T
E
L
KO
M
NI
K
A
,
V
ol
.
14,
N
o.
3,
S
ept
em
ber
20
16,
pp.
11
92
~
120
2
I
S
S
N
:
1
693
-
6
930
,
ac
c
r
edi
t
ed
A
b
y
D
IK
T
I,
D
e
c
r
e
e
N
o
:
58/
D
I
K
T
I
/
K
ep/
2013
D
O
I
:
10.
12928/
T
E
LK
O
M
N
I
K
A
.
v
1
4
i
3
.
3556
11
92
R
ec
ei
v
ed
Mar
c
h 10
,
20
1
6
;
R
ev
i
s
ed
J
u
ne 2
,
20
1
6
;
A
c
c
ept
e
d
J
une
1
7
,
20
1
6
A
c
t
i
o
n R
e
c
og
ni
t
i
on o
f H
uma
n’
s
Lo
w
e
r
Li
mbs
B
a
s
e
d o
n
a
H
u
ma
n J
oi
n
t
F
e
n
g
L
i
a
n
g
*
,
Z
h
i
l
i
Z
h
a
n
g
,
X
i
a
n
g
y
a
n
g
L
i
,
Y
o
n
g
L
o
n
g
,
Z
h
a
o
T
o
n
g
X
i
'
an R
es
ear
c
h I
n
s
t
i
t
ut
e
of
H
i
g
h
-
T
ec
hnol
ogy
,
X
i
'
an,
S
ha
anx
i
71002
5,
C
hi
na
*
C
or
r
es
po
ndi
ng a
ut
hor
,
e
-
ma
i
l
:
5245
2760
6
@
qq
.
c
om
A
b
st
r
act
T
o r
ec
o
gni
z
e t
he ac
t
i
on
s
of
hum
an’
s
l
ow
er
l
i
m
bs
ac
c
ur
at
el
y
and q
ui
c
k
l
y
,
a
n
ov
el
a
c
t
ion
r
ec
og
ni
t
i
on m
et
hod
bas
ed o
n
a hum
an
j
oi
nt
w
a
s
pr
o
pos
ed.
F
i
r
s
t
l
y
,
hi
p
j
oi
nt
w
a
s
c
hos
en a
s
t
he
r
e
c
o
g
ni
t
ion
obj
e
c
t
,
i
t
s
y
c
oor
di
n
at
e
s
w
er
e as
r
ec
o
gni
t
i
on p
ar
am
et
er
,
a
nd
hum
an ac
t
i
o
n c
har
ac
t
er
i
s
t
i
c
s
w
er
e ac
hi
ev
e
d
bas
e
d on
B
ut
t
er
w
or
t
h f
i
l
t
er
a
nd w
av
el
et
t
r
an
s
f
or
m
.
S
e
c
on
dl
y
,
an
i
m
pr
ov
e
d s
el
f
-
or
gan
i
z
i
ng c
om
pet
i
t
i
v
e
neur
al
n
et
w
or
k
w
as
pr
op
os
e
d,
w
hi
c
h c
oul
d c
l
a
s
s
i
f
y
t
he ac
t
i
on
c
har
a
c
t
er
i
s
t
i
c
s
a
ut
om
at
i
c
al
l
y
a
c
c
or
di
ng t
o t
h
e
c
l
a
s
s
i
f
i
c
at
i
on num
ber
.
T
he
c
l
a
s
s
i
f
i
c
at
i
o
n r
es
ul
t
s
of
m
ot
i
on
c
apt
ur
e
dat
a
pr
ov
ed t
h
e v
al
i
d
i
t
y
of
t
h
e ne
ur
al
net
w
or
k
.
F
i
nal
l
y
,
an
a
c
t
i
o
n
r
ec
ogni
t
i
on
m
et
ho
d
b
as
e
d
o
n
h
i
d
den
M
ar
k
ov
m
ode
l
(
H
M
M
)
w
as
i
nt
r
od
uc
e
d
t
o
r
eal
i
z
e
t
he
r
ec
ogn
i
t
i
o
n
o
f
c
l
a
s
s
i
f
i
c
at
i
on
r
es
ul
t
s
o
f
h
um
an
a
c
t
i
on
c
har
a
c
t
er
i
s
t
i
c
s
w
i
t
h
t
he c
ha
nge di
r
e
c
t
i
on o
f
y
c
oor
di
na
t
es
.
T
he
pr
o
pos
ed
ac
t
i
o
n
r
ec
o
gni
t
i
on
m
et
ho
d
n
eeds
l
es
s
ac
t
i
on
i
n
f
or
m
at
i
o
n
and
ha
s
a
f
as
t
c
al
c
ul
at
i
o
n s
pee
d.
E
x
p
er
i
m
en
t
s
pr
ov
ed t
he m
et
ho
d
had
a
hi
gh r
e
c
ogn
i
t
i
on r
at
e an
d
a
good app
l
i
c
at
i
o
n
pr
os
p
ec
t
.
Ke
y
w
o
rd
s
:
hum
an ac
t
i
on c
har
a
c
t
er
i
s
t
i
c
s
,
c
h
a
r
ac
t
e
r
is
t
ic
c
l
as
s
i
f
i
c
a
t
i
on,
i
m
pr
o
v
ed
s
el
f
-
o
r
g
an
i
z
ing
c
om
pet
i
t
i
v
e ne
ur
al
net
w
or
k
;
a
c
t
i
on r
ec
o
gni
t
i
on
C
o
p
y
r
i
g
h
t
©
20
16
U
n
i
ver
si
t
a
s A
h
mad
D
ah
l
an
.
A
l
l
r
i
g
h
t
s r
eser
ved
.
1
.
I
n
tr
o
d
u
c
ti
o
n
W
i
t
h t
he
de
v
e
l
opm
ent
a
nd
popu
l
ar
i
t
y
of
op
t
i
c
a
l
m
ot
i
on
c
apt
ur
e
equ
i
pm
ent
,
h
um
an ac
t
i
on
r
ec
ogni
t
i
o
n
t
ec
hno
l
o
g
y
ha
s
r
ec
ei
v
ed
m
uc
h
at
t
ent
i
o
n
i
n
r
ec
e
nt
y
ear
s
.
Ma
n
y
s
c
hol
ar
s
ha
v
e
s
t
udi
e
d t
he t
ec
hno
l
o
g
y
ex
t
ens
i
v
e
l
y
a
nd i
nt
r
od
uc
ed
m
an
y
m
et
hods
t
o r
ec
ogn
i
z
e t
h
e hum
a
n
ac
t
i
on
pr
oc
es
s
es
i
n r
ec
en
t
20
y
ear
s
.
C
hen
et
al
.
[
1
]
ex
t
r
ac
t
ed s
pat
i
o
-
t
em
por
al
i
n
t
er
es
t
po
i
nt
s
a
nd
3D
-
S
I
F
T
des
c
r
i
pt
or
s
ar
oun
d
eac
h
i
nt
er
es
t
poi
nt
i
n
t
h
e
v
i
de
os
and
i
nt
r
oduc
e
d
a
h
um
an
behav
i
or
c
l
as
s
i
f
i
c
at
i
on
m
odel
bas
e
d
on
d
y
n
am
i
c
B
a
y
es
i
a
n
net
w
or
k
.
B
ar
na
c
h
on
et
a
l
.
[
2
]
ex
t
r
ac
t
e
d
hi
s
t
ogr
am
s
of
ac
t
i
on
pos
es
f
r
o
m
m
ot
i
on
c
apt
ur
e
d
at
a
t
o
r
ec
ogn
i
z
e
ong
oi
ng
ac
t
i
v
i
t
i
es
bas
e
d
on
d
y
n
am
i
c
t
i
m
e pl
ann
i
ng.
K
a
m
al
et
al
.
[
3
]
r
ec
e
i
v
ed a
s
e
quenc
e
of
dept
h
m
aps
t
o
ex
t
r
ac
t
hum
an
s
i
l
ho
uet
t
es
,
f
r
om
w
hi
c
h
h
y
b
r
i
d f
eat
ur
es
as
o
pt
i
c
a
l
f
l
o
w
m
ot
i
on f
eat
ur
es
a
nd
di
s
t
an
c
e par
am
et
er
s
w
er
e ex
t
r
ac
t
ed an
d us
ed t
o
w
or
k
as
s
pat
i
o
-
t
em
por
al
f
e
at
ur
es
;
t
hes
e f
eat
ur
es
w
er
e
c
l
us
t
er
ed an
d
s
y
m
bol
i
z
e
d b
y
s
e
l
f
-
or
gani
z
at
i
o
n m
aps
and H
MMs
w
e
r
e t
r
ai
ne
d t
o r
ec
ogn
i
z
e hu
m
an ac
t
i
v
i
t
i
es
.
E
v
en
s
o,
t
her
e
i
s
no
m
a
t
ur
e
t
ec
hn
ol
o
g
y
w
hi
c
h
c
a
n
s
at
i
s
f
y
t
he
r
equ
i
r
em
ent
s
of
r
eal
-
t
im
e
ac
qui
s
i
t
i
o
n
a
nd
hi
gh
ac
c
ur
ac
y
i
n
t
he
r
ec
o
gni
t
i
o
n
pr
o
g
r
es
s
due
t
o
t
h
e
d
i
v
er
s
i
t
y
of
hum
an
bod
i
es
and t
he c
om
pl
ex
i
t
y
of
ac
t
i
on pr
oc
es
s
es
[
4
]
.
A
n i
m
por
t
ant
r
eas
o
n i
s
ac
t
i
o
n c
har
ac
t
er
i
s
t
i
c
s
ar
e
d
i
f
f
i
c
ul
t
t
o r
epr
es
e
nt
.
M
an
y
s
c
hol
ar
s
pr
opos
e
d m
an
y
m
et
ho
ds
,
s
uc
h as
k
e
y
f
r
am
es
[
5
,
6
], s
p
a
ti
o
-
t
em
por
al
at
t
i
t
u
de m
odel
[
7
]
,
et
c
.
B
ut
t
he
i
r
ex
pr
es
s
i
on m
anner
s
ar
e a
l
w
a
y
s
c
om
pl
i
c
at
ed
an
d us
e
d
i
n s
pec
i
f
i
c
ac
t
i
on pr
oc
es
s
es
.
B
es
i
d
es
,
t
he c
l
as
s
i
f
i
c
at
i
on
m
et
hods
of
hu
m
an ac
t
i
on c
har
ac
t
er
i
s
t
i
c
s
al
s
o ha
v
e an
i
m
pac
t
on r
ec
ogn
i
t
i
on r
es
ul
t
s
,
an
d t
he
y
c
an be di
v
i
de
d i
nt
o t
w
o c
l
as
s
es
i
n g
ener
a
l
.
O
ne i
s
t
o c
l
as
s
i
f
y
ac
t
i
o
n c
har
ac
t
er
i
s
t
i
c
s
bas
ed o
n t
r
a
i
ne
d m
at
hem
at
i
c
al
m
odel
s
,
s
uc
h as
BP
neur
a
l
net
w
or
k
[
8
]
,
s
uppor
t
v
ec
t
or
m
ac
hi
ne (
S
V
M)
[
9
]
,
r
el
e
v
anc
e
v
ec
t
or
m
ac
hi
ne (
R
V
M)
[
1
0
]
,
et
c
.
T
he ot
her
i
s
b
as
ed
on
s
el
f
-
l
ear
n
i
ng
m
et
hods
,
m
ai
nl
y
i
nc
l
ud
i
ng
s
e
lf
-
or
gan
i
z
i
n
g c
om
pet
i
t
i
v
e
neur
a
l
ne
t
w
or
k
[1
1
], K
-
Me
ans
[
1
2
, 1
3
]
,
et
c
.
T
he f
or
m
er
m
et
hods
nee
d t
o
l
e
a
r
n t
he
ac
t
i
o
n
c
har
ac
t
er
i
s
t
i
c
s
i
nf
or
m
at
i
on
w
h
i
c
h
has
bee
n
m
ar
k
ed,
s
o
t
he
ex
pec
t
ed
o
ut
put
s
of
t
he
c
or
r
es
pondi
ng
i
n
put
s
n
eed
t
o b
e ac
h
i
e
v
ed
i
n
ad
v
a
nc
e.
W
hen
as
,
bec
aus
e of
t
h
e l
i
m
i
t
at
i
on
of
hum
an
c
ogni
t
i
v
e
ab
i
l
i
t
y
or
env
i
r
onm
ent
t
he
ex
pec
t
e
d
out
p
ut
s
ar
e
h
ar
d
t
o
ac
hi
e
v
ed
s
om
et
i
m
es
.
T
he l
at
t
er
m
et
hods
d
on
’
t
need
t
o k
no
w
t
h
e ex
p
ec
t
e
d out
put
s
,
and c
a
n r
ea
l
i
z
e
t
he
aut
om
at
i
c
cl
a
ssi
f
i
ca
t
i
o
n
o
f
a
ct
i
o
n
ch
a
r
a
ct
e
r
i
st
i
cs
b
y
c
om
par
i
ng
and
di
f
f
er
ent
i
a
t
i
n
g a
l
l
t
r
ai
ni
n
g s
am
pl
es
.
H
o
w
e
v
er
,
w
hen
t
he
i
ni
t
i
a
l
c
ondi
t
i
o
ns
of
s
el
f
-
or
gani
z
i
n
g
c
om
pet
i
t
i
v
e
neur
a
l
n
et
w
o
r
k
ar
e
s
et
,
t
h
e
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
IS
S
N
:
1
693
-
6
930
A
c
t
i
o
n R
ec
o
gn
i
t
i
on o
f
H
u
m
an
’
s
L
ow
er
Li
m
bs
B
as
ed o
n A
H
um
an J
o
i
nt
(
F
e
ng L
i
a
ng
)
1193
c
l
as
s
i
f
i
c
at
i
on
num
ber
of
ac
t
i
on
c
har
ac
t
er
i
s
t
i
c
s
i
s
har
d
t
o
c
ont
r
o
l
;
t
he
m
et
hod
s
uc
h
as
K
-
M
eans
ca
n
cl
a
ssi
f
y
ac
t
i
on c
har
ac
t
er
i
s
t
i
c
s
ac
c
or
di
ng t
o
t
he s
et
c
l
as
s
i
f
i
c
at
i
on
nu
m
ber
,
but
t
he
c
l
as
s
i
f
i
c
at
i
on
ef
f
ec
t
i
s
i
nf
l
uenc
ed
b
y
t
he
i
ni
t
i
a
l
s
et
t
i
ng
s
,
and
t
he
m
et
hod
i
s
eas
y
t
o
f
al
l
i
nt
o
l
oc
a
l
m
i
ni
m
u
m
v
al
ue
an
d s
ens
i
t
i
v
e t
o n
oi
s
e.
T
o r
eal
i
z
e t
he
f
as
t
and ac
c
ur
at
e ac
t
i
on
r
ec
ogn
i
t
i
on
of
hu
m
an’
s
l
o
w
er
l
i
m
bs
,
a nov
e
l
ac
t
i
on
r
ec
o
gni
t
i
o
n
m
et
hod
f
or
hum
an’
s
l
o
w
er
l
i
m
bs
i
s
pr
opos
ed.
O
n
l
y
t
he
hi
p
j
oi
nt
i
s
adopt
ed
as
t
he
r
ec
ogn
i
t
i
on
o
bj
ec
t
and
i
t
s
y
c
oor
di
na
t
es
ar
e
as
r
ec
ogn
i
t
i
on
par
am
et
er
.
B
ec
au
s
e
t
he
c
hang
e
c
ur
v
es
of
y
c
oor
di
n
a
t
es
i
n
di
f
f
er
ent
ac
t
i
ons
c
an r
epr
es
ent
di
f
f
er
ent
gr
ap
hi
c
s
c
har
ac
t
er
i
s
t
i
c
s
,
w
av
el
e
t
t
r
ans
f
or
m
i
s
i
nt
r
od
uc
ed t
o c
a
l
c
ul
a
t
e t
h
e ac
t
i
o
n c
har
ac
t
er
i
s
t
i
c
s
af
t
er
f
i
l
t
er
i
ng t
he ac
t
i
on
par
am
et
er
.
T
hen,
an
i
m
pr
o
v
ed
s
e
l
f
-
or
gan
i
z
i
ng
c
om
pet
i
t
i
v
e
neur
a
l
n
et
w
or
k
bas
ed
on
s
i
m
ul
at
e
d
anne
al
i
ng
a
l
g
or
i
t
hm
i
s
pr
opos
ed
t
o
c
l
as
s
i
f
y
ac
t
i
on
c
har
ac
t
er
i
s
t
i
c
s
aut
om
at
i
c
al
l
y
ac
c
or
di
ng
t
o
t
h
e
c
l
as
s
i
f
i
c
at
i
on n
um
ber
s
et
i
ni
t
i
al
l
y
.
F
i
na
l
l
y
,
a
n ac
t
i
o
n
r
ec
ogn
i
t
i
on
m
et
hod bas
ed
on H
MM
[
1
4]
is
i
nt
r
od
uc
ed t
o r
ea
l
i
z
e t
h
e ac
t
i
on r
ec
og
ni
t
i
o
n of
hu
m
a
n’
s
l
o
w
er
l
i
m
bs
w
i
t
h t
he c
ha
ng
e di
r
ec
t
i
on of
y
c
oor
di
n
at
es
.
B
ec
aus
e t
he
m
ot
i
on i
nf
or
m
at
i
on of
onl
y
a j
oi
nt
i
s
ado
pt
e
d,
t
he ac
t
i
on r
ec
ogn
i
t
i
on
m
et
hod c
an f
ac
e t
he per
s
o
nne
l
of
di
f
f
er
ent
s
i
z
es
,
an
d i
t
has
a f
as
t
c
al
c
ul
at
i
on s
p
e
ed
an
d a go
od
r
ec
ogni
t
i
o
n
ef
f
ec
t.
T
he hum
an
ac
t
i
on
d
at
a
c
om
es
f
r
o
m
C
MU
hum
an
ac
t
i
on dat
abas
e
of
C
ar
neg
i
e
Me
l
l
on U
ni
v
er
s
i
t
y
.
2.
R
e
sea
r
ch
M
et
h
o
d
2.
1.
A
c
t
i
o
n
C
h
a
r
a
c
te
r
i
s
ti
c
s
o
f H
u
m
a
n
’
s
L
o
w
e
r
L
i
m
b
s
2
.1
.1
.
T
h
e
S
e
l
e
c
ti
o
n
o
f
A
c
ti
o
n
P
a
r
a
m
e
te
r
I
n or
der
t
o
r
e
al
i
z
e t
h
e
ac
t
i
on
r
ec
ogn
i
t
i
on
of
hum
an’
s
l
o
w
er
l
i
m
bs
,
a
s
i
m
pl
i
f
i
ed s
k
el
et
on
s
t
r
uc
t
ur
e of
hum
an’
s
l
o
w
e
r
l
i
m
bs
i
s
i
nt
r
od
uc
ed a
nd
s
ho
w
n
i
n F
i
g
ur
e
1.
I
n t
h
e
f
i
gur
e,
W
C
S
deno
t
es
w
or
l
d
c
oor
d
i
na
t
e
s
y
s
t
em
and
LC
S
d
eno
t
e
s
l
oc
a
l
c
oor
d
i
n
at
e
s
y
s
t
em
.
B
ec
aus
e
t
he
m
ot
i
on pr
oc
es
s
es
of
ot
h
er
j
oi
nt
s
ar
e ar
oun
d h
i
p j
o
i
nt
w
hi
c
h c
a
n b
e as
t
he r
o
ot
j
oi
n
t
,
w
e c
h
oos
e
i
t
t
o
r
epr
es
e
nt
t
h
e
ac
t
i
on
pr
o
c
es
s
es
of
hu
m
an’
s
l
o
w
er
l
i
m
b
s
.
W
hen
peop
l
e
w
al
k
or
r
un
i
n
a
r
oom
,
t
he r
eg
ul
at
i
ons
ar
e
di
f
f
i
c
ul
t
f
ound i
n t
h
e c
ha
nges
of
t
he
l
oc
at
i
ons
of
h
i
p j
o
i
nt
i
n t
he
x
and
z
d
ir
ec
t
i
ons
due
t
o
t
h
e
unc
er
t
ai
n
t
y
of
m
ot
i
on
di
r
ec
t
i
on
.
T
her
ef
or
e,
y
c
oor
di
nat
es
of
hi
p
j
oi
n
t
i
n
t
he
W
C
S
ar
e
c
hos
en
t
o
r
ec
og
ni
z
e
e
ac
h
ac
t
i
on
of
hum
an’
s
l
o
w
er
l
i
m
bs
and
t
he
y
c
o
or
di
n
at
e
of
hi
p
j
oi
nt
at
t
i
m
e
t
i
s
deno
t
ed
a
s
y
(
t
)
.
I
n
t
he
pap
er
,
t
h
e
s
a
m
pl
e
f
r
equen
c
y
of
m
ot
i
on
c
apt
ur
e
d
at
a
i
s
120 H
z
and
t
he
un
i
t
of
t
i
m
e
t
i
s
t
h
e s
am
pl
e i
nt
er
v
al
1/
12
0 s
.
F
ig
ur
e
1
.
A
s
i
m
pl
i
f
i
ed s
k
el
et
on s
t
r
uc
t
ur
e
of
hum
an’
s
lo
w
e
r
lim
b
s
.
I
n or
der
t
o f
ac
i
l
i
t
a
t
e ex
t
r
ac
t
i
ng ac
c
ur
at
e ac
t
i
o
n c
har
ac
t
er
i
s
t
i
c
s
,
B
ut
t
er
w
or
t
h f
i
l
t
er
of
w
h
i
c
h
t
he c
ut
of
f
f
r
equenc
y
i
s
s
et
as
0.
1 r
ad/
s
i
s
ad
opt
ed t
o f
i
l
t
er
y
(
t
)
,
and t
h
e f
i
l
t
er
i
ng r
es
ul
t
s
ar
e
deno
t
ed
w
i
t
h
y’
(
t
).
T
ak
e
mul
t
i
pl
e
s
et
s
ac
t
i
o
n
dat
a
of
hum
an’
s
l
ow
er
l
i
m
bs
f
or
ex
a
m
pl
e,
s
uc
h
as
w
al
k
i
ng,
r
u
nn
i
ng
and
j
um
pi
ng a
nd t
hei
r
f
i
l
t
er
i
ng r
es
u
l
t
s
ar
e s
ho
w
n i
n F
i
g
ur
e
2.
l
_
to
l
_
an
l
_
kn
l
_
th
r
_
to
r
_
an
r
_
kn
r
_
th
hip
(
root
)
x
y
z
WCS
LCS
r
_
to
y
z
LCS
l
_
to
LCS
l
_
an
x
y
z
x
y
z
x
y
z
x
y
z
x
y
z
x
y
z
x
y
z
x
LCS
l
_
kn
LCS
r
_
an
LCS
r
_
kn
LCS
r
_
th
LCS
l
_
th
o
y
z
x
LCS
hip
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
14
,
N
o
.
3,
S
ept
em
ber
2016
:
11
92
–
1
202
1194
F
i
gur
e
2.
C
han
ge c
ur
v
es
of
y
’(
t
)
i
n
di
f
f
er
ent
a
c
t
io
n
pr
oc
es
s
es
I
n F
i
g
ur
e
2,
t
h
er
e
ar
e s
i
m
i
l
ar
gr
ap
hi
c
s
c
har
ac
t
er
i
s
t
i
c
s
of
t
he s
am
e ac
t
i
ons
,
and
gr
aph
i
c
s
c
har
ac
t
er
i
s
t
i
c
s
of
di
f
f
er
ent
ac
t
i
ons
h
av
e o
bv
i
ous
d
i
s
t
i
nc
t
i
on.
T
he c
ha
nge
di
r
ec
t
i
on
of
y
’(
t
)
i
s
c
al
c
ul
a
t
ed
as
f
ol
l
o
w
s
:
(
1)
(
)
()
yt
yt
y
t
T
′′
+
−
∆=
(
1
)
()
,
()
0
()
()
0,
(
)
0
y
t
y
t
Dy
t
y
t
y
t
∆
∆≠
=
∆
∆=
(
2
)
W
h
er
e,
()
y
t
∆
m
eans
t
he v
e
l
oc
i
t
y
of
hi
p j
oi
nt
i
n t
he
y
di
r
ec
t
i
on at
t
i
m
e
t
;
()
Dy
t
r
epr
es
ent
s
t
he
c
hange
di
r
ec
t
i
o
n of
y
’(
t
).
2
.1
.2
.
T
h
e
R
e
p
r
e
s
e
n
ta
ti
o
n
o
f
A
c
ti
o
n
C
h
ar
act
er
i
st
i
c
s
S
i
g
na
l
s
’
f
r
ac
t
al
c
har
ac
t
er
i
s
t
i
c
s
r
epr
es
ent
t
hei
r
s
el
f
-
s
im
ila
r
it
y
.
T
im
e
-
f
r
equenc
y
pr
o
p
er
t
y
of
s
i
gna
l
s
c
an
be
obs
er
v
ed
ex
ped
i
ent
l
y
b
as
ed
on
w
av
el
et
t
r
ans
f
or
m
,
and t
h
e s
i
gna
l
s
’
s
el
f
-
s
i
m
ila
r
it
y
c
oef
f
i
c
i
ent
s
i
n
d
i
f
f
er
ent
s
c
al
es
c
an
be
r
epr
es
e
nt
e
d
w
i
t
h
t
he
t
i
m
e
-
f
r
equenc
y
pr
o
per
t
y
.
T
he
hi
gher
s
e
lf
-
s
i
m
i
l
ar
i
t
y
i
s
,
t
he
l
ar
ger
i
t
s
c
oef
f
i
c
i
ent
s
ar
e.
T
her
ef
or
e,
s
i
gnal
s
’
s
e
l
f
-
s
i
m
ila
r
it
y
c
o
e
f
f
ic
ie
n
t
s
in
di
f
f
er
ent
s
c
al
es
ar
e
i
nt
r
o
du
c
ed
t
o
r
epr
es
ent
t
he
c
ha
ng
es
of
s
i
gnal
m
or
phol
og
y
[1
5
]
.
T
he
f
or
m
ul
a
of
w
a
v
e
l
et
t
r
ans
f
or
m
i
s
as
f
ol
l
o
w
s
:
(
)
,
1
()
d
a
b
Q
tb
F
ft
t
a
a
ψ
−
=
∫
(
3
)
W
h
er
e
()
ft
m
eans
t
he
ac
t
i
o
n
p
ar
am
et
er
y’
(
t
)
,
,
a
b
F
m
eans
t
he
w
av
el
e
t
t
r
ans
f
or
m
c
oef
f
i
c
i
ent
s
of
()
ft
,
a
m
eans
t
he s
c
al
e f
ac
t
or
,
b
m
eans
t
he t
r
ans
l
at
i
on f
ac
t
or
,
Q
m
eans
t
he s
i
gn
al
s
p
ac
e and
ψ
m
eans
t
he w
a
v
el
et
b
as
i
s
f
unc
t
i
on
.
F
ig
ur
e
3
.
W
av
el
et
c
o
ef
f
i
c
i
ent
s
of
y
’(
t
)
i
n d
i
f
f
er
ent
ac
t
i
o
n pr
oc
es
s
es
.
b
/(
1
/
120
s
)
a
100
200
300
400
500
2
10
18
26
34
42
50
58
Walk
a
b
/(
1
/
120
s
)
50
100
150
2
10
18
26
34
42
50
58
Run
a
b
/(
1
/
120
s
)
100
200
300
2
10
18
26
34
42
50
58
Jump
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
IS
S
N
:
1
693
-
6
930
A
c
t
i
o
n R
ec
o
gn
i
t
i
on o
f
H
u
m
an
’
s
L
ow
er
Li
m
bs
B
as
ed o
n A
H
um
an J
o
i
nt
(
F
e
ng L
i
a
ng
)
1195
C
hoos
e
c
oi
f
3
w
a
v
e
l
et
t
o
t
r
a
ns
f
or
m
t
he
ac
t
i
on
p
ar
am
et
er
s
y’
(
t
)
i
n
di
f
f
er
ent
s
c
al
es
of
2,
4,
6,
…
,
64.
12
,,
,
(
(
),
(
),
,
(
))
n
a
b
a
b
a
b
Yt
Yt
Yt
i
s
us
ed t
o de
n
ot
e t
he t
r
a
ns
f
or
m
at
i
on r
es
ul
t
s
of
y’
(
t
)
,
and
n
m
eans
t
he s
c
al
e num
ber
w
hi
c
h
i
s
equ
al
t
o 32
i
n
t
he pa
per
.
T
he t
r
ans
f
or
m
at
i
on r
es
u
l
t
s
i
n
di
f
f
er
ent
ac
t
i
on
pr
oc
es
s
es
ar
e s
h
o
w
n
i
n
F
i
g
ur
e
3.
I
n t
h
e f
i
gur
e
,
i
t
i
s
f
ou
nd t
h
at
t
h
e
w
av
el
e
t
t
r
ans
f
or
m
c
oef
f
i
c
i
ent
s
of
di
f
f
er
ent
ac
t
i
on
pr
oc
es
s
es
c
an
r
epr
es
ent
di
f
f
er
ent
v
ar
i
at
i
o
n
r
ul
es
.
2.
2
.
T
h
e
C
l
a
s
s
i
fi
c
a
ti
o
n
M
e
th
o
d
o
f
A
c
ti
o
n
C
h
a
r
a
c
t
e
r
i
s
ti
c
s
2
.2
.1
.
A
n
I
m
p
r
o
v
e
d
S
e
l
f
-
O
r
g
a
n
i
z
i
n
g
N
e
u
r
a
l
N
e
tw
o
r
k
B
a
s
e
d
o
n
S
i
m
u
l
a
t
e
d
A
n
n
e
a
l
i
n
g
A
l
g
o
r
i
th
m
T
her
e
ar
e
t
w
o
r
eq
ui
r
em
ent
s
i
n
t
he
c
l
as
s
i
f
i
c
at
i
on
pr
oc
e
s
s
of
t
he
ac
t
i
on
c
h
ar
ac
t
er
i
s
t
i
c
s
o
f
hum
an’
s
l
i
m
bs
.
O
ne i
s
ac
t
i
on c
har
ac
t
er
i
s
t
i
c
s
of
hi
gh s
i
m
i
l
ar
i
t
y
c
an be c
l
as
s
i
f
i
e
d
as
t
he s
a
m
e
c
l
as
s
;
anot
her
i
s
c
l
as
s
i
f
i
c
at
i
on r
es
ul
t
s
ar
e r
el
at
i
v
el
y
ev
en
,
bec
a
us
e i
t
c
an r
ed
uc
e t
h
e di
s
t
ur
banc
e
of
noi
s
e
s
i
gn
al
s
a
nd
ens
ur
e
t
h
e
r
epr
es
e
nt
at
i
v
en
es
s
of
eac
h
c
h
ar
ac
t
er
i
s
t
i
c
s
pac
e.
T
her
ef
or
e,
an
i
m
pr
ov
ed s
e
l
f
-
or
gani
z
i
ng
n
eur
al
ne
t
w
or
k
bas
ed
on
s
i
m
ul
at
ed a
nne
al
i
ng
a
l
gor
i
t
h
m
i
s
pr
opos
ed
t
o
cl
a
ssi
f
y
t
h
e
a
ct
i
o
n
ch
a
r
a
ct
e
r
i
st
i
cs o
f
hum
an’
s
l
i
m
bs
.
F
i
r
s
t
of
al
l
,
s
et
c
l
as
s
i
f
i
c
at
i
o
n n
um
ber
S
, th
e
n
us
e
c
l
as
s
i
f
i
c
at
i
on
s
pac
e
n
um
ber
s
X
=
{
x
1
,
x
2
,
…
,
x
S
}
t
o
d
enot
e
c
l
as
s
i
f
i
c
at
i
o
n
r
es
ul
t
s
an
d
c
l
as
s
i
f
i
c
at
i
on ent
r
o
p
y
Ec
t
o
r
epr
es
ent
t
h
e un
i
f
or
m
i
t
y
o
f
c
l
as
s
i
f
i
c
at
i
on
r
es
ul
t
s
.
T
he
c
al
c
ul
at
i
o
n of
Ec
i
s a
s f
o
l
l
o
w
s:
2
1
lo
g
S
ii
i
E
c
x
x
=
=
−
∑
(
4
)
T
he t
r
ai
ni
ng
pr
oc
es
s
of
t
he
neur
a
l
n
et
w
or
k
i
s
f
ol
l
o
w
i
ng
:
S
t
ep 1
:
I
ni
t
i
a
l
i
z
e n
et
w
or
k
.
T
he i
n
put
l
a
y
er
i
s
c
o
m
pr
i
s
ed of
R
neur
ons
and t
h
e c
om
p
et
i
t
iv
e
l
a
y
er
i
s
c
om
pr
i
s
ed of
S
n
eu
r
ons
,
t
he
i
n
put
m
at
r
i
x
of
t
r
ai
ni
n
g s
am
pl
es
i
s
de
not
e
d as
:
11
12
1
21
22
2
12
R
R
Q
Q
QR
p
p
p
p
p
p
P
pp
p
=
(
5)
W
h
er
e
Q
i
s
t
he nu
m
ber
of
t
r
ai
ni
ng s
am
pl
es
;
R
equa
l
s
t
he s
c
al
e num
ber
of
w
av
el
e
t
t
r
ans
f
or
m
;
P
ij
deno
t
es
t
he
j
t
h
i
np
ut
of
t
he
i
t
h t
r
a
i
ni
ng s
am
pl
e
12
[]
i
i
i
iR
p
pp
p
=
,
1,
2
,
,
i
Q
=
and
1,
2
,
,
jR
=
.
S
et
i
n
i
t
i
al
t
em
per
at
ur
e
T
1
,
c
ool
i
ng
t
em
per
at
ur
e
T
2
,
u
ni
t
t
em
per
at
ur
e
v
ar
i
at
i
on
Δ
T
,
i
s
ot
her
m
al
c
hang
e num
ber
t
1
,
w
ei
g
ht
adj
us
t
m
ent
c
oef
f
i
c
i
ent
N
,
an
d t
he i
dea
l
i
nf
or
m
at
i
on num
ber
of
c
har
ac
t
er
i
s
t
i
c
s
pac
e
[
/]
a
QS
=
.
L
et
t
he
c
ur
r
ent
t
em
per
at
ur
e
1
TT
=
,
ut
i
l
i
z
e
Ψ
t
o d
en
ot
e
t
he
c
ol
l
ec
t
i
on
of
al
l
t
h
e ac
t
i
on
c
har
ac
t
er
i
s
t
i
c
s
,
den
ot
e t
he d
i
v
i
d
ed c
h
ar
ac
t
er
i
s
t
i
c
s
pac
es
12
{,
,
,
}
S
θθ
θ
Φ=
w
h
er
e
i
θ
=
∅
at
t
he
i
ni
t
i
a
l
t
i
m
e
and
1,
2
,
,
i
S
=
,
and
l
et
t
h
e
i
n
i
t
i
al
i
nf
or
m
at
i
on
num
ber
of
eac
h c
har
ac
t
er
i
s
t
i
c
s
pac
e
0
i
x
=
,
1,
2
,
,
i
S
=
.
T
he i
ni
t
i
a
l
c
on
nec
t
i
on
w
ei
g
ht
s
of
t
he n
et
w
or
k
ar
e de
no
t
ed as
:
12
[]
R
SR
IW
w
w
w
×
=
(
6)
W
h
er
e
1
[
1/
(
)
1/
(
)
1/
(
)
]
iS
w
SR
SR
SR
×
′
=
××
×
,
1,
2
,
,
iR
=
.
T
he i
ni
t
i
a
l
t
hr
es
h
ol
ds
of
t
he
net
w
or
k
ar
e de
not
e
d as
:
1
l
n
(1
/
)
1
l
n
(1
/
)
1
l
n
(1
/
)
1
[]
SS
S
S
be
e
e
−−
−
×
′
=
(
7)
S
et
t
he l
ear
ni
n
g
s
pe
ed of
c
onnec
t
i
on
w
e
i
ght
s
as
α
and
t
he
l
ear
n
i
n
g s
pe
ed of
t
hr
es
hol
ds
as
β
.
S
t
ep
2:
C
a
l
c
ul
a
t
e t
he
w
i
n
n
i
ng
ne
ur
on.
S
e
l
ec
t
eac
h
t
r
ai
n
i
ng
s
am
pl
e
m
p
(
1,
2
,
,
mQ
=
)
or
der
l
y
.
A
c
c
or
di
n
g t
o
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
14
,
N
o
.
3,
S
ept
em
ber
2016
:
11
92
–
1
202
1196
12
1
()
R
i
mj
i
j
i
j
n
p
IW
b
=
=
−
−+
∑
,
1,
2
,
,
i
S
=
(
8)
C
al
c
u
l
at
e
t
h
e
i
np
ut
s
of
c
o
m
pet
i
t
i
v
e
n
eur
ons
.
I
n
f
or
m
ul
a
(
8)
,
1
i
n
de
not
es
t
he
ou
t
p
ut
of
t
h
e
i
th
c
o
m
pet
i
t
i
v
e neur
on;
mj
p
deno
t
es
t
he
j
t
h i
npu
t
of
t
he t
r
ai
n
i
ng s
am
pl
e
m
p
;
ij
IW
denot
es
t
he
c
onnec
t
i
on
w
e
i
g
ht
bet
w
een
t
he
i
t
h
c
om
pet
i
t
i
v
e
n
eur
o
n
and
t
he
j
t
h
i
n
put
n
eur
on
;
i
b
deno
t
es
t
he
t
hr
es
hol
d of
t
h
e
i
t
h c
om
pet
i
t
i
v
e n
eur
on.
W
h
en t
he
k
t
h c
om
pet
i
t
i
v
e n
eur
on s
at
i
s
f
i
es
t
he
f
ol
l
o
w
i
n
g
:
11
ma
x
(
)
k
i
nn
=
,
1,
2
,
,
i
S
=
,
{1
,
2
,
,
}
kS
∈
(
9)
I
t
’
s
s
een as
t
he
w
i
nni
ng n
eur
on.
U
p
dat
e t
he r
el
at
e
d
c
har
ac
t
er
i
s
t
i
c
s
pac
e and
i
t
s
i
nf
or
m
at
i
on
num
ber
as
f
ol
l
o
w
s
:
kk
m
p
θθ
=
,
1
kk
x
x
=
+
(
10)
S
t
ep
3:
U
pd
at
e
t
he
w
e
i
gh
t
s
and t
hr
es
ho
l
ds
b
as
ed
o
n s
i
m
ul
at
ed
an
nea
l
i
ng
al
g
or
i
t
hm
.
U
pdat
e t
h
e
w
e
i
g
ht
s
an
d t
hr
es
hol
d of
t
he
w
i
n
i
n
g n
eur
on
k
s
epar
at
e
l
y
as
f
ol
l
o
w
s
:
(
)
k
j
k
j
m
j
k
j
IW
IW
p
IW
r
a
n
d
α
=
+
−×
,
1,
2
,
,
jR
=
(
11)
1
ln
[
(
1
)
ln
(
)
]
k
e
b
r
and
k
b
e
β
βα
−
−
−
+
××
=
(
12)
W
h
er
e
r
and
i
s
a
r
a
ndom
num
ber
,
bel
ongs
t
o
[
0,
1]
an
d
f
ol
l
o
w
s
un
i
f
or
m
di
s
t
r
i
but
i
on.
T
he
i
nt
r
od
uc
t
i
on
of
t
he
r
and
om
num
ber
m
a
k
es
t
he
v
ar
i
at
i
o
n
pr
oc
es
s
es
of
t
h
e
w
ei
ght
s
c
an
s
i
m
ul
at
e
t
he r
an
dom
c
hange
pr
oc
es
s
es
of
m
ol
ec
ul
es
i
n t
h
er
m
al
ac
t
i
o
n.
A
f
t
er
al
l
t
h
e
t
r
a
i
n
i
ng
s
am
pl
es
ar
e
s
t
ud
i
e
d
o
nc
e
,
c
a
l
c
ul
at
e
c
l
us
t
er
c
ent
er
s
1
i
w
an
d
2
i
w
of
c
har
ac
t
er
i
s
t
i
c
s
pac
e
i
θ
and i
t
s
c
o
m
pl
em
ent
ar
y
s
et
i
θ
a
s
f
o
llo
w
s
:
1
1,
2
(
),
,
1,
2
,
,
(
),
i
m
im
i
S
ij
j
ji
i
m
im
i
w
P
l
e
ngt
h
P
i
S
w
P
l
e
ngt
h
P
θθ
θθ
θθ
=
≠
=
∈
=
=
=
∈
∑
∑
(
13)
T
hen,
ac
c
or
di
n
g
t
o
t
he
i
nf
o
r
m
at
i
on
n
um
ber
of
eac
h
c
har
ac
t
er
i
s
t
i
c
s
pac
e
,
u
pdat
e
al
l
t
he
c
onnec
t
i
on
w
ei
ght
s
a
nd t
hr
es
hol
ds
.
T
he m
et
hod i
s
as
f
ol
l
o
w
s
:
a)
W
hen
()
i
l
e
ngt
h
a
θ
>
,
{1
,
2
,
,
}
iS
∈
1
1
1
(
)
1
(
)(
)
/
l
e
ngt
h
a
T
ij
ij
ij
ij
IW
IW
e
e
IW
w
r
a
n
d
N
θ
−
−
−
=
+
+
−×
(
14)
1
1
ln
[
(
1
)
ln
(
)
(
(
)
)
]
k
e
b
r
and
l
e
ngt
h
a
i
be
β
βα
θ
−
−
−
+
××
×
−
=
(
15)
W
h
er
e
1
ij
w
i
s
t
he
j
t
h
w
ei
g
ht
of
1
i
w
and
1,
2
,
,
jR
=
. If
T
i
s
hi
g
h or
()
i
l
e
ngt
h
a
θ
−
is
la
r
g
e
,
ij
IW
w
i
ll
c
hange
r
el
at
i
v
e
l
y
qu
i
c
k
l
y
i
n t
he
di
r
ec
t
i
o
n a
w
a
y
f
r
om
t
he c
l
us
t
er
c
ent
er
1
ij
w
. W
i
th
T
dr
ops
or
()
i
l
e
ngt
h
a
θ
−
d
ec
r
eas
es
,
t
h
e c
ha
nge
pr
o
c
es
s
w
i
l
l
bec
om
e s
l
o
w
.
T
he adj
us
t
m
ent
pr
oc
es
s
of
i
b
is
onl
y
r
e
l
at
e
d t
o
()
i
l
e
ngt
h
a
θ
−
.
I
f
t
he a
dj
us
t
m
ent
a
m
ount
i
s
t
o
o l
ar
ge,
t
h
e c
ha
nge
pr
oc
es
s
o
f
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
IS
S
N
:
1
693
-
6
930
A
c
t
i
o
n R
ec
o
gn
i
t
i
on o
f
H
u
m
an
’
s
L
ow
er
Li
m
bs
B
as
ed o
n A
H
um
an J
o
i
nt
(
F
e
ng L
i
a
ng
)
1197
w
ei
g
ht
s
w
i
l
l
be
di
f
f
i
c
ul
t
t
o
c
onv
er
genc
e,
s
o
w
e
i
gh
t
adj
us
t
m
ent
c
oef
f
i
c
i
ent
N
i
s
i
nt
r
od
uc
ed t
o
c
ont
r
ol
t
he
adj
us
t
m
ent
s
peed of
w
e
i
ght
s
.
b)
W
hen
()
i
l
e
ngt
h
a
θ
<
,
{1
,
2
,
,
}
iS
∈
1
1
1
(
)
2
(
)(
)
/
i
a
l
e
ngt
h
T
ij
ij
ij
IW
IW
e
e
w
IW
r
a
n
d
N
θ
−
−
−
=
++
−
×
(
16)
1
1
ln
[
(
1
)
ln
(
)
(
(
)
]
k
e
b
r
and
a
l
e
ngt
h
i
be
β
βα
θ
−
−
−
−
××
×
−
=
(
17)
W
h
er
e
2
ij
w
i
s
t
he
j
t
h w
ei
gh
t
o
f
2
i
w
and
1,
2
,
,
jR
=
. If
T
is
h
ig
h
o
r
()
i
l
e
ngt
h
a
θ
−
i
s
l
ar
ge
,
ij
IW
w
il
l
c
hange
r
el
at
i
v
el
y
q
ui
c
k
l
y
i
n
t
he
d
i
r
ec
t
i
o
n c
l
os
e t
o t
he
c
l
us
t
er
c
ent
er
2
i
w
.
W
i
t
h t
he t
e
m
per
at
ur
e
dr
ops
or
()
i
l
e
ngt
h
a
θ
−
dec
r
eas
es
,
t
he
c
hange
pr
oc
es
s
w
i
l
l
bec
om
e
s
l
ow
.
T
he
adj
us
t
m
e
nt
pr
oc
es
s
of
i
b
i
s
onl
y
r
e
l
at
ed t
o
()
i
l
e
ngt
h
a
θ
−
.
c
)
W
hen
()
i
l
e
ngt
h
a
θ
=
,
{1
,
2
,
,
}
iS
∈
1
s
gn(
0.5
)
/
T
ij
ij
I
W
I
W
e
r
and
r
and
N
−
=
+×
×
−
(
18)
1
l
n[
(
1
)
l
n(
)
s
gn(
0.
5
)
]
k
e
b
r
and
r
and
i
be
β
βα
−
−
−
+
××
×
−
=
(
19)
W
h
er
e
1,
2
,
,
jR
=
.
W
hen
T
i
s
hi
gh,
ij
IW
w
i
l
l
c
hang
e
r
andom
l
y
an
d
r
el
at
i
v
el
y
qu
i
c
k
l
y
;
w
hen
T
dr
ops
,
ij
IW
w
i
l
l
c
ha
nge s
l
o
w
l
y
.
T
he adj
us
t
m
ent
pr
o
ce
ss o
f
i
b
i
s
n’
t
r
el
a
t
ed
t
o
T
.
S
t
ep
4:
I
t
er
at
e t
h
e c
om
put
at
i
ons
of
t
he n
eur
a
l
ne
t
w
or
k
.
a)
A
c
c
or
di
ng
t
o
t
he
i
s
ot
h
er
m
al
c
hange
num
ber
t
1
,
upd
at
e
t
h
e
n
eur
a
l
n
et
w
or
k
b
y
ex
ec
ut
i
n
g S
t
ep
2 a
nd
3,
a
n
d c
al
c
u
l
at
e
t
he
c
han
ges
of
t
he c
l
as
s
i
f
i
c
at
i
on
ent
r
o
p
y
Ec
.
b)
Let
TT
T
=
−∆
. If
2
TT
>
,
t
ur
n t
o a)
;
i
f
no,
end t
h
e i
t
er
at
i
v
e c
om
put
at
i
on pr
oc
es
s
,
t
hen
out
p
ut
w
ei
gh
t m
a
tr
i
x
IW
,
t
hr
e
s
hol
ds
b
,
t
he i
nf
or
m
at
i
on nu
m
ber
o
f
eac
h c
har
ac
t
er
i
s
t
i
c
s
pac
e
and t
he c
l
as
s
i
f
i
c
at
i
o
n e
nt
r
o
p
y
Ec
.
2
.2
.2
.
T
h
e
C
l
a
s
s
i
fi
c
a
ti
o
n
o
f
A
c
ti
o
n
C
h
a
r
a
c
te
r
i
s
ti
c
s
T
a
k
e ac
t
i
on
c
har
ac
t
er
i
s
t
i
c
s
of
s
om
e
m
ot
i
on c
a
pt
ur
e
d
at
a f
or
ex
am
pl
e
t
o
i
nt
r
odu
c
e t
he
c
l
as
s
i
f
i
c
at
i
on m
et
hod b
as
e
d on t
h
e i
m
pr
ov
ed s
el
f
-
or
g
ani
z
i
ng n
eur
a
l
net
w
or
k
.
T
r
ai
n
i
ng
dat
a
i
s
c
o
m
pr
i
s
ed
of
di
f
f
er
ent
k
i
nds
of
hu
m
an
ac
t
i
o
n
da
t
a,
a
n
d
eac
h
k
i
nd
c
o
ns
i
s
t
s
of
3
s
et
s
of
hum
an
ac
t
i
on d
at
a.
T
her
e ar
e 1
1
90 gr
ou
ps
of
w
al
k
i
ng c
ha
r
ac
t
er
i
s
t
i
c
dat
a
,
505 gr
ou
ps
of
r
unni
ng
c
har
ac
t
er
i
s
t
i
c
dat
a
,
a
nd
59
2 gr
ou
ps
of
j
um
pi
ng c
har
ac
t
er
i
s
t
i
c
d
at
a.
F
i
r
s
t
of
a
l
l
,
t
he
t
r
ai
n
i
ng
dat
a
i
s
nor
m
al
i
z
e
d as
f
ol
l
o
w
s
:
1
mi
n
(
)
(
ma
x
(
)
mi
n
(
)
)
j
ij
ij
jj
pp
p
p
pb
−
′
=
−
×
(
20)
W
h
er
e
12
[]
j
j
j
Qj
p
pp
p
′
=
, i
t i
s
th
e
j
t
h c
ol
u
m
n
v
ec
t
or
of
t
he i
nput
m
at
r
i
x
P
,
and
b
1
is
c
ont
r
ac
t
i
o
n c
oef
f
i
c
i
ent
.
Let
1
1.1
b
=
i
n t
he p
aper
.
T
he i
ni
t
i
al
s
et
t
i
n
gs
ar
e as
t
he
f
o
llo
w
in
g
s
:
0.02
α
=
,
0.01
β
=
,
1
1
T
=
,
2
0.95
T
=
,
0.01
T
∆
=
,
1
400
t
=
and
10
S
=
.
T
r
ai
n
t
h
e i
m
pr
ov
e
d
s
el
f
-
or
gani
z
i
ng neur
al
ne
t
w
or
k
.
T
he c
hang
e of
t
he
c
l
as
s
i
f
i
c
at
i
on
e
nt
r
op
y
Ec
is
s
h
o
w
n
in
F
ig
ur
e
4 a
nd
t
he
c
hange
of
t
h
e c
har
ac
t
er
i
s
t
i
c
num
ber
Ns
of
a c
om
pet
i
t
i
v
e
neur
o
n i
s
s
h
o
w
n i
n F
i
g
ur
e
5.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
14
,
N
o
.
3,
S
ept
em
ber
2016
:
11
92
–
1
202
1198
F
ig
ur
e
4
.
T
he c
ha
nge of
c
la
s
s
if
ic
a
t
io
n
e
nt
r
op
y
Ec
F
ig
ur
e
5
.
T
he
c
ha
nge of
c
har
ac
t
er
i
s
t
i
c
n
um
ber
Ns
I
n
F
ig
ur
e
4
a
nd
5,
s
om
e
r
egu
l
at
i
ons
c
an
be
f
ound
.
E
ar
l
y
i
n
t
he
i
t
er
at
i
v
e
c
om
put
at
i
o
ns
,
t
he c
l
as
s
i
f
i
c
at
i
on en
t
r
op
y
Ec
and t
he c
har
ac
t
er
i
s
t
i
c
nu
m
ber
Ns
c
hange qu
i
c
k
l
y
;
w
i
t
h t
he i
nc
r
eas
e
of
t
r
ai
ni
ng
t
i
m
e,
t
hei
r
c
hang
e
pr
oc
es
s
es
bec
om
e
s
l
ow
and
t
en
d
t
o
c
on
v
er
ge
.
A
f
t
er
t
he
i
m
pr
ov
ed
s
e
lf
-
or
gani
z
i
ng ne
ur
al
ne
t
w
or
k
i
s
t
r
ai
ned,
t
h
e net
w
or
k
i
s
us
ed t
o c
l
as
s
i
f
y
ot
her
ac
t
i
on
c
har
ac
t
er
i
s
t
i
c
s
of
hum
an’
s
l
o
w
er
l
i
m
bs
.
T
he c
l
as
s
i
f
i
c
at
i
on r
es
ul
t
s
ar
e s
ho
w
n
i
n F
i
g
ur
e
6.
F
i
gur
e 6.
S
om
e
c
la
s
s
if
ic
a
t
io
n
r
es
ul
t
s
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
IS
S
N
:
1
693
-
6
930
A
c
t
i
o
n R
ec
o
gn
i
t
i
on o
f
H
u
m
an
’
s
L
ow
er
Li
m
bs
B
as
ed o
n A
H
um
an J
o
i
nt
(
F
e
ng L
i
a
ng
)
1199
S
om
e c
onc
l
us
i
ons
c
an b
e
ac
hi
e
v
ed
i
n F
i
g
ur
e
6.
T
he c
l
as
s
i
f
i
c
at
i
o
n r
es
ul
t
s
of
s
a
m
e
ac
t
i
ons
ha
v
e s
i
m
i
l
ar
c
han
g
e c
har
ac
t
er
i
s
t
i
c
,
b
ut
t
he c
ha
nge pr
oc
es
s
es
of
t
he c
l
as
s
i
f
i
c
at
i
on r
es
ul
t
s
f
or
di
f
f
er
ent
ac
t
i
ons
ar
e
qu
i
t
e d
i
f
f
er
ent
.
2.
3
.
A
c
t
i
o
n
R
e
c
o
g
n
i
ti
o
n
o
f H
u
m
a
n
’
s
L
o
w
e
r
L
i
m
b
s
2
.
3
.
1
.
A
c
ti
o
n
R
e
c
o
g
n
i
ti
o
n
B
a
s
e
d
o
n
H
M
M
B
ec
aus
e H
MM has
a s
t
r
ong abi
l
i
t
y
of
bui
l
d
i
ng s
eq
u
ent
i
al
m
odel
and i
t
i
s
a k
i
nd of
d
y
n
am
i
c
i
nf
or
m
at
i
on pr
oc
e
s
s
i
ng m
et
hod bas
e
d o
n
s
eque
nt
i
al
ac
c
um
ul
at
i
v
e pr
obab
i
l
i
t
y
,
i
t
i
s
i
nt
r
od
uc
ed t
o r
ec
og
ni
z
e
t
he
ac
t
i
o
n pr
oc
es
s
es
of
hum
an’
s
l
o
w
er
l
i
m
bs
.
a)
T
he s
t
at
es
of
H
MM ar
e deno
t
es
as
12
,,
,
N
θθ
θ
,
w
her
e
N
i
s
t
h
e
num
ber
of
t
he s
t
at
es
.
T
he s
t
at
e of
t
he m
odel
at
t
i
m
e
t
i
s
denot
e
d
as
t
q
,
an
d
12
(,
,
,
)
t
N
q
θθ
θ
∈
.
b)
T
he obs
er
v
at
i
o
ns
w
hi
c
h
eac
h s
t
at
e i
s
r
el
a
t
ed t
o ar
e
deno
t
ed as
1
2
,,
,
M
VV
V
,
w
her
e
M
i
s
t
he num
ber
of
t
he obs
er
v
at
i
o
ns
.
T
he obs
er
v
a
t
i
on a
t
t
i
m
e
t
i
s
denot
e
d as
t
O
,
and
1
2
(,
,
,
)
tM
O
VV
V
∈
.
c
)
T
he
pr
obab
i
l
i
t
y
di
s
t
r
i
but
i
on
of
t
he
i
n
i
t
i
al
s
t
a
t
es
i
s
r
e
pr
es
ent
e
d
w
i
t
h
12
(,
,
,
)
N
π
π
π
π
=
,
w
her
e
1
()
ii
Pq
πθ
=
=
(
1,
2
,
,
iN
=
)
,
and
1
q
is
t
h
e
s
t
at
e a
t
t
he i
n
i
t
i
al
t
i
m
e.
d)
T
he s
t
at
e t
r
ans
i
t
i
on
pr
obab
i
l
i
t
y
m
at
r
i
x
i
s
deno
t
ed as
*
()
ij
N
N
Aa
=
,
w
her
e
1
()
ij
t
j
t
i
a
Pq
q
θθ
+
=
=
=
.
e)
T
he
pr
oba
bi
l
i
t
y
di
s
t
r
i
b
ut
i
on
m
at
r
i
x
of
obs
er
v
at
i
o
ns
i
s
denot
ed
as
*
()
jk
N
M
Bb
=
,
w
h
er
e
()
jk
t
k
t
j
b
PO
q
θθ
=
=
=
.
T
her
ef
or
e,
H
MM
c
an
be
de
not
e
d
as
(
,
,,
,
)
N
M
A
B
λπ
=
.
T
her
e
ar
e
t
w
o
b
as
i
c
al
gor
i
t
hm
s
w
hen
us
i
ng
H
M
M:
B
a
um
-
W
el
c
h
al
gor
i
t
hm
and
V
i
t
er
b
i
al
g
or
i
t
hm
.
B
aum
-
W
el
c
h
al
gor
i
t
hm
i
s
us
ed
t
o
s
t
ud
y
t
he
ex
i
s
t
i
ng
o
bs
er
v
at
i
on
d
at
a
a
nd
t
r
ai
n
t
he
r
el
e
v
a
nt
H
MM
.
V
i
t
er
b
i
al
g
or
i
t
hm
i
s
us
ed
t
o
c
al
c
ul
a
t
e t
h
e m
os
t
pr
obabl
e s
eque
nc
e of
hi
dde
n s
t
at
e
s
and t
he pr
o
bab
i
l
i
t
y
gi
v
e
n an obs
er
v
at
i
on
s
equenc
e
bas
ed o
n t
he
t
r
ai
ne
d H
M
M.
I
n t
h
e pa
per
,
t
he c
l
as
s
i
f
i
c
at
i
on r
es
ul
t
s
of
ac
t
i
on
c
har
ac
t
er
i
s
t
i
c
s
ar
e as
t
he
o
bs
er
v
at
i
o
ns
,
Ba
u
m
-
W
el
c
h al
gor
i
t
hm
i
s
us
ed t
o t
r
ai
n t
h
e
H
MM of
eac
h
ac
t
i
on k
i
nd and
V
i
t
er
bi
a
l
go
r
i
t
hm
i
s
us
ed t
o c
al
c
ul
at
e t
h
e ac
t
i
on k
i
nd of
t
he
m
os
t
pr
obab
i
l
i
t
y
gi
v
e
n
an obs
er
v
at
i
on s
eq
ue
nc
e.
B
ec
aus
e t
her
e m
a
y
b
e a a
m
ount
of
s
i
m
i
l
ar
i
t
y
am
ong di
f
f
er
ent
ac
t
i
ons
and
t
he
m
ot
i
on
di
r
ec
t
i
ons
of
m
an
y
ac
t
i
ons
ar
e o
p
pos
i
t
e
s
uc
h
as
s
i
t
t
i
ng
a
n
d s
t
an
di
ng,
t
he
r
ec
ogni
t
i
o
n
r
es
u
l
t
b
as
ed
on
H
MM
an
d
Dy
(
t
)
ar
e
us
ed
t
o
j
udg
e
ac
t
i
on
k
i
nd
c
om
pr
e
hens
i
v
e
l
y
a
nd
t
he ac
t
i
on
r
ec
ogn
i
t
i
on m
et
h
od of
hum
an’
s
l
o
w
er
l
i
m
bs
i
s
s
how
n i
n F
i
g
ur
e
7.
F
ig
ur
e
7
.
T
he ac
t
i
on r
ec
o
gn
i
t
i
on
m
et
hod
of
hum
an’
s
l
o
w
er
l
i
m
bs
Butterworth filter
y
(
t
)
Wavelet transform
Action characteristics
Improved
self
-
organizing
competitive neural network
Baum
-
Welch
algorithm
HMM of each action
Train
Viterbi
algorithm
Recognize
Classification results
Training data
Test data
The most probable hidden state
sequence and the probability
The most probable action kind
Compare
D
y
(
t
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
14
,
N
o
.
3,
S
ept
em
ber
2016
:
11
92
–
1
202
1200
2.
3.
2
.
E
xp
er
i
m
en
t
s
S
om
e t
y
p
i
c
al
ac
t
i
ons
of
hum
an’
s
l
o
w
er
l
i
m
bs
ar
e c
hos
en
t
o
v
al
i
dat
e t
he
ac
t
i
on
r
ec
ogni
t
i
o
n
m
et
hod,
s
uc
h
as
w
al
k
i
ng,
r
u
nni
ng,
j
um
pi
n
g,
s
i
t
t
i
ng
,
s
t
and
i
n
g
,
c
l
i
m
bi
ng
up
an
d
c
l
i
m
bi
ng d
o
w
n.
T
he ac
t
i
o
n
pr
oc
es
s
es
ar
e s
ho
w
n
i
n F
i
g
ur
e
8.
F
ig
ur
e
8
.
A
c
t
i
o
n pr
oc
es
s
es
of
hum
an’
s
l
o
w
er
l
i
m
bs
T
he ex
per
i
m
ent
al
dat
a
i
s
l
i
s
t
ed i
n T
abl
e 1,
and f
i
g
ur
es
i
n br
ac
k
et
s
r
epr
es
ent
t
he n
u
m
b
e
r
of
ac
t
i
on
pr
oc
es
s
es
i
n t
he s
et
of
hum
an ac
t
i
o
n d
at
a.
T
abl
e 1.
E
x
per
i
m
ent
al
d
at
a
A
ct
i
o
n
T
r
ai
ni
ng dat
a
T
es
t
dat
a
W
al
k
02_01,
02_02,
05_01
06_01,
07_01,
07_02,
08_01,
08_
02,
16_12,
16_13,
16_14,
16_15
,
16_16,
16_17
Ru
n
16_35,
16_36,
16_37
02_03,
09_01,
09_02,
09_03,
35_
17,
35_18,
35_19,
35_20,
16_38
,
16_39,
16_40,
16_41,
16_42,
16_43
J
um
p
16_01,
16_02,
16_03
01_01(
2)
,
02_04,
13_11,
13_13,
1
3_39,
13_40,
13_41,
16_04,
16_0
5,
16_06,
16_07,
16_09
S
it
13_01(
2)
,
13_02
13_03(
3)
,
13_04(
4)
,
13_05(
2)
,
13_06(
3)
,
14_27,
14_28,
14_29(
3)
,
14
_3
0
(2
),
14_31(
2)
,
14_32(
3)
S
t
and
13_01(
2)
,
13_02
13_03(
2)
,
13_04(
4)
,
13_05(
2)
,
13_06(
3)
,
14_27,
14_28,
14_29(
2)
,
14
_30,
14_31(
2)
,
14_32(
2)
C
l
i
m
b up
01_02(
3)
01_03(
2)
,
01_04(
2)
,
01_05(
2)
,
01_06(
2)
,
01_07(
3)
,
13_35,
13_36,
13
_37,
13_38,
14_21,
14_22,
13_23
C
l
im
b
dow
n
01_02(
3)
01_03(
2)
,
01_04(
2)
,
01_05(
2)
,
01_06,
01_07,
13_35,
13_36,
13_37
,
13_38,
14_21,
14_22,
14_23
T
he H
MM t
r
ai
ni
ng
par
am
et
er
s
i
n
t
he
ac
t
i
o
n
r
ec
ogn
i
t
i
o
n
pr
oc
es
s
of
y’
(
t
)
ar
e
as
f
o
l
l
o
w
s
:
t
he num
ber
of
s
t
at
es
1
7
K
=
,
t
he
num
ber
of
obs
er
v
at
i
o
ns
2
10
K
=
,
a
nd t
he m
ax
i
m
u
m
num
ber
o
f
cy
cl
e
s
1
40
C
=
.
A
f
t
er
t
he H
MMs
ar
e t
r
ai
n
ed
w
i
t
h t
he t
r
a
i
ni
ng d
at
a,
t
es
t
d
at
a
i
s
r
ec
ogn
i
z
ed
bas
ed
on t
h
e pr
o
pos
ed
ac
t
i
on r
ec
ogn
i
t
i
on m
et
hod
and
t
he f
i
n
al
r
ec
og
ni
t
i
o
n r
es
u
l
t
s
ar
e
l
i
s
t
ed
i
n T
abl
e
2.
Walk
Run
Jump
Stand
Sit
Climb up
Climb down
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
IS
S
N
:
1
693
-
6
930
A
c
t
i
o
n R
ec
o
gn
i
t
i
on o
f
H
u
m
an
’
s
L
ow
er
Li
m
bs
B
as
ed o
n A
H
um
an J
o
i
nt
(
F
e
ng L
i
a
ng
)
1201
T
abl
e 2.
T
he f
i
na
l
r
ec
o
gni
t
i
on r
es
ul
t
s
W
al
k
Ru
n
J
um
p
S
it
S
t
and
C
l
i
m
b up
C
l
i
m
b dow
n
W
al
k
11
0
0
0
0
0
0
Ru
n
1
13
0
0
0
0
0
J
um
p
0
1
11
0
0
0
0
S
it
0
0
0
20
0
4
0
S
t
and
1
0
1
0
14
0
4
C
l
i
m
b up
0
0
0
1
0
17
0
C
l
i
m
b dow
n
0
0
0
0
1
0
14
3
.
D
i
s
cu
ssi
o
n
C
al
c
u
l
at
i
ons
i
n
t
h
e
ex
per
i
m
ent
w
er
e
per
f
or
m
ed
us
i
ng
a
c
om
put
er
w
i
t
h
a
gua
d
-
c
or
e I
nt
el
E
5 2
.
80G
H
z
C
P
U
.
T
he a
v
e
r
age c
al
c
u
l
at
i
o
n t
i
m
e of
r
ec
ogn
i
t
i
on pr
oc
es
s
es
of
C
MU
hum
an ac
t
i
on
dat
a
bas
e i
s
a
bou
t
0.
8
2s
.
C
om
par
ed w
i
t
h
ot
her
ac
t
i
on r
ec
ogn
i
t
i
on m
et
hods
,
t
he r
ec
ogn
i
t
i
on
r
es
ul
t
s
ar
e
l
i
s
t
e
d i
n T
abl
e
3.
T
abl
e 3.
T
he C
om
par
i
s
on o
f
R
ec
ogni
t
i
on
R
es
u
l
t
s
.
M
et
hod
A
c
c
ur
ac
y
D
y
nam
i
c
t
e
m
por
al
w
ar
pi
ng
[
2
]
0.
8421
T
r
aj
ec
t
or
y
P
r
oj
ec
t
i
on
[
7
]
0.
8684
Lear
ni
ng ac
t
i
on
e
n
s
em
bl
e
[1
6
]
0.
9035
P
r
opos
ed m
e
t
hod
0.
8772
I
t
c
an b
e f
ound
t
hat
t
he
pr
opos
e
d m
et
hod has
a
hi
g
h r
ec
ogn
i
t
i
on r
at
e
i
n
T
abl
e 3.
C
om
par
ed t
o
t
he ot
her
ac
t
i
on r
ec
og
ni
t
i
o
n m
et
hod
s
,
t
he pr
op
os
ed m
et
hod ne
ed
s
l
es
s
m
ot
i
on
i
nf
or
m
at
i
on;
i
t
do
es
not
ne
ed t
o a
dj
us
t
t
he m
ot
i
on t
r
aj
ec
t
or
i
es
of
m
ul
t
i
p
l
e h
um
an j
oi
nt
s
at
t
he
s
a
m
e
t
i
m
e
and
c
an
obt
ai
n
t
h
e
ac
t
i
on
c
har
ac
t
er
i
s
t
i
c
s
qui
c
k
l
y
bas
ed
on
w
a
v
e
l
et
t
r
a
ns
f
or
m
.
T
her
ef
or
e,
t
he pr
op
os
ed m
et
ho
d c
an
per
f
or
m
a f
as
t
c
al
c
ul
a
t
i
o
n s
pe
ed
.
4
.
C
o
n
c
l
u
s
i
o
n
T
o
r
eal
i
z
e
t
he
f
as
t
and
ac
c
ur
at
e
ac
t
i
on
r
ec
og
ni
t
i
o
n
of
hum
an’
s
l
o
w
er
l
i
m
bs
c
apt
u
r
ed
b
y
opt
i
c
a
l
m
ot
i
on
c
apt
ur
e
e
qu
i
pm
ent
,
a
nov
e
l
ac
t
i
o
n
r
ec
ogni
t
i
o
n
m
et
hod
of
hu
m
an’
s
l
o
w
er
l
i
m
bs
i
s
pr
opos
e
d.
A
f
t
er
f
i
l
t
er
ed,
t
h
e
y
c
oor
di
nat
es
of
hi
p
j
oi
nt
i
n
t
he
W
C
S
ar
e
us
ed
t
o c
al
c
ul
at
e ac
t
i
on
c
har
ac
t
er
i
s
t
i
c
s
a
nd t
he
y
ar
e c
l
as
s
i
f
i
ed
bas
e
d
on
an
i
m
pr
ov
ed s
el
f
-
or
gan
i
z
i
n
g n
eur
al
net
w
or
k
pr
opos
e
d i
n t
he pa
per
.
A
n
ac
t
i
on r
ec
o
gn
i
t
i
on m
et
hod bas
ed o
n H
MM
i
s
i
nt
r
o
duc
e
d t
o r
eal
i
z
e t
h
e
r
ec
o
gni
t
i
o
n
of
t
he
ac
t
i
o
ns
of
hu
m
an’
s
l
o
w
er
l
i
m
bs
i
n
c
onj
unc
t
i
o
n
w
i
t
h
t
he
c
han
g
e
di
r
ec
t
i
on
of
y
c
oor
di
n
at
es
.
T
he ac
t
i
on r
ec
ogn
i
t
i
on m
et
hod of
hu
m
an’
s
l
ow
er
l
i
m
bs
onl
y
ut
i
l
i
z
es
t
h
e
y
c
oor
di
n
at
es
of
hi
p
j
o
i
nt
t
o
c
al
c
ul
at
e
o
t
her
ac
t
i
on
i
nf
or
m
at
i
o
n;
i
t
has
a
f
as
t
c
al
c
ul
a
t
i
on
s
pee
d
an
d
c
an
s
at
i
s
f
y
t
he
ac
t
i
on
r
ec
o
gni
t
i
o
n
n
eed
s
of
di
f
f
er
ent
p
er
s
on
ne
l
.
E
x
per
i
m
ent
s
pr
o
v
e
t
he
m
et
hod
has
a g
ood
r
ec
ogn
i
t
i
on
ef
f
e
c
t
and
a go
od
app
l
i
c
at
i
o
n p
r
os
pec
t
.
A
c
k
n
o
w
l
e
d
g
e
m
e
n
ts
T
hi
s
w
or
k
w
as
s
uppor
t
e
d b
y
t
he
N
a
t
i
o
na
l
N
at
ur
a
l
S
c
i
enc
e
F
o
un
dat
i
on of
C
hi
n
a
(
G
r
ant
N
o:
4
117
416
2
).
R
ef
er
en
ces
[1
]
C
hen
W
Q
, X
i
ao
G
Q
, L
in
X,
et
al
.
O
n a
hum
an be
hav
i
or
s
c
l
a
s
s
i
f
i
c
at
i
o
n m
o
del
bas
ed o
n at
t
r
i
b
ut
e
-
B
ay
es
i
an
net
w
or
k
.
J
our
nal
of
S
out
hw
e
s
t
U
ni
v
er
s
i
t
y
.
2
014
;
3
9
(3
):
7
-
11.
[2
]
B
ar
nac
hon M
,
B
ouak
az
S
,
B
o
uf
am
a B
,
e
t
al
.
O
ngoi
ng h
um
a
n ac
t
i
on r
e
c
og
ni
t
i
on w
i
t
h m
o
t
i
o
n c
apt
ur
e
.
P
at
t
er
n R
e
c
o
gni
t
i
on
.
201
4
;
4
7:
238
-
2
47.
[3
]
K
am
al
S,
J
al
al
A
.
A
hy
br
i
d
f
eat
ur
e
ex
t
r
ac
t
i
on
appr
o
ac
h
f
o
r
hum
a
n
de
t
ec
t
i
on
,
t
r
a
c
k
i
ng
a
nd
ac
t
i
v
i
t
y
r
ec
og
ni
t
i
on
us
i
ng d
ept
h
s
e
ns
or
s
.
A
r
ab
i
an J
o
ur
nal
f
or
S
c
i
e
nc
e
a
nd E
n
gi
ne
er
i
n
g
.
2
015
:
1
-
9.
[4
]
H
u
Y,
Z
heng
W
.
H
um
an
ac
t
i
on
r
ec
og
ni
t
i
on
ba
s
ed
on
k
ey
f
r
am
es
.
P
r
oc
ee
di
ng
s
of
C
on
f
er
en
c
e
o
n
C
om
put
er
S
c
i
en
c
e a
nd E
d
uc
a
t
i
on
.
Q
i
ngda
o
.
2
011
:
535
-
542.
Evaluation Warning : The document was created with Spire.PDF for Python.