t
a
n
ce
b
et
w
een
i
n
i
t
i
a
l
an
d
fi
n
a
l
d
ra
w
i
n
g
p
o
i
n
t
s
,
an
d
g
en
era
t
e
d
n
o
i
s
e,
w
h
i
c
h
w
a
s
u
n
d
er
1
0
%
,
w
i
t
h
a
9
9
%
p
r
o
b
a
b
i
l
i
t
y
o
f
d
ra
w
i
n
g
a
cl
o
s
ed
p
o
l
y
g
o
n
.
A
Fas
t
reg
i
o
n
-
b
as
e
d
co
n
v
o
l
u
t
i
o
n
a
l
n
eu
r
al
n
e
t
w
o
rk
(
Fas
t
R
-
CN
N
)
n
et
w
o
r
k
d
et
ec
t
e
d
each
arm
s
ect
i
o
n
w
i
t
h
6
0
.
2
%
accu
racy
,
p
r
o
d
u
ci
n
g
d
et
ect
i
o
n
b
o
x
es
w
i
t
h
p
reci
s
i
o
n
ran
g
i
n
g
fro
m
1
7
%
t
o
5
9
%
.
A
n
y
reco
g
n
i
t
i
o
n
s
h
o
rt
c
o
mi
n
g
s
w
ere
ad
d
re
s
s
e
d
t
h
ro
u
g
h
mat
h
ema
t
i
ca
l
es
t
i
mat
i
o
n
o
f
mi
s
s
i
n
g
p
o
i
n
t
s
an
d
n
o
i
s
e
fi
l
t
er
s
,
res
u
l
t
i
n
g
i
n
a
9
0
.
4
%
i
mi
t
at
i
o
n
rat
e
o
f
h
u
ma
n
u
p
p
er
l
i
mb
mo
v
eme
n
t
.
K
e
y
w
o
r
d
s
:
C
onvolut
ional
ne
ur
a
l
ne
twor
ks
F
a
s
t
R
-
C
NN
Ge
s
tur
e
s
r
e
c
ognit
ion
s
ys
tem
human
mot
ions
M
oti
on
c
a
ptur
e
da
ta
R
e
a
l
-
ti
me
im
it
a
ti
on
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
:
R
obins
on
J
im
é
ne
z
-
M
or
e
no
M
e
c
ha
tr
onics
E
nginee
r
ing
P
r
ogr
a
m
,
F
a
c
ult
y
of
E
n
ginee
r
ing,
Unive
r
s
idad
M
il
it
a
r
Nue
va
Gr
a
na
da
C
a
r
r
e
r
a
11
#
101
-
80,
B
ogotá
D.
C
.
,
C
olom
bia
E
mail:
r
obins
on.
ji
mene
z
@unim
il
it
a
r
.
e
du.
c
o
1.
I
NT
RODU
C
T
I
ON
R
e
s
e
a
r
c
h
or
iente
d
to
the
de
tec
ti
on
a
nd
t
r
a
c
king
of
the
human
body
[
1]
a
ll
ows
e
xplor
ing
a
ppli
c
a
ti
ons
a
s
s
oc
iate
d
with
im
pr
oving
the
li
v
ing
c
ondit
ions
of
pe
ople,
f
or
e
xa
mpl
e
,
with
the
moni
to
r
ing
o
f
pa
in
t
he
r
a
pies
thr
ough
yoga
a
s
mentioned
in
[
2]
.
Anothe
r
a
ppr
oa
c
h
to
body
tr
a
c
king
is
in
human
mac
hine
int
e
r
a
c
ti
on
to
a
void
c
oll
is
ions
be
twe
e
n
thes
e
[
3]
.
Als
o
by
pe
r
f
or
mi
ng
s
p
e
c
if
ic
identif
ica
ti
ons
s
uc
h
a
s
ha
nd
a
nd
e
lbow,
dr
on
e
c
ontr
ol
c
a
n
be
pe
r
f
or
med
a
s
dis
c
us
s
e
d
in
[
4]
.
As
pe
c
t
th
a
t
a
t
the
s
pe
c
if
ic
leve
l
of
pa
tt
e
r
n
lea
r
ning
a
ll
ows
to
ge
ne
r
a
te
tr
a
jec
tor
y
tr
a
c
king
[
5]
.
Now
a
da
ys
,
pa
tt
e
r
n
lea
r
ning
a
nd/or
r
e
inf
or
c
e
ment
lea
r
ning
tec
hniques
a
r
e
us
e
d
by
de
e
p
lea
r
ning
a
lgor
it
hms
to
tr
a
c
k,
f
or
e
xa
mpl
e
,
c
onvolut
ional
ne
u
r
a
l
ne
twor
ks
(
C
NN
)
[
6
]
a
ppli
e
d
to
t
r
a
c
king
mobi
le
s
ys
tems
unde
r
the
c
onc
e
pt
of
int
e
r
ne
t
o
f
thi
ngs
[
7]
or
s
a
tel
li
te
in
f
or
mation
[
8
]
.
F
o
r
the
topi
c
of
int
e
r
e
s
t
C
NN
s
c
a
n
be
us
e
d
to
tr
a
c
k
gr
oups
of
pe
ople
[
9
]
or
a
va
r
ied
g
r
o
up
of
tar
ge
ts
[
10]
.
W
ha
t
is
im
po
r
tant
in
thi
s
c
a
s
e
is
to
ha
ve
one
or
s
e
ve
r
a
l
tr
a
c
king
tar
ge
ts
a
nd
to
tr
a
in
the
C
N
N
a
de
qua
tely
with
s
uf
f
icie
nt
da
ta
a
s
de
s
c
r
ibed
in
[
11]
a
nd
r
obus
t
e
nough
to
s
uppor
t
oc
c
lus
ions
[
12]
.
I
f
r
e
quir
e
d,
many
tr
a
c
king
s
ys
tems
ha
ve
de
pth
c
a
mer
a
s
f
or
dis
tanc
e
-
b
a
s
e
d
tr
a
c
king
[
13]
,
[
14]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
1
,
F
e
br
ua
r
y
2025
:
19
3
-
203
194
As
pe
c
ts
s
uc
h
a
s
the
t
r
a
c
king
o
f
ge
ometr
ies
f
or
r
a
il
wa
y
s
a
f
e
ty
by
mea
ns
of
C
NN
s
[
15
]
a
nd
the
t
r
a
c
king
of
body
pa
r
ts
s
uc
h
a
s
the
e
ye
[
16]
,
s
how
the
potential
of
C
NN
s
in
tr
a
c
king
tas
ks
.
I
n
thi
s
wor
k,
th
e
us
e
of
c
onvolut
ional
ne
twor
ks
to
identif
y
a
nd
tr
a
c
k
pa
r
ts
of
the
human
a
r
m
to
r
e
pli
c
a
te
it
s
moveme
nt
in
a
r
oboti
c
manipulator
to
pe
r
f
o
r
m
s
pe
c
if
ic
c
utt
ing
ge
ometr
i
e
s
or
iente
d
to
indus
tr
ial
wor
k
f
or
e
xa
mpl
e
s
tee
l
s
he
e
ts
or
s
im
il
a
r
is
e
xpos
e
d.
T
he
tr
a
c
king
of
the
s
houlder
,
e
lbow
a
nd
ha
nd
to
r
e
pli
c
a
te
a
pa
r
ti
c
ular
ge
ometr
y
with
the
r
oboti
c
manipulator
us
ing
c
onvolut
ional
ne
twor
ks
by
r
e
gions
,
to
know
it
s
s
pa
ti
a
l
loca
ti
on
a
nd
matc
h
it
with
that
of
the
r
obo
t,
thus
pr
opos
ing
the
ne
twor
k
a
r
c
hi
tec
tur
e
a
nd
the
tr
a
c
king
method
.
Ne
two
r
ks
by
r
e
gi
ons
s
uc
h
a
s
the
F
a
s
ter
r
e
gion
-
ba
s
e
d
c
onvolut
ional
ne
ur
a
l
ne
twor
k
(
R
-
C
NN
)
[
17]
a
r
e
us
e
d
f
or
objec
t
c
las
s
if
ica
ti
on
a
nd
loca
li
z
a
ti
on,
with
a
ppli
c
a
ti
ons
s
uc
h
a
s
r
oa
d
s
a
f
e
ty
[
18]
,
objec
t
c
oll
e
c
ti
on
by
r
obo
ts
[
19]
,
a
mong
o
th
e
r
s
[
20]
.
T
his
a
ll
ows
the
us
e
of
thes
e
ne
twor
ks
f
or
tr
a
c
king
objec
ts
in
c
ompl
e
x
e
nvir
onments
whe
r
e
t
he
y
ne
e
d
to
be
loca
ted.
T
he
f
o
ll
owing
is
the
de
ve
lopm
e
nt
of
a
n
a
ppli
c
a
ti
on
f
oc
us
e
d
on
r
e
c
ognizing
the
pos
it
ion
of
a
us
e
r
's
a
r
m
in
r
e
a
l
ti
me
,
by
mea
ns
of
a
n
r
e
d,
gr
e
e
n
a
nd
blue
(
R
GB
)
s
upe
r
vis
or
y
c
a
mer
a
a
nd
C
NN
-
ba
s
e
d
pa
tt
e
r
n
r
e
c
ognit
ion
methods
.
T
he
objec
ti
v
e
is
to
ge
ne
r
a
te
a
t
r
a
c
king
s
ys
tem
in
or
de
r
to
move
a
vir
tual
r
oboti
c
manipulator
r
e
s
pons
ibl
e
f
or
mi
m
icking
the
movem
e
nts
of
the
us
e
r
,
a
nd
t
r
a
c
e
the
pa
th
of
it
s
e
nd
e
f
f
e
c
t
or
a
s
if
it
we
r
e
a
c
utt
ing
tool
.
T
his
a
lgor
it
hm
p
r
e
s
e
nts
a
n
innovative
method
of
r
e
c
o
gnizing
a
nd
tr
a
c
king
the
m
ove
ment
of
the
human
uppe
r
li
mb
,
by
us
ing
F
a
s
t
R
-
C
NN
to
e
xtr
a
c
t
r
e
leva
nt
s
e
c
ti
ons
of
the
a
r
m
s
uc
h
a
s
the
s
houlder
,
e
lbow,
wr
is
t
a
nd
f
inger
s
,
a
nd
f
r
om
their
r
e
lative
po
s
it
ions
,
c
ontr
ol
the
r
otation
a
ngles
of
e
a
c
h
joi
nt
of
t
he
r
obot
to
a
c
hieve
the
de
s
ir
e
d
dis
plac
e
ment,
in
a
ddit
ion
to
e
s
ti
mating
the
mi
s
s
ing
point
s
in
the
de
tec
ti
on
of
the
ne
twor
k
to
s
olve
it
s
low
pe
r
c
e
ntage
o
f
a
c
c
ur
a
c
y.
T
he
a
r
ti
c
le
is
divi
de
d
in
to
f
ive
s
e
c
ti
ons
:
in
the
f
ir
s
t
s
e
c
ti
on,
a
br
ief
int
r
oduc
ti
on
to
human
movem
e
nt
im
it
a
ti
on
s
ys
tems
,
the
us
e
o
f
C
NN
s
a
nd
thei
r
d
if
f
e
r
e
nt
va
r
iations
a
nd
a
ppli
c
a
ti
ons
is
given.
I
n
the
s
e
c
ond
s
e
c
ti
on,
the
ope
r
a
ti
on
of
the
a
lgo
r
it
hm
f
r
om
the
f
ir
s
t
c
ontac
t
of
the
us
e
r
with
the
int
e
r
f
a
c
e
to
the
c
omp
letion
of
the
pr
ogr
a
m
with
the
f
ull
y
r
e
p
li
c
a
ted
s
tr
oke
is
e
xplaine
d.
I
n
the
thi
r
d
s
e
c
ti
on,
the
a
r
c
hit
e
c
tur
e
of
the
F
a
s
t
R
-
C
N
N
tr
a
ined
f
or
f
e
a
tur
e
e
xtr
a
c
ti
on
is
p
r
e
s
e
nted.
2.
M
E
T
HO
D
Ne
xt,
the
methodology
that
a
ll
ows
obtaining
the
a
lgor
it
hm
of
im
it
a
ti
on
of
moveme
nts
o
f
the
human
uppe
r
li
mb
,
whos
e
pur
pos
e
is
to
r
e
pli
c
a
te
the
a
c
tu
a
l
pos
it
ion
of
the
a
r
m
dur
ing
a
video
s
hot,
a
nd
to
ge
ne
r
a
te
s
tr
oke
s
with
the
f
inal
e
f
f
e
c
tor
,
s
im
ulating
a
c
utt
ing
tool
,
is
pr
e
s
e
nted.
T
he
de
ve
loped
a
lgor
it
hm
is
r
e
s
pons
ibl
e
f
or
de
tec
ti
ng
a
nd
tr
a
c
king
the
loca
ti
on
of
th
e
r
e
leva
nt
r
e
gions
of
the
a
r
m
,
s
uc
h
a
s
the
s
houlder
,
e
lb
ow,
wr
is
t
a
nd
f
inger
ti
ps
.
F
or
t
his
pu
r
pos
e
,
F
a
s
t
R
-
C
NN
a
nd
R
GB
s
upe
r
vis
ing
c
a
mer
a
a
r
e
us
e
d,
in
o
r
de
r
to
e
xtr
a
c
t
a
r
e
lations
hip
of
a
ngles
be
twe
e
n
e
a
c
h
of
the
r
e
gions
f
ound
a
nd
thus
e
xtr
a
c
t
the
de
gr
e
e
s
o
f
r
otation
o
f
e
a
c
h
joi
nt
of
a
vi
r
tual
r
obo
t
to
ge
ne
r
a
te
the
i
mi
tation
o
f
the
m
ove
ment.
S
im
ult
a
ne
ous
ly,
the
e
nd
e
f
f
e
c
to
r
of
the
r
obot
ge
n
e
r
a
tes
a
s
e
r
ies
of
s
tr
oke
s
dur
ing
it
s
dis
plac
e
ment,
whic
h
r
e
pr
e
s
e
nt
the
c
ut
made
by
a
too
l
that
would
be
loca
ted
a
t
it
s
f
a
r
e
nd
.
T
he
s
e
que
nc
e
of
ope
r
a
ti
on
of
the
a
lgor
it
hm
is
s
hown
in
the
diagr
a
m
in
F
ig
ur
e
1,
whe
r
e
e
a
c
h
s
tep
wa
s
de
mar
c
a
ted
with
number
s
f
r
om
0
to
4
a
nd
dif
f
e
r
e
nt
f
unc
ti
ons
we
r
e
us
e
d
f
or
e
a
c
h
of
the
tas
ks
r
e
quir
e
d
by
the
pr
ogr
a
m
,
whic
h
a
r
e
e
xplaine
d
in
e
a
c
h
of
the
f
oll
owing
s
tage
s
.
T
he
pr
ogr
a
m
int
e
r
f
a
c
e
is
pr
e
s
e
nted
in
F
igu
r
e
2.
2.
1
.
S
t
e
p
0:
I
n
it
ial
iza
t
ion
o
f
v
ar
iab
les
T
he
de
ve
lopm
e
nt
r
e
quir
e
d
the
us
e
of
a
n
R
GB
c
a
mer
a
to
c
a
ptur
e
the
wor
king
e
nvir
onment
a
nd
the
input
im
a
ge
wa
s
r
e
s
ize
d
f
r
om
480x640
pixels
to
32
0x427
pixels
,
in
o
r
de
r
to
s
pe
e
d
up
the
a
r
m
pos
it
ion
de
tec
ti
on
pr
oc
e
s
s
by
r
e
duc
ing
the
number
o
f
pixels
to
be
e
v
a
luate
d.
T
o
ge
ne
r
a
te
the
F
a
s
t
R
-
C
NN
tr
a
ini
ng
da
taba
s
e
,
the
e
leme
nts
of
the
wor
king
e
nvir
onment
we
r
e
va
r
ie
d
a
nd
the
ba
r
e
a
r
m
wa
s
c
a
ptu
r
e
d
f
r
om
the
s
hould
e
r
to
the
f
inger
ti
ps
,
a
s
the
e
xa
mpl
e
s
hown
in
F
igu
r
e
3,
whe
r
e
it
is
pos
s
ibl
e
to
obs
e
r
ve
the
R
OI
s
mar
ke
d
on
e
a
c
h
s
e
c
ti
on
of
the
a
r
m
(
s
houlder
,
e
lbow
,
wr
is
t,
f
inger
s
).
As
s
hown
in
F
igur
e
3
,
the
a
r
m
mus
t
a
lwa
ys
e
nter
f
r
o
m
the
r
ight
s
ide
of
the
im
a
ge
,
a
nd
the
ha
nd
mus
t
ha
ve
it
s
f
inge
r
s
f
ull
y
e
xtende
d
in
or
de
r
to
de
tec
t
the
f
inger
s
(
F
in
ge
r
s
)
.
F
igur
e
1.
Algor
i
thm
f
lowc
ha
r
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
I
mitat
ion
of
the
human
uppe
r
li
mb
by
c
onv
olut
iona
l
ne
ur
al
ne
tw
or
k
s
(
P
aula
Us
e
c
he
)
195
F
igur
e
2.
R
e
a
l
a
nd
vi
r
tual
pos
it
ion
o
f
the
a
r
m
withi
n
the
int
e
r
f
a
c
e
(
GU
I
)
F
igur
e
3.
Ar
m
c
a
ptur
e
with
labe
ls
s
houlder
(
blue
)
,
e
lbow
(
ye
ll
ow)
,
wr
is
t
(
gr
e
e
n)
,
a
nd
f
inger
s
(
r
e
d)
T
he
f
i
r
s
t
s
tep
of
the
a
lgo
r
it
hm
c
ons
is
ts
of
ini
ti
a
li
z
in
g
the
va
r
iable
s
to
be
us
e
d
th
r
oughout
the
pr
og
r
a
m,
whic
h
a
r
e
li
s
ted
in
T
a
ble
1,
whe
r
e
thei
r
f
unc
ti
o
n,
na
me
a
nd
ini
ti
a
li
z
a
ti
on
va
lue
a
r
e
mentioned,
with
ti
m
e
L
be
ing
the
c
ompl
e
ti
on
ti
me
e
nte
r
e
d
by
the
us
e
r
in
the
gr
a
phica
l
int
e
r
f
a
c
e
.
B
oth
the
length
of
the
li
nks
(
L
)
a
nd
the
a
ngles
of
the
r
obot
(
Q)
we
r
e
take
n
in
or
de
r
f
r
om
t
he
point
a
nc
hor
e
d
to
the
gr
ound
(
e
quivale
nt
to
s
h
oulder
in
F
igur
e
3)
,
to
the
e
nd
e
f
f
e
c
tor
(
e
quivale
nt
to
f
inger
s
in
F
igur
e
3)
.
T
he
va
lues
of
L
we
r
e
take
n
a
s
a
n
a
ve
r
a
ge
of
the
lengths
of
a
n
a
dult
a
r
m
a
s
mea
s
ur
e
d
e
xpe
r
im
e
ntally
with
dif
f
e
r
e
nt
tes
t
s
ubjec
ts
.
I
n
a
ddit
ion
to
the
va
r
iable
s
men
ti
one
d
in
T
a
ble
1
,
the
c
a
mer
a
,
the
vi
r
tual
e
n
vir
onment,
a
nd
the
F
a
s
t
R
-
C
NN
tr
a
ined
f
o
r
the
a
lgor
it
hm
(
c
a
ll
e
d
de
tec
tor
)
a
r
e
ini
ti
a
li
z
e
d.
T
a
ble
1.
V
a
r
ibale
i
nit
ial
iza
ti
on
V
a
r
ia
bl
e
F
unt
io
n
V
a
lu
e
L
L
e
ngt
h of
t
he
l
in
ks
of
t
he
vi
r
tu
a
l
ma
ni
pul
a
to
r
[
L
(
1
)
L
(
2
)
L
(
3
)
]
(
me
te
r
s
)
[
0.36 0.3 0.2]
Q
A
ngl
e
s
of
t
he
j
oi
nt
s
of
t
he
vi
r
tu
a
l
ma
ni
pul
a
to
r
[
Q
(
1)
Q
(
2
)
Q
(
3)
]
(
r
a
di
a
ns
)
[
0 0 0]
C
ount
T
im
e
T
im
e
r
t
o de
te
r
mi
ne
t
he
e
nd of
t
he
pr
ogr
a
m (
s
e
c
onds
)
0
ti
me
L
T
im
e
l
im
it
f
or
pr
ogr
a
m c
ompl
e
ti
on
U
s
e
r
-
de
f
in
e
d
ite
I
te
r
a
to
r
f
or
pl
o
tt
in
g t
he
t
r
a
c
e
ge
ne
r
a
te
d by the
r
obot
1
numN
N
umbe
r
of
i
nt
e
r
me
di
a
te
poi
nt
s
be
twe
e
n a
n e
nd e
f
f
e
c
to
r
poi
nt
a
nd t
he
ne
xt
one
4
P
A
r
r
a
y c
ont
a
in
in
g t
he
pos
it
io
n of
t
he
e
nd e
f
f
e
c
to
r
a
t
e
a
c
h de
te
c
ti
on [
P
(
it
e
,1)
P
(
it
e
,2)
P
(
it
e
,3)
]
--
boxe
s
D
e
te
c
ti
on f
r
a
me
s
on t
he
i
nput
i
ma
ge
[
x y width height]
(
pi
xe
ls
)
--
[x
c
y
c
]
G
e
ome
tr
ic
c
e
nt
e
r
of
t
he
boxe
s
--
D
im
I
nput
i
ma
ge
di
me
ns
io
ns
[
r
ow
c
ol
umn]
.
(
pi
xe
ls
)
[
320 427]
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
1
,
F
e
br
ua
r
y
2025
:
19
3
-
203
196
2.
2
.
S
t
e
p
1
:
Wait
f
or
t
h
e
ar
m
Dur
ing
the
pr
oc
e
s
s
of
de
tec
ti
on
a
nd
tr
a
c
king
of
the
uppe
r
li
mb,
the
s
e
c
ti
ons
f
ound
a
nd/or
e
s
ti
mate
d
a
r
e
plot
ted
in
"
Us
e
r
's
Ar
m"
,
a
s
s
hown
in
F
igur
e
4
(
whe
r
e
e
a
c
h
c
olor
c
o
r
r
e
s
ponds
to
the
labe
ls
in
F
ig
ur
e
3)
,
in
or
de
r
to
give
f
e
e
dba
c
k
to
the
us
e
r
a
bout
the
r
e
c
ognit
ion
pr
oc
e
s
s
,
a
nd
the
c
ur
r
e
nt
loca
ti
on
o
f
t
he
a
r
m.
S
ubs
e
que
ntl
y,
the
a
ngles
f
or
the
moveme
nt
of
the
vir
tual
r
obo
t
a
r
e
c
a
lcula
ted
us
ing
the
"
Ar
mAngles
"
,
whic
h
is
e
xplaine
d
be
low.
T
he
di
r
e
c
t
kinema
ti
c
s
o
f
the
m
a
nipul
a
tor
is
us
e
d
to
obtain
the
pos
it
ion
of
the
e
nd
e
f
f
e
c
tor
a
c
c
or
ding
to
the
a
ngles
f
ou
nd,
s
tor
ing
in
the
matr
i
x
P
(
i
te
,
:)
that
point
.
F
inally,
ite
is
incr
e
a
s
e
d
by
one
(
in
e
a
c
h
it
e
r
a
ti
on)
a
nd
S
tep
3
is
e
xe
c
uted.
F
igur
e
4.
De
tec
ti
on
of
a
r
m
s
e
c
ti
ons
2.
3
.
S
t
e
p
2
:
Ob
t
ain
in
g
a
n
gles
I
n
the
s
e
c
ond
s
tep,
the
de
tec
tor
is
us
e
d
to
obtain
t
he
a
r
m
s
e
c
ti
ons
in
e
a
c
h
i
ter
a
ti
on
of
the
"
while"
of
F
igur
e
1
a
nd
the
f
unc
ti
ons
"
uniqB
oxe
s
"
a
nd
"
Ang
ulos
B
r
a
z
z
o"
a
r
e
a
ppli
e
d
to
s
e
lec
t
the
de
tec
ti
on
bo
xe
s
with
the
highes
t
pe
r
c
e
ntage
of
c
onf
idenc
e
in
c
a
s
e
of
de
tec
ti
ng
mor
e
than
one
e
leme
nt
pe
r
c
a
tegor
y,
a
nd
to
c
onve
r
t
the
inclination
a
ngles
of
the
f
inger
s
,
w
r
is
t,
e
lbow
a
nd
s
houlder
point
s
,
ge
ne
r
a
ted
on
the
im
a
ge
,
int
o
r
otation
a
ngles
f
or
the
joi
nts
of
the
manipulator
.
T
he
de
tec
tor
r
e
c
e
ives
a
s
input
the
im
a
ge
c
a
ptur
e
d
by
the
c
a
mer
a
a
nd
ge
ne
r
a
tes
a
s
output
the
c
la
s
s
if
ica
ti
on
of
the
s
e
c
ti
o
ns
(
labe
ls
)
,
the
pe
r
c
e
ntage
of
c
onf
idenc
e
with
wh
ich
e
a
c
h
e
leme
nt
wa
s
c
las
s
if
ied
(
s
c
or
e
s
)
a
nd
the
de
tec
ti
on
bo
xe
s
ge
ne
r
a
ted
on
the
im
a
ge
,
whe
r
e
the
a
r
m
s
e
c
ti
ons
(
boxe
s
)
we
r
e
f
ound.
T
he
"
uniqB
oxe
s
"
f
unc
ti
on
e
va
luate
s
t
he
number
o
f
d
e
tec
ti
on
boxe
s
ge
ne
r
a
ted
pe
r
c
a
tego
r
y,
a
nd
if
mor
e
than
one
is
f
ound,
it
c
ompar
e
s
the
s
c
or
e
s
of
t
he
s
e
boxe
s
a
nd
s
e
lec
ts
the
one
with
the
highes
t
pe
r
c
e
ntage
.
A
human
c
e
nter
e
d
mot
ion
de
tec
ti
on
(
HC
M
D
)
m
a
tr
ix
is
c
a
lcula
ted
to
s
tor
e
the
c
oo
r
dinate
s
of
the
ge
ome
tr
ic
c
e
nter
of
the
de
tec
ti
on
boxe
s
o
f
e
a
c
h
c
a
t
e
gor
y,
whe
r
e
e
a
c
h
r
ow
of
the
mat
r
ix
c
ontains
the
p
os
it
ions
[
xc
yc
]
o
f
the
s
houlder
,
e
lbow,
wr
is
t
,
a
nd
f
inge
r
s
s
e
c
t
ions
,
in
that
o
r
de
r
,
e
a
c
h
s
e
c
ti
on
be
ing
a
d
if
f
e
r
e
nt
r
ow
,
number
e
d
f
r
om
1
to
4
.
T
he
c
oo
r
dinate
s
[
xc
yc
]
a
r
e
obtaine
d
a
s
s
hown
in
(
1
)
a
nd
(
2)
,
yc
be
ing
the
ve
r
ti
c
a
l
c
omponent
a
nd
xc
the
ho
r
izonta
l
one
.
E
a
c
h
r
ow
o
f
boxe
s
is
c
ompos
e
d
o
f
the
e
leme
nts
pr
e
vious
ly
ind
ica
ted
in
T
a
ble
1
,
a
nd
a
s
in
HC
M
D,
it
s
r
ows
c
or
r
e
s
pond
to
e
a
c
h
of
the
boxe
s
de
t
e
c
ted
in
the
c
las
s
if
ica
ti
on
c
a
tegor
ies
.
=
(
:
,
1
)
−
(
:
,
3
)
2
(
1)
=
(
:
,
2
)
+
(
:
,
4
)
2
(
2)
T
he
f
unc
ti
on
"
E
s
ti
mate
dAngle
s
"
s
e
e
ks
to
e
s
ti
mate
the
a
ngle
of
r
otation
of
the
point
whos
e
c
oor
dinate
s
of
the
ne
ighbor
ing
s
e
c
ti
ons
a
r
e
kno
wn
,
e
.
g.
,
the
c
a
s
e
whe
r
e
the
s
houlder
a
nd
wr
is
t
we
r
e
de
tec
ted,
bu
t
t
he
e
lbow
wa
s
not
(
P
C
1,
P
C
2
a
nd
P
e
f
or
F
igu
r
e
5
,
r
e
s
pe
c
ti
ve
ly)
.
F
o
r
whic
h,
it
is
r
e
qui
r
e
d
to
know
the
c
oor
dinat
e
s
of
the
ne
ighbor
ing
point
s
a
nd
the
a
nter
io
r
pos
it
ion
o
f
t
he
mi
s
s
ing
point
(
P
C
a
1
or
P
C
a
2
in
F
igu
r
e
5
).
W
it
h
thi
s
inf
or
mation,
a
tr
iangle
r
e
lation
is
made
a
s
s
hown
in
F
igur
e
5
,
whe
r
e
the
di
mens
ions
of
e
a
c
h
r
obot
li
nk
i
n
pixels
(
L
m
c
a
lcula
ted
in
(
3
)
a
s
s
umi
ng
that
the
a
r
m
s
pa
ns
the
e
nti
r
e
length
Dim(
2
)
of
the
im
a
ge
)
a
r
e
kn
own
a
nd,
the
r
e
f
or
e
,
the
int
e
r
na
l
a
ngles
of
the
tr
iangle
c
a
n
be
known,
us
ing
the
c
os
ine
theor
e
m
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
I
mitat
ion
of
the
human
uppe
r
li
mb
by
c
onv
olut
iona
l
ne
ur
al
ne
tw
or
k
s
(
P
aula
Us
e
c
he
)
197
F
igur
e
5.
R
a
ti
o
of
tr
iangle
s
a
pplyi
ng
the
c
os
ine
the
or
e
m
[
(
1
)
(
2
)
(
3
)
]
=
[
(
1
)
(
2
)
(
3
)
]
∗
(
2
)
(
1
)
+
(
2
)
+
(
3
)
(
3)
W
he
r
e
qd
is
the
a
ngle
be
twe
e
n
the
ve
r
ti
c
a
l
a
nd
the
li
ne
ge
ne
r
a
ted
be
twe
e
n
the
point
s
P
C
1=
[
P
C
1x
P
C
1y]
a
nd
P
C
2=
[
P
C
2x
P
C
2y]
,
c
a
ll
e
d
d,
whic
h
c
a
n
be
known
by
a
pplyi
ng
the
d
is
tanc
e
e
qua
ti
on
be
twe
e
n
two
p
oint
s
(
4)
.
P
e
is
the
point
to
be
e
s
ti
mate
d,
P
C
a
1
o
r
P
C
a
2
is
t
he
pr
e
vious
pos
it
ion
o
f
the
point
to
be
e
s
ti
mate
d,
whic
h
is
known,
L
m
a
r
e
the
lengths
of
the
r
obot
li
nks
in
pi
xe
ls
(
L
m(
1)
a
nd
L
m(
2)
f
o
r
the
c
a
s
e
in
whic
h
the
e
lbow
is
e
s
ti
mate
d,
L
m(
2)
a
nd
L
m(
3)
in
c
a
s
e
of
e
s
ti
mating
the
wr
is
t)
a
nd
β
is
the
a
ngle
be
twe
e
n
d
a
nd
L
m(
1)
,
whic
h
will
a
ll
ow
e
s
ti
mating
the
f
i
r
s
t
a
ngle
of
the
q
ve
c
tor
.
=
√
(
1
−
2
)
2
+
(
1
−
2
)
2
(
4)
C
a
lcula
te
β
with
the
c
os
ine
theor
e
m
a
s
s
hown
in
(
5)
,
a
nd
e
s
ti
mate
the
a
ngle
be
twe
e
n
P
e
a
nd
P
C
1
(
qe
1
a
nd
qe
2)
a
s
s
hown
in
(
6)
a
nd
(
7)
,
whe
r
e
qe
1
or
qe
2
is
s
e
lec
ted
a
s
the
e
s
ti
mate
d
a
ngle,
a
c
c
or
ding
to
it
s
dif
f
e
r
e
nc
e
with
r
e
s
pe
c
t
to
the
p
r
e
vious
a
ngle
f
o
r
t
h
a
t
li
nk,
c
hoos
ing
the
one
that
is
c
los
e
r
.
=
co
s
−
1
(
(
1
)
2
+
2
−
(
2
)
2
2
∗
(
1
)
∗
)
(
5)
1
=
−
(
6)
2
=
+
(
7)
Af
ter
e
s
ti
mating
the
f
ir
s
t
a
ngle
(
qe
)
,
we
p
r
oc
e
e
d
to
obtain
the
e
s
ti
mate
d
c
oor
dinate
s
of
the
m
is
s
ing
point
(
P
e
)
us
ing
(
8)
,
whe
r
e
h
is
1
f
or
the
c
a
s
e
in
whic
h
the
e
l
bow
is
e
s
ti
mate
d,
a
nd
2
f
o
r
e
s
ti
mating
the
w
r
is
t.
I
n
(
8)
wha
t
is
done
is
to
take
the
tr
iangle
f
o
r
med
by
L
m(
h
)
,
the
ve
r
ti
c
a
l
a
nd
a
ho
r
izonta
l
pa
s
s
ing
th
r
ough
P
e
a
nd
the
ve
r
ti
c
a
l,
a
nd
c
a
lcula
te
it
s
two
legs
,
in
or
de
r
to
a
dd
them
to
the
c
ur
r
e
nt
pos
it
ion
of
po
int
P
C
1
a
nd
thus
obtain
P
e
.
Onc
e
the
e
s
ti
mate
d
c
oor
dinate
s
of
the
m
is
s
ing
point
a
r
e
known,
P
e
is
s
tor
e
d
in
HC
M
D(
h+
1,
:)
,
c
onve
r
ti
ng
the
unknown
point
int
o
a
n
e
s
ti
mate
d
point
.
=
[
]
=
[
1
+
(
ℎ
)
∗
co
s
(
)
1
−
(
ℎ
)
∗
s
in
(
)
]
(
8)
I
n
or
de
r
to
s
ubtr
a
c
t
the
a
ngles
of
a
r
m
inclination
,
is
c
a
lcula
t
e
d
to
q(
1)
(
a
ngle
be
twe
e
n
L
(
1)
a
nd
the
ve
r
ti
c
a
l)
to
obtain
the
f
ir
s
t
e
s
ti
mate
d
a
ngle.
Af
ter
q
(
1)
,
the
e
s
ti
mate
d
pos
it
ion
of
e
lbow
is
c
a
lcula
ted
us
ing
(
9)
,
whe
r
e
a
r
otation
o
f
e
lbow
,
q(
1)
de
gr
e
e
s
with
r
e
s
pe
c
t
to
s
houlder
is
made
,
taking
a
s
ini
ti
a
l
pos
it
ion
of
e
lbow
a
point
jus
t
a
bove
s
houlder
,
a
s
s
umi
ng
that
the
dis
t
a
nc
e
both
is
L
m(
1)
,
a
s
s
hown
in
(
10
)
,
be
ing
R
otZ
the
r
otation
matr
ix
with
r
e
s
pe
c
t
to
Z
in
(
11)
,
in
HC
M
D
matr
ix
.
(
2
,
:
)
=
(
(
1
)
)
∗
(
(
2
,
:
)
−
(
1
,
:
)
)
+
(
1
,
:
)
(
9)
(
2
,
:
)
=
[
(
1
,
1
)
,
(
1
,
2
)
+
(
1
)
]
(
10)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
1
,
F
e
br
ua
r
y
2025
:
19
3
-
203
198
(
(
1
)
)
=
[
co
s
(
1
)
−
s
in
(
1
)
s
in
(
1
)
co
s
(
1
)
]
(
11)
Af
ter
obtaining
a
nd
e
s
ti
mating
the
la
r
ge
s
t
number
of
a
r
m
s
e
c
ti
ons
,
the
pos
it
ion
point
s
HC
M
D(
i,
:
)
a
nd
HC
M
D(
i+1,
:)
,
be
ing
i
the
i
ter
a
tor
of
a
f
o
r
f
r
om
on
e
to
thr
e
e
,
a
r
e
us
e
d
to
c
a
lcula
te
the
ti
lt
a
ngle
of
e
a
c
h
pa
ir
of
point
s
us
ing
the
tange
nt
a
s
s
hown
in
(
12)
,
whe
r
e
a
ll
a
ngles
s
tor
e
d
in
q
we
r
e
take
n
with
r
e
s
pe
c
t
t
o
the
ve
r
ti
c
a
l.
Onc
e
the
a
ngles
a
r
e
obtaine
d
(
e
it
he
r
c
a
lcula
ted
or
e
s
ti
mate
d)
,
they
a
r
e
a
djus
ted
to
be
take
n
with
r
e
s
pe
c
t
to
the
r
obot's
c
oor
dinate
a
xis
,
whe
r
e
q(
1)
is
take
n
with
r
e
s
pe
c
t
to
the
ve
r
ti
c
a
l,
q(
2)
with
r
e
s
pe
c
t
to
q(
1)
,
a
nd
q(
3)
with
r
e
s
pe
c
t
to
q(
2)
,
us
ing
(
13)
thr
ough
(
15
)
,
f
or
mi
ng
t
he
ve
c
tor
Q
with
the
a
ngles
ne
e
de
d
to
mi
mi
c
the
mot
ion
of
the
human
a
r
m
in
the
vir
tual
e
nvir
onment
.
F
inall
y,
a
nois
e
f
il
ter
wa
s
a
dde
d
to
a
void
a
br
upt
c
ha
ng
e
s
in
the
pos
it
ion
of
the
vi
r
tual
a
r
m
be
twe
e
n
e
a
c
h
it
e
r
a
ti
on,
de
nying
i
ts
moveme
nt
in
c
a
s
e
s
whe
r
e
the
va
r
iatio
n
of
the
a
ngles
be
twe
e
n
the
pr
e
vious
a
nd
the
c
ur
r
e
nt
it
e
r
a
ti
on
is
gr
e
a
ter
than
15º
,
whos
e
va
lue
wa
s
e
xpe
r
i
menta
ll
y
s
e
lec
ted
to
e
ns
ur
e
s
moot
h
moveme
nts
a
nd
a
void
e
r
r
one
ous
de
tec
ti
ons
a
lt
e
r
ing
the
polygonal
tr
ace.
(
)
=
ta
n
(
(
+
1
,
2
)
−
(
,
2
)
(
+
1
,
1
)
−
(
,
1
)
)
(
12)
(
1
)
=
(
1
)
(
13)
(
2
)
=
(
2
)
−
(
1
)
(
14)
(
3
)
=
(
3
)
−
(
1
)
−
(
2
)
(
15)
2.
4
.
S
t
e
p
3
:
Dr
aw
s
t
r
ok
e
T
o
dr
a
w
the
c
ut
li
ne
made
by
the
vi
r
tual
a
r
m
with
t
he
e
nd
e
f
f
e
c
tor
,
s
tr
a
ight
li
ne
s
we
r
e
dr
a
wn
be
twe
e
n
e
a
c
h
pa
ir
of
point
s
of
the
P
matr
ix
a
nd
the
s
tar
t
a
nd
e
nd
point
s
of
the
tr
a
jec
tor
y
we
r
e
c
ha
nge
d
to
g
r
e
e
n
a
nd
r
e
d
,
a
s
s
hows
F
igur
e
6
(
a
)
.
T
he
f
ir
s
t
joi
nt
,
a
nc
hor
e
d
to
th
e
gr
ound,
r
e
p
r
e
s
e
nts
the
s
houlder
,
the
s
e
c
ond
the
e
l
bow
a
nd
the
thi
r
d
the
w
r
is
t,
with
the
f
inal
e
f
f
e
c
tor
be
ing
th
e
us
e
r
's
f
inger
ti
ps
.
T
he
dr
a
wing
is
pe
r
f
or
med
du
r
i
ng
e
a
c
h
it
e
r
a
ti
on,
a
ll
owing
the
us
e
r
to
obs
e
r
ve
the
c
utt
ing
p
r
oc
e
s
s
a
s
he
/s
he
moves
the
a
r
m.
T
o
p
r
e
ve
nt
the
number
of
point
s
dr
a
wn
in
"
Vir
tual
r
obot
a
r
m"
f
r
om
a
lt
e
r
ing
the
e
xe
c
uti
on
ti
me
of
the
a
lgor
it
hm,
the
c
utt
ing
im
a
ge
wa
s
r
e
f
r
e
s
he
d
e
ve
r
y
100
point
s
,
a
nd
the
c
ompl
e
te
s
tr
oke
wa
s
dr
a
wn
a
t
the
e
nd
of
the
pr
ogr
a
m
.
T
he
Q
a
ngles
a
r
e
s
e
nt
dir
e
c
tl
y
to
the
r
obot's
joi
nts
to
move
it
a
r
ound
the
e
nvir
on
ment,
n
o
matter
whic
h
a
r
m
of
the
us
e
r
is
be
ing
r
e
c
ognize
d,
a
nd
if
a
ny
of
the
a
r
m
s
e
c
ti
ons
a
r
e
not
de
tec
ted,
the
mi
s
s
ing
a
ngle
is
r
e
plac
e
d
by
the
a
ngle
o
f
the
pr
e
vious
it
e
r
a
ti
on
.
A
F
a
s
t
R
-
C
NN
type
ne
twor
k
wa
s
u
s
e
d
f
or
the
de
tec
ti
on
a
nd
c
las
s
if
ica
ti
on
of
the
a
r
m
s
e
c
ti
ons
:
s
houlder
,
e
lbow,
w
r
is
t
,
a
nd
f
inger
s
,
whic
h
wa
s
tr
a
i
ne
d
with
a
tot
a
l
o
f
4900
tr
a
ini
ng
im
a
ge
s
,
f
or
250
e
poc
hs
,
a
M
ini
B
a
tchSi
z
e
of
49,
a
n
input
im
a
ge
of
320x427
p
ixels
a
nd
de
tec
ti
on
f
r
a
mes
of
dim
e
ns
ions
40x40
pixels
.
An
ini
ti
a
l
a
c
quis
it
ion
of
700
tr
a
ini
ng
im
a
ge
s
,
a
nd
10
0
tes
t
im
a
ge
s
we
r
e
made
,
whe
r
e
the
us
e
r
's
ge
nde
r
,
a
r
m
pos
it
ion,
he
ight
r
e
lative
to
the
c
a
mer
a
,
a
nd
wo
r
king
e
nvir
on
ment
we
r
e
va
r
ied,
a
nd
then
the
im
a
ge
s
we
r
e
a
ugme
nted
with
the
Da
ta
Augme
ntation
pr
og
r
a
m
f
r
om
[
21]
,
a
lt
e
r
in
g
the
li
ght
int
e
ns
it
y,
c
olor
s
,
a
nd
a
dding
nois
e
to
the
im
a
ge
.
T
he
a
r
c
hit
e
c
tur
e
f
or
the
ne
twor
k
C
NN
[
22
]
is
pr
e
s
e
nted
in
s
e
c
ti
on
3
(
Ne
twor
k
1)
,
whe
r
e
s
mall
a
nd
s
qua
r
e
f
il
ter
s
we
r
e
us
e
d
to
e
xt
r
a
c
t
the
de
tails
of
e
a
c
h
s
e
c
ti
on
of
the
a
r
m
,
a
nd
to
di
f
f
e
r
e
nti
a
te
e
a
c
h
c
a
t
e
gor
y
by
a
s
pe
c
ts
s
uc
h
a
s
the
f
inger
s
of
the
ha
nd,
the
c
ha
nge
of
a
r
e
a
a
t
the
w
r
is
t
be
twe
e
n
the
ha
nd
a
nd
the
a
r
m
,
t
he
f
olds
of
the
e
lbow,
a
nd
the
c
ontac
t
of
the
s
houlder
wit
h
the
c
lot
hing.
T
he
F
a
s
t
R
-
C
NN
us
e
d
(
Ar
m_F
a
s
t
R
C
NN
)
a
c
hieve
d
60.
2%
a
c
c
ur
a
c
y
f
or
a
tot
a
l
of
700
tes
t
im
a
ge
s
.
F
igur
e
6
(
b
)
s
hows
a
n
e
xa
mpl
e
of
de
tec
ti
on
a
nd
c
las
s
if
ica
ti
on
of
the
human
uppe
r
li
mb
whe
r
e
it
is
pos
s
ibl
e
to
obs
e
r
ve
the
de
tec
ti
on
of
e
a
c
h
of
the
s
e
c
ti
ons
of
the
a
r
m,
a
nd
the
manipulato
r
's
im
it
a
ti
on
o
f
the
r
e
c
ognize
d
pos
it
ion,
in
a
ddi
ti
on
to
the
c
ut
t
r
a
c
e
ge
ne
r
a
ted
f
r
om
the
pr
e
vious
pos
it
ion
to
the
c
ur
r
e
nt
one
.
2.
5
.
S
t
e
p
4
:
Calc
u
lat
io
n
of
t
i
m
e
I
n
e
a
c
h
it
e
r
a
ti
on
of
the
"
while"
in
F
igu
r
e
1,
it
is
e
va
luate
d
whe
ther
the
a
r
m
is
r
e
c
ognize
d
or
not
,
us
ing
a
c
ondit
ional
that
c
ounts
the
number
of
s
e
c
ti
ons
de
tec
ted
by
the
de
tec
tor
,
whe
r
e
a
c
ount
o
f
the
ti
me
that
the
a
ppli
c
a
ti
on
ha
s
be
e
n
e
xe
c
uted
is
s
tar
ted,
in
c
a
s
e
no
s
e
c
ti
on
is
f
ound.
I
f
the
a
r
m
is
not
de
tec
ted
w
it
hin
the
wor
king
a
r
e
a
,
the
ti
me
that
ha
s
e
laps
e
d
f
r
om
the
c
a
ptur
e
of
the
e
nvir
onment
to
the
de
t
e
c
ti
on
of
the
up
pe
r
li
mb
is
s
tor
e
d
in
the
va
r
iable
C
ountT
im
e
a
nd
is
a
dde
d
to
the
f
oll
owing
it
e
r
a
ti
ons
unti
l
a
s
e
c
ti
on
is
de
tec
ted,
i
n
whic
h
c
a
s
e
,
C
ountT
im
e
is
e
qua
led
to
z
e
r
o
a
nd
s
teps
2
a
n
d
3
a
r
e
e
xe
c
uted.
T
his
c
ondit
ional
is
only
e
xe
c
uted
a
f
ter
the
f
ir
s
t
de
tec
ti
on
o
f
the
a
r
m
(
S
tep
1
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
I
mitat
ion
of
the
human
uppe
r
li
mb
by
c
onv
olut
iona
l
ne
ur
al
ne
tw
or
k
s
(
P
aula
Us
e
c
he
)
199
(
a
)
(
b)
F
igur
e
6.
S
t
r
oke
tes
t
(
a
)
s
tr
oke
e
xa
mpl
e
a
nd
(
b)
a
r
m
tr
a
c
king
f
or
s
tr
oke
3.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
T
he
a
lgor
it
h
m
wa
s
tes
ted
wi
th
the
r
e
a
l
iza
ti
on
of
12
polygons
a
s
s
hown
in
F
igur
e
7
,
whe
r
e
the
a
c
c
ur
a
c
y
of
the
s
tr
oke
wa
s
e
va
luate
d
ba
s
e
d
on
the
c
los
ur
e
of
the
dr
a
wing
a
nd
the
number
o
f
li
ne
s
that
r
e
maine
d
outs
ide
the
polygon,
a
s
s
hown
in
T
a
ble
2
,
whe
r
e
"
C
los
ur
e
"
r
e
pr
e
s
e
nts
the
pe
r
c
e
ntage
o
f
c
los
ur
e
o
f
the
po
lygo
n,
be
ing
0%
c
ompl
e
tely
ope
n
(
a
s
tr
a
ight
li
ne
)
a
nd
100
%
c
o
mpl
e
tely
c
los
e
d.
"
Nois
e
"
r
e
pr
e
s
e
nts
the
pe
r
c
e
ntage
of
li
ne
s
that
r
e
maine
d
outs
ide
the
polygon
a
nd
do
not
be
lo
ng
to
the
f
igu
r
e
,
obtaine
d
with
r
e
s
pe
c
t
to
the
dim
e
ns
io
ns
of
the
polygon
(
c
or
r
e
s
ponding
to
100%
)
a
nd
wi
thout
c
ons
ider
ing
the
s
tar
t
a
nd
e
nd
li
ne
s
o
f
the
s
tr
oke
.
"
Ac
c
ur
a
te
c
los
ur
e
"
r
e
pr
e
s
e
nts
the
dis
tanc
e
be
tw
e
e
n
the
e
nd
point
a
nd
the
ini
ti
a
l
point
of
the
f
igu
r
e
,
100
%
be
ing
c
o
mpl
e
tely
togethe
r
,
a
nd
0
%
be
ing
the
dr
a
wing
o
f
a
c
ompl
e
tel
y
s
tr
a
ight
li
ne
be
twe
e
n
both
point
s
.
T
he
"
Ave
r
a
ge
C
los
ur
e
"
wa
s
obtaine
d
a
s
the
a
ve
r
a
ge
be
twe
e
n
the
r
e
s
ult
s
of
"
C
los
ur
e
"
a
nd
"
Ac
c
ur
a
te
C
los
ur
e
"
,
a
nd
the
"
S
tr
oke
Ac
c
ur
a
c
y"
wa
s
c
a
lcula
ted
by
s
ubtr
a
c
ti
ng
the
a
ve
r
a
ge
nois
e
pe
r
c
e
ntage
f
r
om
the
"
Ave
r
a
ge
C
los
ur
e
"
.
F
igur
e
7
.
P
olygons
made
with
the
pr
ogr
a
m
T
a
bl
e
2
.
P
r
e
c
is
ion
of
the
p
r
ogr
a
m
f
or
the
r
e
a
li
z
a
ti
o
n
of
mu
lt
ipol
ygona
l
t
r
a
c
e
s
P
ol
ygon
C
lo
s
in
g
(%)
N
oi
s
e
(%)
A
c
c
ur
a
te
c
lo
s
in
g
(%)
1
100
0
95
2
96
10
90
3
100
15
90
4
100
10
100
5
96
0
98
6
100
15
90
7
100
10
50
8
100
5
98
9
95
5
95
10
100
0
80
11
100
15
98
12
100
30
80
A
ve
r
a
ge
98
.
92
9
.
58
88
.
67
A
ve
r
a
ge
c
lo
s
in
g
93
.
79
S
tr
oke
pr
e
c
is
io
n
84
.
21
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
1
,
F
e
br
ua
r
y
2025
:
19
3
-
203
200
Ac
c
or
ding
to
the
r
e
s
ult
s
in
T
a
ble
2,
in
94%
o
f
th
e
c
a
s
e
s
it
will
be
pos
s
ibl
e
to
make
a
99%
f
igu
r
e
c
los
ur
e
,
with
a
c
los
ing
a
c
c
ur
a
c
y
o
f
89%
.
99
%
of
th
e
c
a
s
e
s
it
will
be
pos
s
ibl
e
to
c
los
e
the
f
igu
r
e
,
with
a
c
los
ing
a
c
c
ur
a
c
y
of
89%
.
Additi
ona
ll
y
,
in
84%
o
f
the
c
a
s
e
s
it
will
be
pos
s
ibl
e
to
d
r
a
w
a
polygon
without
nois
e
.
nois
e
,
a
nd
in
c
a
s
e
of
ha
ving
nois
e
,
it
will
be
les
s
than
10%
.
On
the
other
ha
nd
,
the
pe
r
c
e
ntage
of
a
c
c
ur
a
c
y
o
f
i
mi
tation
of
uppe
r
li
mb
moveme
nts
wa
s
c
a
lcula
ted,
ba
s
e
d
o
n
the
pe
r
c
e
ntage
the
ti
me
dur
ing
whic
h
the
pos
it
ion
of
the
vir
tual
a
r
m
r
e
s
e
mbl
e
s
that
o
f
the
upp
e
r
l
im
b.
the
vir
tual
a
r
m
pos
it
ion
r
e
s
e
mbl
e
s
the
r
e
a
l
one
by
mor
e
th
a
n
90%
.
F
or
thi
s
pur
pos
e
,
f
ive
polygons
we
r
e
5
polygons
w
e
r
e
pe
r
f
or
med
,
two
with
the
r
ight
a
r
m
,
two
with
the
lef
t
a
r
m
a
nd
one
with
both
a
r
ms
.
the
r
ight
a
r
m,
two
wi
th
the
lef
t
a
r
m
a
nd
one
wi
th
both
a
r
ms
,
obtaining
th
e
r
e
s
ult
s
s
hown
in
T
a
ble
3,
whe
r
e
a
ll
ti
me
va
lues
a
r
e
in
s
e
c
onds
.
ti
me
va
lues
a
r
e
in
s
e
c
onds
.
T
he
"
I
mi
tation
pe
r
c
e
ntage
"
wa
s
obtaine
d
by
divi
ding
the
"
S
uc
c
e
s
s
f
ul
im
it
a
ti
on
ti
me"
by
the
"
T
otal
im
it
a
ti
on
ti
me"
a
nd
mul
t
ipl
y
ing
it
by
100%
,
a
nd
the
"
Ave
r
a
ge
"
is
the
a
ve
r
a
ge
pe
r
c
e
ntage
f
or
the
5
c
a
s
e
s
pr
e
s
e
nted.
T
he
"
T
otal
I
mi
tation
T
i
me"
wa
s
s
tar
ted
c
ounti
ng
f
r
om
the
f
i
r
s
t
de
tec
ti
on
o
f
the
4
a
r
m
s
e
c
ti
ons
,
unti
l
the
be
ginni
ng
o
f
the
c
ount
f
or
th
e
e
nd
of
the
pr
ogr
a
m
.
T
a
ble
3.
P
r
og
r
a
m
a
c
c
ur
a
c
y
f
o
r
the
im
it
a
ti
on
o
f
hu
man
uppe
r
li
mb
moveme
nts
T
e
s
t
a
r
m
T
ot
a
l
im
it
a
ti
on
tim
e
S
uc
c
e
s
s
f
ul
im
it
a
ti
on t
im
e
P
e
r
c
e
nt
a
ge
of
im
it
a
ti
on
(%)
A
ve
r
a
ge
R
ig
ht
31
27
87
.
10
90
.
39
R
ig
ht
2
20
20
100
.
00
L
e
f
t
30
28
93
.
33
L
e
f
t
2
24
21
87
.
50
B
ot
h
25
21
84
.
00
I
n
a
ll
tes
ts
,
the
vir
tual
a
r
m
f
oll
owe
d
the
r
e
a
l
one
w
it
h
ve
r
y
s
mall
dif
f
e
r
e
nc
e
s
in
pos
it
ion,
de
ter
mi
ning
a
s
a
not
a
c
c
ur
a
te
im
it
a
ti
on
the
pos
it
ions
in
whic
h
2
or
mor
e
s
e
c
ti
ons
of
the
a
r
m
we
r
e
not
de
tec
ted
a
nd
e
s
ti
mate
d
f
or
mor
e
than
ha
l
f
a
s
e
c
ond,
whic
h
we
r
e
the
only
s
it
ua
ti
ons
in
whic
h
a
dif
f
e
r
e
nc
e
in
the
im
it
a
ti
on
wa
s
a
ppr
e
c
iate
d.
As
c
a
n
be
s
e
e
n
in
T
a
ble
3
,
the
pr
o
gr
a
m
a
c
hieve
d
90.
39
%
a
c
c
ur
a
c
y
in
the
im
it
a
ti
o
n
of
the
moveme
nts
of
the
human
uppe
r
li
mb
dur
ing
the
r
e
a
li
z
a
ti
on
of
a
mul
ti
polygonal
f
igu
r
e
,
than
ks
to
the
e
s
ti
mation
of
unde
tec
ted
point
s
,
whic
h
made
up
f
or
the
low
p
e
r
c
e
ntage
of
a
c
c
ur
a
c
y
of
the
F
a
s
t
R
-
C
NN
.
Ne
xt
,
a
c
o
mpa
r
is
on
be
twe
e
n
t
hr
e
e
di
f
f
e
r
e
n
t
R
-
C
NN
a
r
c
h
it
e
c
tur
e
s
tr
a
ined
f
o
r
the
r
e
c
og
nit
ion
of
a
r
m
pa
r
ts
is
p
r
e
s
e
n
ted
,
wh
os
e
r
e
s
ult
s
ob
taine
d
b
y
c
o
nf
us
ion
mat
r
ice
s
a
nd
r
e
c
a
ll
Vs
p
r
e
c
is
i
on
plo
ts
we
r
e
us
e
d
to
s
e
lec
t
the
f
inal
ne
tw
or
k
a
r
c
hit
e
c
tu
r
e
.
All
ne
tw
or
ks
w
e
r
e
tr
a
ined
f
o
r
250
e
poc
hs
w
it
h
a
M
in
iB
a
tch
S
ize
o
f
4
9,
w
hich
wa
s
s
e
lec
ted
a
s
hi
ghe
r
va
lu
e
s
inc
r
e
a
s
e
d
t
r
a
in
ing
ti
me
pe
r
e
p
oc
h
a
nd
r
e
d
uc
e
d
a
c
c
u
r
a
c
y
.
F
igu
r
e
s
8
(
a
)
to
8
(
c
)
s
h
ows
the
c
onf
us
ion
matr
ice
s
[
23]
–
[
25
]
(
whe
r
e
1
is
s
houlder
,
2
is
e
lbow
,
3
is
w
r
is
t
,
a
nd
4
is
f
inger
s
)
f
or
the
ne
twor
ks
tr
a
ined
with
the
a
r
c
hit
e
c
tu
r
e
s
in
T
a
ble
4
.
W
he
r
e
the
f
ir
s
t
ne
two
r
k
is
a
n
R
-
C
NN
a
s
s
hown
in
F
ig
ur
e
8
(
a
)
Ar
m_R
C
NN
a
nd
the
s
e
c
ond
one
s
a
r
e
F
a
s
t
R
-
C
NN
a
s
s
hown
in
F
ig
ur
e
8
(
b
)
,
whic
h
p
r
e
s
e
nted
a
lowe
r
d
e
tec
ti
on
ti
me
than
the
R
-
C
NN
with
be
tt
e
r
de
tec
ti
on
r
e
s
ult
s
(
T
hr
id
ne
twor
k
F
ig
u
r
e
8
(
c
)
)
,
r
e
a
s
on
why
it
wa
s
de
c
ided
to
us
e
a
F
a
s
t
R
-
C
NN
f
or
the
a
ppli
c
a
ti
on.
(
a
)
(
b)
(
c
)
F
ig
ur
e
8.
C
onf
us
ion
m
a
tr
i
x
(
a
)
r
e
d
a
r
m
_R
C
NN
,
(
b
)
r
e
d
a
r
m
_F
a
s
tR
C
NN
,
a
nd
(
c
)
r
e
d
a
r
m
_F
a
s
tR
C
NN
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
I
mitat
ion
of
the
human
uppe
r
li
mb
by
c
onv
olut
iona
l
ne
ur
al
ne
tw
or
k
s
(
P
aula
Us
e
c
he
)
201
T
a
ble
4
pr
e
s
e
nt
the
a
r
c
hit
e
c
tur
e
s
o
f
the
t
r
a
ined
ne
t
wor
ks
[
26]
,
whe
r
e
we
tr
ied
to
va
r
y
the
s
ize
of
the
f
il
ter
s
,
the
or
de
r
o
f
the
laye
r
s
a
nd
the
de
pth
o
f
the
n
e
twor
ks
.
F
or
a
ll
c
a
s
e
s
,
the
dim
e
ns
ions
of
the
de
tec
ti
on
f
r
a
me
we
r
e
40x40
pixels
,
s
ince
thi
s
is
the
mi
nim
um
s
ize
of
the
labe
ls
ge
ne
r
a
ted
on
the
t
r
a
ini
ng
im
a
ge
s
.
W
he
r
e
C
c
or
r
e
s
ponds
to
a
c
onvolut
ion
laye
r
,
R
to
t
he
r
e
c
ti
f
i
e
d
li
ne
a
r
unit
s
,
B
to
a
ba
tch
nor
maliza
ti
on
,
M
AX
P
to
one
of
M
a
xpooli
ng,
F
C
to
a
f
ull
y
c
onne
c
ted
,
DR
OP
to
a
d
r
opout
,
a
nd
S
OFT
to
a
s
of
tm
a
x
.
T
a
ble
4.
C
NN
a
r
c
hit
e
c
tur
e
s
N
E
T
W
O
R
K
1
N
E
T
W
O
R
K
2
N
E
T
W
O
R
K
3
L
A
Y
E
R
S
F
il
te
r
S
iz
e
S
tr
id
e
N
umbe
r
F
il
te
r
s
L
A
Y
E
R
S
F
il
te
r
S
iz
e
S
tr
id
e
N
umbe
r
F
il
te
r
s
L
A
Y
E
R
S
F
il
te
r
S
iz
e
S
tr
id
e
N
umbe
r
F
il
te
r
s
C
+
B
+
R
6x6
1
16
C
+
B
+
R
6x6
1
16
C
+
B
+
R
3x3
1
16
C
+
B
+
R
5x5
1
64
C
+
B
+
R
5x5
1
64
M
A
X
P
2x2
2
--
C
+
B
+
R
4x4
1
128
M
A
X
P
3x3
2
--
C
+
B
+
R
3x3
1
256
M
A
X
P
3x3
2
--
C
+
B
+
R
4x4
1
128
M
A
X
P
3x3
2
--
C
+
B
+
R
3x3
1
256
C
+
B
+
R
3x3
1
256
C
+
B
+
R
3x3
1
512
C
+
B
+
R
3x3
1
512
F
C
1+
R
E
L
U
+
D
R
O
P
F
C
2+
R
E
L
U
+
D
R
O
P
F
C
3+
S
O
F
T
Obs
e
r
ving
that
none
of
the
ne
twor
ks
e
xc
e
e
de
d
61%
,
the
ne
twor
k
with
the
highes
t
pe
r
c
e
ntage
wa
s
take
n
a
nd
tr
a
ined
f
or
a
n
a
ddit
ional
200
e
poc
hs
,
h
owe
ve
r
,
it
s
a
c
c
ur
a
c
y
dr
oppe
d
to
59
.
5%
,
with
no
n
oti
c
e
a
ble
im
pr
ove
ment
in
f
r
a
me
de
tec
ti
on
pe
r
c
e
ntage
.
W
e
tr
ied
to
tr
a
in
that
s
a
me
ne
twor
k
with
a
lar
ge
r
M
ini
B
a
tchSi
z
e
,
a
nd
with
a
s
maller
one
,
obtaining
pe
r
c
e
ntage
s
of
4
3
%
a
nd
43.
1
%
,
r
e
s
pe
c
ti
ve
ly
(
Ar
m_F
a
s
tR
C
NN
_3
be
ing
the
ne
twor
k
with
43%
a
c
c
ur
a
c
y)
.
T
o
ve
r
if
y
that
the
p
e
r
c
e
ntage
r
e
duc
ti
on
wa
s
not
due
to
ove
r
-
tr
a
ini
ng,
the
ne
w
tr
a
ined
ne
twor
ks
we
r
e
tes
ted
with
the
tr
a
ini
ng
da
taba
s
e
,
obtaining
a
bout
44%
a
c
c
ur
a
c
y,
whic
h
a
ll
ow
s
us
to
de
duc
e
that
the
ne
twor
k
did
not
memor
ize
.
Due
to
thes
e
r
e
s
ult
s
,
it
wa
s
de
c
ided
to
e
va
lu
a
te
the
a
c
ti
va
ti
ons
[
27]
of
the
t
r
a
ined
ne
twor
ks
to
obs
e
r
ve
the
qua
li
ty
of
the
inf
o
r
mation
e
xtr
a
c
ted
by
e
a
c
h
o
f
t
he
m
a
nd
f
r
om
ther
e
de
ter
mi
ne
the
s
tr
uc
tur
e
r
e
quir
e
d
f
o
r
the
f
inal
ne
twor
k
.
T
he
a
c
ti
va
ti
ons
s
howe
d
that
thes
e
a
r
c
hit
e
c
tur
e
s
did
not
mana
ge
to
f
oc
us
on
lea
r
ning
the
a
r
m,
but
c
a
pt
ur
e
d
other
types
of
f
e
a
tur
e
s
s
uc
h
a
s
the
wood
,
the
c
ha
ir
a
nd
the
body,
while
the
A
r
m_F
a
s
tR
C
NN
,
f
oc
us
e
d
on
the
a
r
m
a
nd
pa
r
t
o
f
the
ba
c
kgr
ound.
F
or
a
ny
o
f
t
he
tr
a
ined
ne
twor
ks
,
the
mos
t
dif
f
icult
s
e
c
ti
on
of
th
e
a
r
m
to
r
e
c
ognize
is
the
e
lbow,
due
to
the
mul
ti
ple
pos
it
ions
it
c
a
n
a
dopt,
f
r
om
f
ull
y
s
tr
e
tche
d
to
tucke
d
tow
a
r
ds
the
c
he
s
t,
in
a
ddit
ion
to
the
upwa
r
d
o
r
downw
a
r
d
ti
lt
(
with
r
e
s
pe
c
t
to
the
g
r
ound)
,
whic
h
c
a
n
oc
c
ur
d
ur
ing
the
moveme
nt
of
the
a
r
m.
S
i
mi
lar
ly
,
the
dis
plac
e
ment
of
W
r
is
t
a
nd
it
s
r
otation
th
r
oughout
the
im
a
ge
a
ls
o
led
to
mul
ti
ple
r
e
pr
e
s
e
ntations
,
s
ome
of
them
s
im
il
a
r
to
o
ther
pa
r
ts
of
the
a
r
m
s
uc
h
a
s
E
lbow
.
Due
to
thes
e
de
tec
ti
on
dif
f
iculti
e
s
,
the
de
c
is
ion
wa
s
made
to
make
the
p
r
oc
e
s
s
of
"
P
oint
E
s
ti
mation"
a
s
a
T
r
a
c
king
s
ys
tem
f
or
the
a
lgor
it
h
m,
in
or
de
r
t
o
ke
e
p
tr
a
c
k
of
a
s
many
pa
r
ts
of
the
a
r
m
a
s
pos
s
i
ble
f
or
a
r
obot
tele
ope
r
a
ti
on
c
los
e
r
to
the
r
e
a
l
one
,
ove
r
the
lac
k
of
in
f
or
mation
ge
ne
r
a
ted
by
the
ne
twor
k.
A
s
s
umi
ng
that
the
point
e
s
ti
mation
c
ondit
ionals
make
up
f
or
the
low
de
tec
ti
on
of
e
lbo
w
,
the
c
onf
us
ion
matr
ix
of
Ar
m_F
a
s
tR
C
NN
wa
s
c
a
lcula
ted
f
or
the
r
e
maining
3
s
e
c
ti
ons
of
the
a
r
m
,
whe
r
e
the
a
c
c
ur
a
c
y
r
os
e
to
6
6.
3%
.
4.
CONC
L
USI
ON
T
he
t
r
a
c
king
s
ys
tem
(
point
e
s
ti
mation)
im
pleme
nt
e
d
in
the
pr
ogr
a
m
,
a
ll
owe
d
r
e
a
c
hing
a
pe
r
c
e
ntage
of
r
e
c
ognit
ion
a
nd
r
e
pli
c
a
ti
on
of
moveme
nts
of
t
he
human
uppe
r
li
mb
of
up
to
90
%
a
c
c
ur
a
c
y,
lea
ding
the
pr
ogr
a
m
to
a
c
hieve
mo
r
e
than
98%
of
c
los
ur
e
in
th
e
mul
ti
polygonal
f
igur
e
s
,
with
a
pe
r
c
e
ntage
of
nois
e
of
les
s
than
10%
,
de
s
pit
e
us
ing
a
F
a
s
t
R
-
C
NN
with
on
ly
60%
a
c
c
ur
a
c
y
.
C
onc
ludi
ng
that
the
a
c
ti
on
o
f
ne
twor
k
r
e
c
ognit
ion
c
a
n
be
s
tr
e
ngthene
d
withi
n
the
a
ppli
c
a
ti
on.
T
he
wide
va
r
iation
of
pos
it
ions
that
the
e
lbow
c
a
n
a
dopt
thr
oughout
the
a
ppli
c
a
ti
on
,
a
nd
the
high
de
gr
e
e
of
s
im
il
a
r
it
y
be
twe
e
n
the
dif
f
e
r
e
nt
s
e
c
ti
ons
of
the
a
r
m
a
nd
it
s
mul
ti
ple
pos
it
ions
(
both
f
or
the
r
ight
a
nd
lef
t
a
r
m
)
,
ge
ne
r
a
ted
dif
f
iculti
e
s
in
the
tr
a
ini
ng
of
the
ne
two
r
k,
due
to
the
de
manding
ne
e
d
not
to
c
onf
us
e
the
s
e
c
ti
ons
w
it
h
e
a
c
h
other
,
nor
wi
th
the
ba
c
kgr
ound,
a
nd
to
a
c
hieve
a
n
a
c
c
ur
a
c
y
in
the
de
tec
ti
o
n
f
r
a
mes
that
would
a
ll
ow
the
pos
it
ion
of
the
vir
tual
a
r
m
to
be
s
im
il
a
r
to
the
r
e
a
l
one
.
Due
to
thes
e
c
ompl
ica
ti
ons
,
it
wa
s
pr
e
f
e
r
r
e
d
to
s
e
lec
t
a
ne
two
r
k
wi
th
a
n
a
c
c
ur
a
c
y
of
60%
ins
tea
d
o
f
a
nother
with
72%
,
be
c
a
us
e
of
the
a
c
c
ur
a
c
y
o
f
the
A
r
m_F
a
s
tR
C
NN
in
th
e
ge
ne
r
a
ti
on
of
the
de
tec
ti
on
f
r
a
mes
f
or
e
a
c
h
c
a
tegor
y,
whe
r
e
a
s
maller
number
of
f
r
a
mes
we
r
e
r
e
c
ognize
d,
but
c
los
e
r
to
wha
t
wa
s
e
xpe
c
ted.
T
he
us
e
of
lar
ge
r
input
im
a
ge
s
than
thos
e
us
e
d
in
the
tr
a
ini
ng
da
taba
s
e
f
or
a
F
a
s
t
R
-
C
NN
,
may
incr
e
a
s
e
it
s
a
c
c
ur
a
c
y
i
n
the
c
las
s
if
ica
ti
on
of
the
ne
twor
k
c
a
tegor
ies
,
howe
ve
r
,
i
t
doe
s
not
mea
n
that
it
inc
r
e
a
s
e
s
in
the
s
a
me
wa
y
the
a
c
c
ur
a
c
y
of
the
de
tec
ti
on
f
r
a
mes
,
nor
the
ove
r
a
ll
p
e
r
f
or
manc
e
o
f
the
p
r
ogr
a
m
pe
r
f
or
med
f
or
the
r
e
pli
c
a
ti
on
of
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
1
,
F
e
br
ua
r
y
2025
:
19
3
-
203
202
tr
a
c
e
s
a
nd
im
it
a
ti
on
o
f
human
uppe
r
li
mb,
s
o
it
is
e
s
s
e
nti
a
l,
f
or
thi
s
type
of
a
ppli
c
a
ti
ons
,
to
e
va
luate
both
the
a
c
c
ur
a
c
y
of
the
ne
twor
k
a
nd
the
a
c
c
ur
a
c
y
in
the
de
tec
ti
on
of
e
a
c
h
o
f
it
s
c
a
tegor
ies
.
AC
KNOWL
E
DGE
M
E
NT
S
T
he
a
uthor
s
a
r
e
gr
a
tef
ul
to
Unive
r
s
idad
M
il
it
a
r
Nu
e
va
Gr
a
na
da
a
nd
Unive
r
s
idad
de
los
L
lanos
.
RE
F
E
RE
NC
E
S
[
1]
Q
.
W
u,
“
R
e
s
e
a
r
c
h
on
h
uma
n
bod
y
d
e
te
c
ti
o
n
a
nd
tr
a
c
k
in
g
a
lg
or
i
t
hm
ba
s
e
d
o
n
ki
n
e
c
t,
”
in
2
021
I
nt
e
r
na
ti
ona
l
C
onf
e
r
e
n
c
e
on
E
le
c
t
r
oni
c
I
nf
or
m
at
io
n T
e
c
hn
ol
og
y
a
nd S
m
a
r
t
A
g
r
ic
ul
tu
r
e
, I
C
E
I
T
SA
2021
,
2021,
pp.
7
–
10
, d
oi
:
1
0.11
09/
I
C
E
I
T
S
A
5422
6.20
21.0
0011
.
[
2]
A
. A
. K
upi
n, S
. B
a
ne
r
je
e
, N
. B
a
n
e
r
je
e
, S
. H
. R
oy, J
. C
. K
li
ne
,
a
nd B
. S
hi
w
a
ni
, “
S
ys
te
m
a
r
c
hi
te
c
tu
r
e
f
or
V
R
yoga
t
he
r
a
py pla
tf
or
m
w
it
h
6
-
D
oF
w
hol
e
-
body
a
va
ta
r
t
r
a
c
k
in
g,”
in
2024
I
E
E
E
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on
A
r
ti
fi
c
ia
l
I
nt
e
ll
ig
e
nc
e
and
e
X
te
nde
d
and
V
ir
tu
al
R
e
al
it
y
, A
I
x
V
R
2024
, 2024, pp. 360
–
366
, doi
:
10.1109/
A
I
xV
R
59861.2024.00062.
[
3]
N
.
A
.
S
.
N
.
N
a
nda
s
e
na
,
W
.
A
.
A
.
V
im
ukt
hi
,
H
.
M
.
K
.
K
.
M
.
B
.
H
e
r
a
th
,
R
.
W
ij
e
s
in
gh
e
,
a
nd
S
.
L
.
P
.
Y
a
s
a
ke
th
u,
“
R
e
a
l
-
ti
me
uppe
r
body
mot
io
n
tr
a
c
ki
ng
us
in
g
c
omput
e
r
vi
s
io
n
f
or
im
pr
ove
d
huma
n
-
r
obot
in
te
r
a
c
ti
on
a
nd
te
le
op
e
r
a
ti
on,”
M
or
at
uw
a
E
ngi
ne
e
r
in
g
R
e
s
e
ar
c
h C
onf
e
r
e
n
c
e
, M
E
R
C
on
, pp. 201
–
206, 2023, doi:
10.1109/M
E
R
C
on60487.2023.10355479.
[
4]
S
. M
. L
.
G
io
i,
G
. L
oi
a
nno,
a
nd F
. C
or
de
ll
a
, “
R
obus
t
uppe
r
l
i
mb
ki
ne
ma
ti
c
r
e
c
ons
tr
uc
ti
on us
in
g a
R
G
B
-
D
c
a
me
r
a
,”
I
E
E
E
R
obo
ti
c
s
and A
ut
om
at
io
n L
e
tt
e
r
s
, vol
. 9, no. 4, pp. 3831
–
3837, 2024, doi:
10.1109/L
R
A
.2024.3373236.
[
5]
H
.
X
u,
J
.
F
a
n,
H
.
M
a
,
a
nd
Q
.
W
a
ng,
“
S
e
mi
pa
r
a
me
tr
ic
m
us
c
ul
os
ke
le
ta
l
mode
l
f
or
r
e
in
f
or
c
e
me
nt
le
a
r
ni
ng
-
ba
s
e
d
tr
a
je
c
to
r
y
tr
a
c
ki
n
g,”
I
E
E
E
T
r
ans
ac
ti
ons
on I
ns
tr
um
e
nt
at
io
n and M
e
as
u
r
e
m
e
nt
, vol
.
73, pp. 1
–
16, 2024, doi:
10.1109/T
I
M
.2024.3370760.
[
6]
Z
.
F
e
ng
a
nd
P
.
W
a
ng,
“
A
mode
l
a
da
pt
iv
e
upda
ti
ng
k
e
r
ne
l
c
or
r
e
la
ti
on
f
il
te
r
tr
a
c
ke
r
w
it
h
de
e
p
C
N
N
f
e
a
tu
r
e
s
,”
E
ngi
ne
e
r
in
g
A
ppl
ic
at
io
ns
of
A
r
ti
fi
c
ia
l
I
nt
e
ll
ig
e
nc
e
, vol
. 123, 2023, doi:
10.1
016/
j.
e
nga
ppa
i.
2023.106250.
[
7]
L
.
A
.
H
us
s
a
in
,
S
.
S
in
gh,
R
.
M
iz
ouni
,
H
.
O
tr
ok,
a
nd
E
.
D
a
mi
a
ni
,
“
A
pr
e
di
c
ti
ve
ta
r
ge
t
tr
a
c
ki
ng
f
r
a
me
w
or
k
f
or
I
oT
us
in
g
C
N
N
–
L
S
T
M
,”
I
nt
e
r
ne
t
of
T
hi
ngs
, vol
. 22, 2023, doi:
10.1016/j
.i
ot
.2023.100744.
[
8]
Q
.
Z
hu,
X
.
H
ua
ng,
a
nd
Q
.
G
ua
n,
“
T
a
bC
tNe
t:
T
a
r
ge
t
-
a
w
a
r
e
bi
la
t
e
r
a
l
C
N
N
-
tr
a
ns
f
or
me
r
ne
twor
k
f
or
s
in
gl
e
obj
e
c
t
tr
a
c
ki
ng
in
s
a
te
ll
it
e
vi
de
os
,”
I
nt
e
r
na
ti
onal
J
ou
r
nal
of
A
ppl
ie
d E
ar
th
O
bs
e
r
v
at
io
n an
d G
e
oi
nf
or
m
at
io
n
, vol
. 128, 2024, doi:
10.1016/j
.j
a
g.2024.1037
23.
[
9]
M
.
B
a
b
a
e
e
,
Z
.
L
i,
a
nd
G
.
R
ig
ol
l,
“
A
du
a
l
C
N
N
–
R
N
N
f
or
mul
ti
pl
e
pe
opl
e
tr
a
c
ki
ng,”
N
e
u
r
oc
om
put
in
g
,
vol
.
368,
pp.
69
–
83,
20
19,
doi
:
10.1016/j
.ne
uc
om.2019.08.008.
[
10]
G
.
L
i
e
t
al
.
,
“
O
ne
-
s
hot
mul
ti
-
obj
e
c
t
tr
a
c
ki
ng
us
in
g
C
N
N
-
ba
s
e
d
ne
twor
ks
w
it
h
s
pa
ti
a
l
-
c
ha
nne
l
a
tt
e
nt
io
n
me
c
h
a
ni
s
m,”
O
pt
ic
s
and
L
as
e
r
T
e
c
hnol
ogy
, vol
. 153, 2022, doi:
10.1016/j
.opt
la
s
te
c
.202
2.108267.
[
11]
Y
.
W
a
ng,
X
.
W
e
i,
X
.
T
a
ng,
H
.
S
h
e
n,
a
nd
L
.
D
in
g,
“
C
N
N
tr
a
c
ki
ng
ba
s
e
d
on
da
ta
a
ugm
e
nt
a
ti
on,”
K
no
w
le
dge
-
B
as
e
d
Sy
s
te
m
s
,
vol
.
194, 2020, doi:
10.1016/j
.knos
ys
.2020.105594.
[
12]
M
. F
. A
s
la
n, A
. D
ur
du, K. S
a
ba
nc
i,
a
nd M
. A
. M
ut
lu
e
r
, “
C
N
N
a
nd H
O
G
ba
s
e
d c
ompa
r
is
on s
tu
dy f
or
c
ompl
e
te
oc
c
lu
s
io
n ha
ndl
in
g
in
huma
n
tr
a
c
ki
ng,”
M
e
as
ur
e
m
e
nt
:
J
our
nal
of
th
e
I
nt
e
r
nat
io
nal
M
e
as
ur
e
m
e
nt
C
onf
e
de
r
at
io
n
,
vol
.
158,
2020,
doi
:
10.1016/j
.me
a
s
ur
e
me
nt
.2020.107704.
[
13]
Y
.
W
a
ng,
X
.
W
e
i,
H
.
S
he
n,
L
.
D
in
g,
a
nd
J
.
W
a
n,
“
R
obu
s
t
f
us
io
n
f
or
R
G
B
-
D
tr
a
c
ki
ng
us
in
g
C
N
N
f
e
a
tu
r
e
s
,”
A
ppl
ie
d
Sof
t
C
om
put
in
g
J
our
nal
, vol
. 92, 2020, doi:
10.1016/j
.a
s
oc
.2020.106302.
[
14]
Y
.
W
a
ng,
X
.
W
e
i,
L
.
L
uo,
W
.
W
e
n,
a
nd
Y
.
W
a
ng,
“
R
obus
t
R
G
B
-
D
tr
a
c
ki
ng
vi
a
c
ompa
c
t
C
N
N
f
e
a
tu
r
e
s
,”
E
ngi
ne
e
r
in
g
A
ppl
ic
at
io
ns
of
A
r
ti
fi
c
ia
l
I
nt
e
ll
ig
e
nc
e
, vol
. 96, 2020, doi:
10.1016/j
.e
nga
ppa
i.
2020.103974.
[
15]
X
.
W
a
ng,
Y
. B
a
i,
a
nd X
.
L
iu
,
“
P
r
e
di
c
ti
on
of
r
a
il
r
o
a
d
tr
a
c
k
ge
o
me
tr
y
c
ha
nge
us
in
g a
hybr
id
C
N
N
-
L
S
T
M
s
pa
ti
a
l
-
te
mpor
a
l
mode
l,
”
A
dv
anc
e
d E
ngi
ne
e
r
in
g I
nf
or
m
at
ic
s
, vol
. 58, 2023, doi:
10.1016
/j
.a
e
i.
2023.102235.
[
16]
N
.
G
una
w
a
r
de
na
,
J
.
A
.
G
in
ig
e
,
B
.
J
a
va
di
,
a
nd
G
.
L
ui
,
“
P
e
r
f
o
r
ma
nc
e
a
na
ly
s
i
s
of
C
N
N
mode
ls
f
or
mob
il
e
de
vi
c
e
e
ye
tr
a
c
ki
ng
w
it
h
e
dge
c
omp
ut
in
g,”
P
r
oc
e
di
a C
om
put
e
r
Sc
ie
n
c
e
, vol
. 207, pp. 22
91
–
2300, 2022, doi:
10.1016/j
.pr
oc
s
.2022.09.288.
[
17]
R
.
G
ir
s
hi
c
k,
“
F
a
s
t
R
-
C
N
N
,”
in
P
r
oc
e
e
di
ng
s
of
th
e
I
E
E
E
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on
C
om
put
e
r
V
is
io
n
,
2015,
I
C
C
V
201
5,
pp.
1440
–
1448
, doi
:
10.1109/I
C
C
V
.2015.169.
[
18]
F
.
C
ui
,
M
.
N
in
g,
J
.
S
he
n,
a
nd
X
.
S
hu,
“
A
ut
oma
ti
c
r
e
c
ogni
ti
on
a
n
d
tr
a
c
ki
ng
of
hi
ghw
a
y
la
ye
r
-
in
te
r
f
a
c
e
us
in
g
F
a
s
te
r
R
-
C
N
N
,
”
J
ou
r
nal
of
A
ppl
ie
d G
e
ophy
s
ic
s
, vol
. 196, J
a
n. 2022, doi:
10.1016/j
.j
a
ppge
o.2021.104
477.
[
19]
J
. O
. P
.
A
r
e
na
s
, M
.
R
. J
im
é
ne
z
,
a
nd P
. C
. U
.
M
ur
il
lo
, “
F
a
s
te
r
R
-
C
N
N
f
or
obj
e
c
t
lo
c
a
ti
on i
n a
vi
r
tu
a
l
e
nvi
r
onme
nt
f
or
s
or
t
in
g t
a
s
k,”
I
nt
e
r
nat
io
nal
J
our
nal
of
O
nl
in
e
E
ngi
ne
e
r
in
g
, vol
. 14, no. 7, pp.
4
–
14, 2018, doi:
10.3991/i
jo
e
.v14i07.8465.
[
20]
L
.
H
.
L
i
a
nd
R
.
T
a
non
e
,
“
E
ns
e
mbl
e
l
e
a
r
ni
ng
ba
s
e
d
on
C
N
N
a
nd
tr
a
ns
f
or
me
r
mode
ls
f
or
le
a
f
di
s
e
a
s
e
s
c
la
s
s
if
ic
a
ti
on
,”
in
P
r
o
c
e
e
di
ngs
of
t
he
2024 18th
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on Ubiquit
ous
I
nf
or
m
at
io
n M
anage
m
e
nt
and C
om
m
uni
c
at
io
n, I
M
C
O
M
2024
, 2024,
pp.
1
–
6
, doi
:
10.1109/I
M
C
O
M
60618.2024.10418393.
[
21]
M
.
P
.
C
.
U
s
e
c
he
,
J
.
P
.
A
r
e
na
s
,
a
nd
R
.
J
.
M
or
e
no,
“
I
mpl
e
me
nt
a
ti
on
of
a
da
ta
a
ugme
nt
a
ti
on
a
lg
or
it
hm
va
li
da
te
d
by
me
a
ns
of
th
e
a
c
c
ur
a
c
y
of
a
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k,”
J
ou
r
nal
of
E
ngi
n
e
e
r
in
g
and
A
ppl
ie
d
S
c
ie
n
c
e
s
,
vol
.
12,
no.
20,
pp.
5323
–
5331,
2
017,
doi
:
10.3923/j
e
a
s
c
i.
2017.5323.5331.
[
22]
P
. A
. J
a
vi
e
r
O
r
la
ndo,
J
. M
.
R
obi
ns
on,
a
nd J
.
E
. M
. B
a
que
r
o, “
C
ompa
r
is
on of
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k mode
ls
f
or
us
e
r
’
s
f
a
c
ia
l
r
e
c
ogni
ti
on,”
I
nt
e
r
nat
io
nal
J
our
nal
of
E
le
c
tr
ic
al
and
C
o
m
put
e
r
E
ngi
ne
e
r
in
g
,
vol
.
14,
no.
1,
pp.
192
–
198,
2024,
doi
:
10.11591/i
je
c
e
.v14i1.pp192
-
198.
[
23]
H
.
Y
un,
“
P
r
e
di
c
ti
on
mode
l
of
a
lg
a
l
bl
ooms
us
in
g
lo
gi
s
ti
c
r
e
gr
e
s
s
io
n
a
nd
c
onf
us
io
n
ma
tr
ix
,”
I
nt
e
r
nat
io
nal
J
our
nal
of
E
le
c
tr
ic
al
and
C
om
put
e
r
E
ngi
ne
e
r
in
g
, vol
. 11, no. 3, pp. 2407
–
2413, 2021, do
i:
10.11591/i
je
c
e
.v11i3.pp2407
-
2413.
[
24]
L
.
E
.
P
omm
e
,
R
.
B
our
qui
,
R
.
G
io
t,
a
nd
D
.
A
ube
r
,
“
R
e
la
ti
ve
c
onf
us
io
n
ma
tr
ix
:
e
f
f
ic
ie
nt
c
ompa
r
is
on
of
de
c
i
s
io
n
mode
ls
,”
in
P
r
oc
e
e
di
ngs
of
th
e
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on
I
nf
or
m
at
io
n
V
is
ual
is
at
io
n
,
2022,
vol
.
2022,
pp.
98
–
103
,
doi
:
10.1109/I
V
56949.2022.00025.
[
25]
A
.
K
r
iz
he
vs
ky,
I
.
S
ut
s
k
e
ve
r
,
a
nd
G
.
E
.
H
in
to
n,
“
I
ma
ge
N
e
t
c
la
s
s
if
ic
a
ti
on
w
it
h
de
e
p
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
s
,”
C
om
m
uni
c
at
i
ons
of
t
he
A
C
M
, vol
. 60, no. 6, pp. 8
4
–
90, 2017, doi:
10.1145/3065
386.
[
26]
T
.
D
.
T
r
in
h,
P
.
C
.
L
.
T
.
V
u,
a
nd
P
.
T
.
B
a
o,
“
C
ompa
r
in
g
m
a
s
k
R
-
C
N
N
ba
c
kbone
a
r
c
hi
te
c
tu
r
e
s
f
or
huma
n
de
te
c
ti
on
u
s
in
g
th
e
r
ma
l
im
a
gi
ng,”
I
nt
e
r
nat
io
nal
J
our
nal
of
E
le
c
tr
ic
al
and
C
om
put
e
r
E
ngi
ne
e
r
in
g
,
vol
.
14,
no.
4,
pp.
3962
–
3970,
2024,
doi
:
10.11591/i
je
c
e
.v14i4.pp3962
-
3970.
[
27]
P
.
K
.
B
a
ll
a
,
A
.
K
uma
r
,
a
nd
R
.
P
a
nde
y,
“
A
4
-
c
ha
nne
ll
e
d
ha
z
y
im
a
ge
in
put
ge
ne
r
a
ti
on
a
nd
d
e
e
p
le
a
r
ni
ng
-
ba
s
e
d
s
in
gl
e
i
ma
ge
de
ha
z
in
g,”
J
ou
r
nal
of
V
is
ual
C
om
m
uni
c
at
io
n and I
m
age
R
e
pr
e
s
e
nt
at
io
n
, vol
. 10
0, 2024, doi:
10.1016/j
.j
vc
ir
.2024.104099.
Evaluation Warning : The document was created with Spire.PDF for Python.