I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
E
lec
t
r
ical
an
d
Com
p
u
t
e
r
E
n
gin
e
e
r
in
g
(
I
JE
CE
)
Vol.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
,
pp.
455
~
466
I
S
S
N:
2088
-
8708
,
DO
I
:
10
.
11591/i
jec
e
.
v
15
i
1
.
pp
4
55
-
466
455
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
e
c
e
.
iaes
c
or
e
.
c
om
Hu
m
an
m
ot
io
n
c
la
ssi
f
ic
at
io
n
b
y m
ic
r
o
-
d
o
p
p
le
r
r
ad
a
r
u
si
n
g
in
t
e
ll
ig
e
n
t
al
gor
it
h
m
s
And
r
e
s
F
e
li
p
e
Ar
ias
B
all
e
n
,
E
d
it
h
P
aola
E
s
t
u
p
i
ñ
án
Cues
t
a,
Ju
an
Car
los
M
ar
t
ín
e
z
Qu
in
t
e
r
o
T
e
le
c
omm
uni
c
a
ti
on
s
E
ngi
ne
e
r
in
g P
r
ogr
a
m, Unive
r
s
id
a
d M
il
it
a
r
N
ue
va
G
r
a
na
da
, B
ogot
a
, C
ol
ombi
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
J
un
12,
2024
R
e
vis
e
d
S
e
p
11,
2024
Ac
c
e
pted
Oc
t
1,
2024
T
h
i
s
ar
t
i
c
l
e
i
n
t
r
o
d
u
ce
s
a
t
ec
h
n
i
q
u
e
f
o
r
d
e
t
e
c
t
i
n
g
f
o
u
r
h
u
m
a
n
m
o
v
e
me
n
t
s
u
s
i
n
g
m
i
c
r
o
-
d
o
p
p
l
er
ra
d
ar
a
n
d
i
n
t
e
l
l
i
g
e
n
t
a
l
g
o
r
i
t
h
m
s
.
M
i
cr
o
-
d
o
p
p
l
e
r
ra
d
ar
e
x
h
i
b
i
t
s
t
h
e
c
a
p
a
b
i
l
i
t
y
t
o
d
e
t
e
c
t
a
n
d
m
ea
s
u
re
o
b
j
e
c
t
m
o
v
e
me
n
t
s
w
i
t
h
i
n
t
r
i
ca
t
e
d
e
t
a
i
l
,
e
v
e
n
c
a
p
t
u
r
i
n
g
c
o
m
p
l
e
x
o
r
n
o
n
-
r
i
g
i
d
m
o
t
i
o
n
s
,
w
h
i
l
e
ac
c
u
ra
t
e
l
y
i
d
e
n
t
i
f
y
i
n
g
d
i
r
e
c
t
i
o
n
,
v
e
l
o
c
i
t
y
,
a
n
d
m
o
t
i
o
n
p
a
t
t
e
r
n
s
.
T
h
e
a
p
p
l
i
c
a
t
i
o
n
o
f
i
n
t
e
l
l
i
g
e
n
t
a
l
g
o
r
i
t
h
m
s
e
n
h
a
n
c
es
d
e
t
e
c
t
i
o
n
e
ff
i
c
i
e
n
c
y
a
n
d
re
d
u
ce
s
fa
l
s
e
a
l
ar
m
s
b
y
d
i
s
c
e
r
n
i
n
g
s
u
b
t
l
e
m
o
v
em
e
n
t
p
a
t
t
er
n
s
,
t
h
e
re
b
y
f
a
c
i
l
i
t
a
t
i
n
g
m
o
r
e
a
c
c
u
ra
t
e
d
e
t
e
c
t
i
o
n
a
n
d
a
d
ee
p
er
u
n
d
e
r
s
t
a
n
d
i
n
g
o
f
o
b
s
er
v
e
d
o
b
j
e
c
t
d
y
n
am
i
c
s
.
A
c
o
n
t
i
n
u
o
u
s
w
a
v
e
ra
d
ar
s
e
t
u
p
w
a
s
i
m
p
l
em
e
n
t
e
d
u
t
i
l
i
z
i
n
g
a
s
p
ec
t
r
u
m
a
n
a
l
y
z
er
a
n
d
ra
d
i
o
f
re
q
u
e
n
c
y
(R
F)
g
e
n
er
a
t
o
r
ca
p
t
u
r
i
n
g
s
i
g
n
a
l
s
i
n
a
s
p
e
c
t
r
o
g
ra
m
c
e
n
t
er
e
d
a
t
2
,
3
9
5
M
H
z.
S
i
x
m
o
d
e
l
s
w
e
re
a
s
s
es
s
e
d
f
o
r
i
m
a
g
e
c
l
a
s
s
i
f
i
ca
t
i
o
n
:
VGG
-
1
6
,
V
G
G
-
1
9
,
M
o
b
i
l
e
N
e
t
,
M
o
b
i
l
e
N
e
t
V
2
,
X
c
e
p
t
i
o
n
,
a
n
d
I
n
c
e
p
t
i
o
n
V
3
.
A
d
a
t
a
s
e
t
c
o
m
p
r
i
s
i
n
g
5
0
0
i
m
a
g
e
s
d
e
p
i
c
t
i
n
g
f
o
u
r
m
o
v
em
en
t
s
-
r
u
n
n
i
n
g
,
w
a
l
k
i
n
g
,
a
r
m
ra
i
s
i
n
g
,
a
n
d
j
u
m
p
i
n
g
-
w
a
s
c
u
ra
t
e
d
.
O
u
r
f
i
n
d
i
n
g
s
re
v
e
a
l
t
h
a
t
t
h
e
m
o
s
t
o
p
t
i
m
a
l
ar
c
h
i
t
e
c
t
u
r
e
i
n
t
e
rm
s
o
f
t
ra
i
n
i
n
g
t
i
me,
a
cc
u
ra
c
y
,
a
n
d
l
o
s
s
i
s
V
G
G
-
1
6
,
a
c
h
i
e
v
i
n
g
a
n
ac
c
u
r
ac
y
o
f
9
6
%
.
F
u
r
t
h
e
rm
o
re,
p
r
ec
i
s
i
o
n
v
a
l
u
e
s
o
f
9
6
%
,
1
0
0
%
,
a
n
d
9
8
%
w
er
e
o
b
t
a
i
n
e
d
f
o
r
t
h
e
m
o
v
e
me
n
t
s
o
f
w
a
l
k
i
n
g
,
r
u
n
n
i
n
g
,
a
n
d
a
rm
r
a
i
s
i
n
g
,
r
e
s
p
e
c
t
i
v
e
l
y
.
N
o
t
a
b
l
y
,
V
G
G
-
1
6
e
x
h
i
b
i
t
e
d
a
t
r
a
i
n
i
n
g
l
o
s
s
o
f
4
.
1
9
1
E
-
0
4
,
a
t
t
r
i
b
u
t
e
d
t
o
t
h
e
u
t
i
l
i
za
t
i
o
n
o
f
t
h
e
A
d
a
m
o
p
t
i
m
i
z
er
w
i
t
h
a
l
e
a
r
n
i
n
g
ra
t
e
o
f
0
.
0
0
1
o
v
e
r
1
5
e
p
o
c
h
s
a
n
d
a
b
a
t
c
h
s
i
z
e
o
f
3
2
.
K
e
y
w
o
r
d
s
:
C
onvolut
ional
ne
ur
a
l
ne
twor
ks
M
icr
o
-
doppler
r
a
da
r
M
oti
on
de
tec
ti
on
VGG
-
16
VGG
-
19
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
:
E
dit
h
P
a
ola
E
s
tupi
ñá
n
C
ue
s
ta
T
e
lec
omm
unica
ti
ons
E
nginee
r
ing
P
r
ogr
a
m
,
Unive
r
s
idad
M
il
it
a
r
Nue
va
Gr
a
na
da
S
e
de
C
a
ll
e
100,
B
ogotá,
C
olom
bia
E
mail:
e
dit
h
.
e
s
tupi
na
n
@unim
il
it
a
r
.
e
du
.
c
o
1.
I
NT
RODU
C
T
I
ON
T
he
de
tec
ti
on
of
human
mot
ion
ha
s
be
c
ome
incr
e
a
s
ingl
y
r
e
leva
nt
ove
r
ti
me
,
e
s
pe
c
ially
i
n
s
ur
ve
il
lanc
e
mec
ha
nis
ms
whe
r
e
the
identif
ica
ti
on
of
s
us
picious
a
c
ti
vit
ies
,
int
r
us
ions
,
o
r
a
bno
r
mal
b
e
ha
vior
s
c
a
n
pr
e
ve
nt
c
r
i
mi
na
l
a
c
ts
s
uc
h
a
s
thef
t
a
nd
pr
ov
ide
a
r
a
pid
r
e
s
pons
e
.
I
n
the
medic
a
l
f
ield
,
moni
tor
ing
a
pa
ti
e
nt's
br
e
a
thi
ng
a
nd
de
tec
ti
ng
a
bnor
mal
body
moveme
nts
c
a
n
f
a
c
il
it
a
te
im
media
te
a
tt
e
nti
on
f
r
o
m
medic
a
l
pe
r
s
ona
l.
I
n
the
r
e
a
lm
of
tr
a
f
f
ic
c
ontr
o
l,
de
tec
ti
ng
ve
hicula
r
moveme
nt
a
nd
c
las
s
if
ying
the
nu
mber
of
indi
viduals
c
a
n
r
e
gulate
tr
a
f
f
ic
f
low
a
nd
e
nha
nc
e
e
f
f
icie
nc
y.
M
icr
o
-
doppler
r
a
da
r
,
in
a
ddit
io
n
to
it
s
a
ppli
c
a
ti
ons
in
de
tec
ti
on
a
nd
s
e
c
ur
it
y,
unf
olds
it
s
potential
in
human
-
c
omput
e
r
int
e
r
a
c
ti
on.
I
n
e
nvi
r
onments
li
ke
vir
tual
r
e
a
li
ty
a
nd
mot
ion
c
ontr
ol
s
ys
tems
,
t
his
r
a
da
r
c
a
ptur
e
s
us
e
r
s
'
ge
s
tur
e
s
a
nd
s
ubtl
e
mo
ve
ments
,
tr
a
ns
lating
them
int
o
digi
tal
c
omm
a
nds
thr
ough
in
telli
ge
nt
a
lgor
it
hms
.
T
his
a
r
ti
c
le
int
r
oduc
e
s
a
tec
hnique
f
o
r
de
tec
ti
ng
f
our
d
is
ti
nc
t
human
moveme
nts
us
ing
mi
c
r
o
-
doppler
r
a
da
r
a
nd
int
e
ll
igent
a
lgor
it
hms
.
M
icr
o
-
doppler
r
a
da
r
,
in
a
ddit
ion
to
it
s
a
ppl
ica
ti
ons
in
de
tec
ti
on
a
nd
s
e
c
ur
it
y,
r
e
ve
a
ls
it
s
potential
in
human
-
c
omput
e
r
int
e
r
a
c
ti
on.
I
n
e
nvir
on
ments
s
uc
h
a
s
vi
r
tual
r
e
a
li
ty
a
nd
mo
ti
on
c
ontr
ol
s
ys
tems
,
t
his
r
a
da
r
c
a
ptur
e
s
us
e
r
s
'
ge
s
tur
e
s
a
nd
s
ubtl
e
moveme
nts
,
tr
a
ns
lating
them
in
to
digi
tal
c
omm
a
nds
th
r
ough
in
telli
ge
nt
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
455
-
466
456
a
lgor
it
hms
.
T
h
is
e
nha
nc
e
s
the
us
e
r
e
xpe
r
ienc
e
a
nd
e
na
bles
na
tur
a
l
a
nd
im
mer
s
ive
ge
s
tur
e
-
ba
s
e
d
in
ter
f
a
c
e
s
.
T
his
im
pr
ove
s
the
us
e
r
e
xpe
r
ienc
e
a
nd
e
na
bles
na
t
ur
a
l
a
nd
im
mer
s
ive
ge
s
tur
e
-
ba
s
e
d
int
e
r
f
a
c
e
s
.
S
tudi
e
s
ha
ve
a
ddr
e
s
s
e
d
thi
s
topi
c
,
pr
opos
ing
va
r
ious
s
c
e
na
r
ios
a
nd
tec
hniques
f
or
de
tec
ti
ng
human
moveme
nt
[
1]
–
[
8]
.
S
e
ve
r
a
l
r
e
s
e
a
r
c
h
e
nde
a
vor
s
s
t
a
nd
out
in
thi
s
f
ield.
I
n
s
tudy
[
3]
,
the
powe
r
of
m
il
li
mete
r
wa
ve
(
mm
W
a
ve
)
r
a
da
r
f
or
p
r
e
c
is
e
mot
ion
da
ta
c
a
ptur
e
is
c
ombi
ne
d
with
de
e
p
im
a
ge
pr
oc
e
s
s
ing
us
ing
c
onvolut
ion
a
l
ne
ur
a
l
ne
twor
ks
(
C
NN
s
)
.
T
he
pr
opos
e
d
a
ppr
oa
c
h
invol
v
e
s
mm
W
a
ve
da
ta
c
oll
e
c
ti
on,
p
r
e
pr
oc
e
s
s
ing,
a
nd
tr
a
ini
ng
of
a
C
NN
de
s
igned
to
r
e
c
ognize
s
pe
c
if
ic
be
ha
vior
p
a
tt
e
r
ns
.
E
va
luate
d
r
e
s
ult
s
de
mons
tr
a
te
the
viabi
li
t
y
of
thi
s
methodology
f
or
r
e
a
l
-
ti
me
de
tec
ti
on
a
nd
c
las
s
if
ica
ti
on,
with
im
pl
ica
ti
ons
f
or
s
ur
ve
il
lanc
e
,
s
e
c
ur
it
y,
a
nd
r
e
late
d
f
ields
in
human
be
ha
vior
moni
to
r
ing.
I
n
s
t
udy
[
6]
,
a
n
ul
tr
a
-
wide
ba
nd
r
a
da
r
tec
hnique
is
e
mp
loyed
to
de
tec
t
human
moveme
nt
thr
ough
wa
ll
s
,
inco
r
po
r
a
ti
ng
a
c
onvolut
ional
ne
ur
a
l
ne
twor
k
(
C
NN
)
f
or
im
a
ge
pr
oc
e
s
s
ing
a
nd
objec
t
c
las
s
if
ica
ti
on.
S
pe
c
tr
ogr
a
ms
take
n
thr
ough
wa
ll
s
a
c
hieve
high
a
c
c
ur
a
c
y
whe
n
a
s
ubjec
t
wa
lks
be
hind
them.
I
n
s
tudy
[
7
]
,
a
n
a
lgor
it
hm
is
de
ve
loped
f
or
the
r
e
c
ognit
ion
o
f
Ame
r
ica
n
s
ign
l
a
ngua
ge
(
ASL
)
s
ignals
,
a
idi
ng
indi
viduals
with
s
pe
e
c
h
dif
f
iculti
e
s
,
a
c
hieving
a
r
e
c
ognit
ion
r
a
te
of
72
.
5%
f
o
r
20
s
ignals
.
I
n
s
tudy
[
8
]
,
the
de
tec
ti
on
of
dr
one
s
u
s
ing
s
of
twa
r
e
de
f
ined
r
a
dio
(
s
dr
)
r
a
da
r
a
nd
r
a
dio
f
r
e
que
nc
y
s
ignals
is
pr
e
s
e
nted,
highl
ight
ing
the
diver
s
e
a
ppli
c
a
ti
ons
of
mi
c
r
o
-
doppler
r
a
da
r
in
mo
ti
on
de
tec
ti
o
n.
T
ivi
ve
e
t
al.
[
9
]
p
r
opos
e
s
a
methodology
f
or
c
a
ptur
in
g
s
ubtl
e
f
e
a
tur
e
s
o
f
hu
man
ga
it
thr
ough
mi
c
r
o
-
doppler
de
tec
ti
on,
whic
h
a
r
e
s
mall
f
luctua
ti
ons
in
r
a
da
r
r
e
tur
n
f
r
e
que
nc
y
c
a
us
e
d
by
body
pa
r
t
moveme
nts
dur
ing
wa
lki
ng.
Us
ing
s
ignal
p
r
oc
e
s
s
ing,
the
a
uthor
s
e
xt
r
a
c
t
r
e
leva
nt
in
f
or
mation
f
r
om
doppler
s
pe
c
tr
ogr
a
ms
a
nd
de
ve
lop
a
c
las
s
if
ica
ti
on
method
that
identif
ies
dis
ti
nc
ti
ve
wa
lki
ng
pa
t
ter
ns
a
mong
dif
f
e
r
e
nt
indi
vidu
a
ls
.
T
his
a
ppr
oa
c
h
de
mons
tr
a
tes
the
f
e
a
s
ibi
li
ty
o
f
us
ing
r
a
da
r
mi
c
r
o
-
doppler
inf
o
r
mation
to
dif
f
e
r
e
nti
a
te
a
n
d
c
las
s
if
y
ga
it
pa
tt
e
r
ns
,
with
potential
a
ppli
c
a
ti
ons
in
bi
ometr
ics
a
nd
he
a
lt
h
moni
to
r
ing.
T
he
s
e
r
e
s
e
a
r
c
h
e
f
f
o
r
ts
unde
r
s
c
or
e
the
us
e
of
mi
c
r
o
-
doppler
r
a
da
r
a
s
a
pr
omi
s
ing
tec
hnique
f
or
mot
ion
de
tec
ti
on,
c
a
pa
ble
of
de
tec
ti
ng
unique
f
e
a
tur
e
s
of
a
movi
ng
objec
t,
s
uc
h
a
s
c
ha
r
a
c
ter
is
ti
c
human
moveme
nt
pa
tt
e
r
ns
li
ke
wa
lki
ng
a
nd
r
unning.
Unlike
other
s
e
ns
or
s
,
s
uc
h
a
s
video
c
a
mer
a
s
,
mi
c
r
o
-
doppler
r
a
da
r
is
mi
nim
a
ll
y
a
f
f
e
c
ted
by
e
nvir
onmenta
l
c
ondit
ions
s
uc
h
a
s
li
ghti
ng,
r
a
in,
or
f
og
,
pr
ovidi
ng
r
e
li
a
bil
i
ty
in
im
-
a
ge
pr
oc
e
s
s
ing
[
10]
.
I
n
s
tudy
[
11]
,
the
t
r
a
ini
ng
of
a
n
int
e
ll
igent
a
lgor
it
hm
f
or
the
c
las
s
if
ica
ti
on
of
e
a
c
h
moveme
nt
is
p
r
e
s
e
nted.
F
r
o
m
the
pr
ojec
t,
it
c
a
n
be
c
onc
luded
that
the
pr
opos
e
d
model
a
c
hieve
s
a
n
a
c
c
ur
a
c
y
of
92
.
65%
f
o
r
human
mot
ion
de
tec
ti
on
us
ing
mi
c
r
o
-
doppler
r
a
da
r
a
nd
int
e
ll
ige
nt
a
lgor
it
hms
f
e
a
tur
ing
long
s
hor
t
-
ter
m
memor
y
(
L
S
T
M
)
with
36
c
e
ll
s
a
nd
82
.
33%
f
o
r
de
e
p
c
onvolut
ional
ne
ur
a
l
ne
twor
ks
(
DC
NN
)
,
s
ur
pa
s
s
ing
pr
e
vious
ly
s
tudi
e
d
e
xis
ti
ng
methods
.
P
r
e
vious
r
e
s
e
a
r
c
h,
s
uc
h
a
s
r
e
f
e
r
e
nc
e
s
[
3]
,
[
9
]
,
[
10
]
f
oc
us
on
s
pe
c
if
ic
a
ppli
c
a
ti
ons
or
e
mpl
oy
pa
r
ti
c
ular
a
ppr
oa
c
he
s
f
or
mot
ion
de
tec
ti
on
(
e
.
g
.
,
C
NN
s
in
[
3]
a
nd
s
pe
c
if
ic
ga
it
c
las
s
if
ica
ti
on
tec
h
niques
in
[
9]
)
,
thi
s
a
r
ti
c
le
s
tands
out
by
e
xha
us
ti
ve
ly
e
va
l
ua
ti
ng
a
nd
c
ompar
ing
s
ix
dif
f
e
r
e
nt
de
e
p
lea
r
nin
g
models
(
vis
ua
l
ge
ometr
y
g
r
oup
-
16
(
VG
G
-
16)
,
VG
G
-
19,
M
obil
e
Ne
t,
M
obil
e
Ne
t
V2,
Xc
e
pti
on,
a
nd
I
nc
e
pti
on
V3)
.
T
his
pr
ovides
a
br
oa
de
r
pe
r
s
pe
c
ti
ve
on
whic
h
m
ode
ls
a
r
e
mos
t
e
f
f
e
c
ti
ve
f
or
de
tec
ti
ng
human
mo
ve
ments
us
ing
mi
c
r
o
-
doppler
r
a
da
r
.
T
his
r
e
s
e
a
r
c
h
dis
ti
ngui
s
he
s
it
s
e
lf
by
f
oc
us
ing
on
the
c
a
pa
bil
it
y
of
mi
c
r
o
-
doppler
r
a
da
r
to
de
tec
t
not
only
c
omm
on
moveme
nts
s
uc
h
a
s
wa
lki
ng
a
nd
r
unning
but
a
ls
o
mo
r
e
s
ubtl
e
a
nd
c
ompl
e
x
moveme
nts
li
ke
jum
ping
or
a
r
m
-
r
a
is
ing.
T
his
c
ontr
a
s
ts
with
wor
ks
li
ke
[
9]
,
whic
h
f
oc
us
on
ga
it
c
las
s
if
ica
ti
on,
a
nd
s
igni
f
ica
ntl
y
e
xpa
nds
the
s
c
op
e
of
the
s
tudy.
T
h
is
r
e
s
e
a
r
c
h
e
mpl
oys
s
ix
de
e
p
lea
r
ning
models
f
or
hu
man
mot
ion
de
tec
ti
on
us
ing
mi
c
r
o
-
doppler
r
a
da
r
.
T
he
models
uti
li
z
e
d
include
V
GG
-
16,
VGG
-
19,
M
obil
e
Ne
t,
M
obil
e
Ne
t
V2,
Xc
e
pti
on,
a
nd
I
nc
e
pti
on
V3.
Ke
y
met
r
ics
s
uc
h
a
s
a
c
c
ur
a
c
y,
tr
a
ini
ng
ti
me,
los
s
,
a
nd
the
c
onf
us
ion
matr
ix
a
r
e
a
s
s
e
s
s
e
d.
A
da
tas
e
t
c
ompr
is
ing
500
s
a
mpl
e
s
na
med
W
AR
J
M
AX
W
E
L
L
of
the
im
pleme
nted
s
c
e
n
a
r
io
wa
s
c
r
e
a
ted,
with
de
f
ined
mot
ion
c
a
tegor
ies
:
wa
lki
ng,
r
unning,
jum
ping,
a
nd
a
r
m
r
a
is
ing.
De
s
pit
e
a
dva
nc
e
s
in
human
mot
ion
de
tec
ti
on
,
a
c
c
ur
a
tely
dis
ti
nguis
hing
s
ubtl
e
a
nd
c
ompl
e
x
moveme
nts
,
e
s
pe
c
ially
in
c
ha
ll
e
nging
e
nvir
onment
s
,
r
e
mains
a
s
igni
f
ica
nt
c
ha
ll
e
nge
.
E
xis
ti
ng
a
pp
r
oa
c
he
s
a
r
e
of
ten
a
f
f
e
c
ted
by
e
nvi
r
onmenta
l
c
ondit
ions
or
e
xhibi
t
a
high
r
a
te
of
f
a
ls
e
pos
it
ives
,
li
mi
ti
ng
their
e
f
f
e
c
ti
ve
ne
s
s
in
c
r
it
ica
l
a
ppli
c
a
ti
ons
s
uc
h
a
s
s
ur
ve
il
lanc
e
a
nd
medic
ine.
M
icr
o
-
doppler
r
a
da
r
e
xh
ibi
ts
the
e
xc
e
pti
ona
l
a
bil
it
y
to
c
a
ptur
e
a
nd
mea
s
ur
e
pr
e
c
is
e
de
tails
of
human
moveme
nts
,
e
ve
n
thos
e
with
c
o
mpl
e
x
or
non
-
r
igi
d
pa
tt
e
r
ns
.
I
ts
int
e
gr
a
ti
on
with
int
e
ll
igent
a
lgor
it
hms
e
nha
nc
e
s
de
tec
ti
on
e
f
f
icie
nc
y
a
nd
r
e
du
c
e
s
f
a
ls
e
pos
it
ives
.
F
or
ins
tanc
e
,
the
r
a
da
r
's
c
a
pa
bil
it
y
to
d
is
ti
nguis
h
be
twe
e
n
the
s
ubtl
e
moveme
nts
of
wa
lki
ng
a
nd
r
unning.
T
h
is
mul
ti
dis
c
ipl
inar
y
a
ppr
oa
c
h
e
na
bles
mor
e
pr
e
c
is
e
de
tec
ti
on
a
nd
a
be
tt
e
r
unde
r
s
tanding
of
human
moveme
nt
pa
tt
e
r
ns
,
with
s
igni
f
ica
nt
im
p
li
c
a
ti
ons
in
a
ppli
c
a
ti
ons
s
uc
h
a
s
s
ur
ve
il
lanc
e
a
nd
medic
i
ne
.
T
he
c
ompar
is
on
of
ke
y
metr
ics
s
uc
h
a
s
pr
e
c
is
ion,
r
e
c
a
ll
,
F
1
s
c
or
e
,
a
nd
tr
a
ini
ng
a
nd
va
li
da
ti
on
los
s
,
a
thor
ough
a
s
s
e
s
s
ment
of
e
a
c
h
model'
s
pe
r
f
or
manc
e
in
s
pe
c
tr
ogr
a
m
c
las
s
if
ica
ti
on
is
pr
ovided.
P
a
r
ti
c
ular
ly
note
wor
thy
is
the
s
upe
r
ior
pe
r
f
or
manc
e
of
the
VG
G
-
16
model,
pos
it
ioni
ng
it
a
s
a
s
tandout
tool
f
or
a
c
c
ur
a
te
s
pe
c
tr
ogr
a
m
c
las
s
if
ica
ti
on
in
thi
s
c
ontext.
F
ur
ther
mor
e
,
the
a
r
ti
c
le
dis
ti
nguis
he
s
it
s
va
lue
by
r
e
c
ognizing
the
f
e
a
s
ibi
li
ty
o
f
r
a
da
r
tec
hniques
f
o
r
de
tec
ti
ng
s
low
-
s
pe
e
d
moveme
nts
,
e
s
pe
c
ially
whe
n
c
ombi
ne
d
with
int
e
ll
igent
a
lgo
r
it
hms
a
nd
a
dva
nc
e
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
Hum
an
motion
c
las
s
if
ication
by
mic
r
o
-
doppler
r
adar
us
ing
…
(
A
ndr
e
s
F
e
li
pe
A
r
ias
B
all
e
n
)
457
tec
hnologi
e
s
.
T
his
e
xplor
a
ti
on
of
potential
r
a
da
r
a
ppli
c
a
ti
ons
a
dds
a
s
igni
f
ica
nt
dim
e
ns
ion
to
th
e
s
tudy's
c
onc
lus
ions
,
br
oa
de
ning
it
s
s
c
ope
be
yond
the
e
va
luation
of
de
e
p
lea
r
ning
models
.
I
n
s
umm
a
r
y,
the
c
ombi
na
ti
on
of
a
thor
ough
e
va
luation
of
de
e
p
lea
r
ning
models
with
innovative
r
e
s
e
a
r
c
h
on
r
a
da
r
a
pp
li
c
a
ti
ons
in
mot
ion
de
tec
ti
on
e
s
tablis
he
s
thi
s
a
r
ti
c
le
a
s
a
va
l
ua
ble
a
nd
dis
ti
nc
ti
ve
c
ontr
ibut
ion
to
the
f
ield.
F
in
a
ll
y,
the
a
r
ti
c
le
is
or
ga
nize
d
int
o
s
e
c
ti
ons
.
S
e
c
ti
on
1
pr
ovid
e
s
a
r
e
view
of
the
s
tate
of
the
a
r
t
.
S
e
c
ti
on
2
int
r
od
uc
e
s
the
pr
opos
e
d
s
e
tup,
e
quipm
e
nt
us
e
d,
the
methodolog
y
a
ppli
e
d,
a
nd
s
ignal
p
r
oc
e
s
s
ing.
S
e
c
ti
on
3
pr
e
s
e
nts
the
r
e
s
ult
s
of
tr
a
ini
ng
int
e
ll
igent
a
lgor
it
h
ms
us
ing
a
d
a
tas
e
t
c
ons
is
ti
ng
of
125
im
a
ge
s
f
o
r
e
a
c
h
mot
ion
c
a
tegor
y,
f
oll
owe
d
by
the
c
onc
lus
ions
dr
a
wn.
2.
M
E
T
HO
D
T
his
r
e
s
e
a
r
c
h
e
mpl
oye
d
a
pha
s
e
d
methodology
,
a
s
il
lus
tr
a
ted
in
F
igur
e
1.
P
ha
s
e
A
identi
f
y
the
pr
im
a
r
y
ha
r
dwa
r
e
a
nd
s
of
twa
r
e
r
e
qui
r
e
ments
to
de
f
ine
the
s
c
e
na
r
io
a
nd
c
onduc
t
mot
ion
de
tec
ti
on
a
lo
ng
with
it
s
s
pe
c
if
ic
c
ha
r
a
c
ter
is
ti
c
s
,
including
dis
tanc
e
,
mo
ve
ments
,
a
nd
ha
r
dwa
r
e
c
onf
igu
r
a
ti
on
pa
r
a
mete
r
s
.
P
ha
s
e
B
de
f
ine
s
ignal
a
c
quis
it
ion
a
nd
pr
oc
e
s
s
ing
f
or
da
tas
e
t
c
r
e
a
ti
on,
uti
li
z
ing
Anr
it
s
u's
mas
ter
tool
s
s
of
twa
r
e
f
or
a
c
quir
ing
s
pe
c
tr
ogr
a
ms
a
nd
mea
s
ur
e
ment
da
ta
f
r
o
m
the
s
pe
c
tr
um
a
na
lyze
r
.
F
inally
,
in
P
ha
s
e
C
is
d
e
s
igned
a
nd
tr
a
ined
the
6
s
e
lec
ted
c
las
s
if
ica
ti
on
models
us
i
ng
metr
ics
s
uc
h
a
s
a
c
c
ur
a
c
y,
t
r
a
ini
ng
ti
me,
los
s
,
r
e
c
a
ll
,
a
nd
F
1
s
c
or
e
.
T
a
ble
1
s
how
the
e
quipm
e
nt
us
e
d
in
thi
s
r
e
s
e
a
r
c
h
a
nd
the
pr
opos
e
d
s
c
e
na
r
io.
F
igur
e
1.
R
e
s
e
a
r
c
h
methodology
wor
kf
low
T
he
s
e
lec
ted
metr
ics
a
ll
ow
f
or
the
mea
s
ur
e
ment
of
a
c
c
ur
a
c
y
in
c
las
s
if
ica
ti
on
models
,
with
a
f
oc
us
on
mi
nim
izing
f
a
ls
e
pos
it
ives
-
c
a
s
e
s
whe
r
e
the
model
incor
r
e
c
tl
y
p
r
e
dicts
a
pos
it
ive
outcome
whe
n
it
is
a
c
tually
ne
ga
ti
ve
.
T
he
a
c
c
ur
a
c
y
metr
ic
is
de
f
ined
a
s
a
pa
r
a
mete
r
that
e
va
luate
s
the
model'
s
a
bil
it
y
to
c
or
r
e
c
tl
y
c
las
s
if
y
s
a
mpl
e
s
int
o
the
de
s
ir
e
d
c
a
tegor
ies
.
I
t
is
c
a
lcula
ted
by
divi
ding
the
number
of
c
or
r
e
c
t
pr
e
dic
ti
ons
by
the
tot
a
l
number
of
pr
e
dictions
a
nd
is
typi
c
a
ll
y
e
xpr
e
s
s
e
d
a
s
a
va
lue
be
twe
e
n
0
a
nd
1,
whe
r
e
1
s
igni
f
ies
pe
r
f
e
c
t
a
c
c
ur
a
c
y
a
nd
0
indi
c
a
tes
no
a
c
c
ur
a
c
y.
T
r
a
ini
ng
ti
me
r
e
f
e
r
s
to
the
pe
r
iod
r
e
qui
r
e
d
f
or
a
mac
hine
lea
r
ning
model
to
pr
oc
e
s
s
a
nd
a
na
lyze
a
t
r
a
ini
ng
d
a
tas
e
t
in
or
de
r
to
lea
r
n
pa
tt
e
r
ns
a
nd
r
e
lations
hips
.
T
he
los
s
pa
r
a
mete
r
s
e
r
ve
s
a
s
a
n
indi
c
a
tor
of
the
model's
lea
r
ning
leve
l
dur
ing
tr
a
ini
ng
,
with
the
goa
l
of
m
ini
mi
z
ing
it
by
a
djus
ti
ng
model
pa
r
a
mete
r
s
[
12
]
.
T
he
r
e
c
a
ll
m
e
tr
ic
e
va
luate
s
the
model's
a
bil
it
y
to
c
or
r
e
c
tl
y
identif
y
a
ll
pos
it
ive
e
xa
mpl
e
s
in
the
da
tas
e
t,
f
oc
us
ing
on
mi
n
im
izing
f
a
ls
e
ne
ga
ti
ve
s
.
T
he
F
1
s
c
or
e
c
ombi
ne
s
p
r
e
c
is
ion
a
nd
r
e
c
a
ll
metr
ics
int
o
a
s
ingl
e
va
lue
to
a
s
s
e
s
s
the
pe
r
f
or
manc
e
of
a
c
las
s
if
ica
ti
on
model,
e
s
pe
c
ially
in
binar
y
c
las
s
if
ica
ti
on
pr
oblems
.
T
a
ble
1.
M
a
ter
ials
E
qui
pme
nt
C
ha
r
a
c
te
r
is
ti
c
s
/i
n
s
ta
ll
a
ti
on r
e
qui
r
e
me
nt
s
O
bj
e
c
ti
ve
S
pe
c
tr
um a
na
ly
z
e
r
A
nr
it
s
u S
332E
F
r
e
que
nc
y:
100 kHz
t
o 4 G
H
z
, A
ve
r
a
ge
noi
s
e
l
e
ve
l:
152
dB
m t
o 10 Hz
R
B
W
, P
ha
s
e
noi
s
e
:
100 dB
c
/Hz
m
ax
,
C
onne
c
ti
ons
:
E
th
e
r
ne
t,
uni
ve
r
s
a
l
s
e
r
ia
l
bu
s
(
U
S
B
)
c
a
bl
e
, me
mor
y U
S
B
, R
S
-
232
[
13]
M
e
a
s
ur
e
t
he
po
w
e
r
di
s
tr
ib
ut
io
n of
a
s
ig
na
l
a
s
a
f
unc
ti
on of
f
r
e
que
nc
y a
nd t
im
e
.
R
F
ge
ne
r
a
to
r
R
&S
S
M
B
100 A
P
ha
s
e
noi
s
e
S
S
B
:
-
108 dB
c
(
tí
p.)
a
t
10 G
H
z
a
nd
c
ompe
ns
a
ti
on of
20 kHz
, B
r
oa
dba
nd noi
s
e
:
-
138 dB
c
a
t
10 G
H
z
a
nd 30 M
H
z
c
ompe
n
s
a
ti
on,
M
a
x output
pow
e
r
+
27 dB
m
[
14]
G
e
ne
r
a
te
t
he
c
ont
in
uous
R
F
w
a
ve
s
ig
na
l
th
a
t
f
unc
ti
ons
a
s
t
he
r
a
da
r
s
ig
na
l.
I
t
is
s
ync
hr
oni
z
e
d
w
it
h t
he
r
e
c
e
iv
e
r
(
S
pe
c
tr
um a
na
ly
z
e
r
)
.
C
a
n a
nt
e
nna
s
F
r
e
que
nc
y r
a
nge
a
nt
e
nna
1:
2.2
-
2.7 G
H
z
;
F
r
e
que
nc
y
r
a
nge
a
nt
e
nna
2:
2.16
-
2.5 G
H
z
. A
ppr
oxi
ma
te
ba
ndw
id
th
:
A
nt
e
nna
1:
500 M
H
z
;
A
nt
e
nn
a
2:
340 M
H
z
S
e
lf
-
im
pl
e
me
nt
e
d a
nt
e
nna
s
t
une
d t
o t
he
r
a
da
r
'
s
ope
r
a
ti
ng f
r
e
que
nc
y.
C
oa
xi
a
l
c
a
bl
e
I
ns
ul
a
ti
on r
e
s
is
ta
nc
e
:
5.000
M
Ω
min
, I
mpe
da
nc
e
50 Ω
, V
S
W
R
:
1,3 m
ax
, R
a
ng
e
f
r
e
que
nc
y:
0
-
4 G
H
z
[
15]
C
a
bl
e
t
o s
ync
hr
oni
z
e
t
he
c
lo
c
k of
t
he
R
F
ge
ne
r
a
to
r
a
nd t
he
s
pe
c
tr
um a
na
ly
z
e
r
.
S
M
A
t
o S
M
A
c
a
bl
e
s
F
r
e
que
nc
y r
a
nge
:
ma
x 18 G
H
z
. I
mpe
da
nc
e
:
50 Ω
[
16]
T
r
a
ns
mi
t
hi
gh
-
f
r
e
que
nc
y s
ig
na
ls
w
it
h l
ow
s
ig
n
a
l
lo
s
s
be
twe
e
n
e
qui
pme
nt
a
nd a
nt
e
nn
a
s
.
S
of
twa
r
e
ma
s
te
r
to
ol
s
A
nr
it
s
u
S
of
twa
r
e
pr
opr
ie
ta
r
y of
A
n
r
it
s
u
[
17]
A
c
qui
s
it
io
n, ha
ndl
in
g, s
to
r
a
ge
, a
nd i
nt
e
r
pr
e
ta
ti
on
of
t
he
da
ta
obt
a
in
e
d dur
in
g t
he
t
e
s
ts
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
455
-
466
458
2.
1.
P
h
as
e
A:
s
c
e
n
ar
io
d
e
f
i
n
it
ion
T
he
s
e
lec
ted
moveme
nts
f
or
the
de
tec
t
a
nd
e
va
luate
of
ne
ur
a
l
ne
twor
k
models
we
r
e
:
r
unning
,
jum
ping,
a
r
m
r
a
is
ing,
a
nd
wa
lki
ng
.
T
he
s
e
c
omm
on
moveme
nts
that
c
a
n
e
na
ble
non
-
ve
r
ba
l
c
omm
unica
ti
on
a
nd
a
r
e
us
e
d
in
e
ve
r
yda
y
li
f
e
.
Huma
ns
r
un
to
c
a
tch
a
bus
,
jum
p
to
a
void
obs
tac
les
in
their
pa
th,
r
a
is
e
their
a
r
ms
f
or
a
c
ti
ve
b
r
e
a
ks
or
s
tr
e
tching
a
f
ter
s
it
ti
ng
f
o
r
a
n
e
xtende
d
pe
r
iod
,
a
nd
wa
lk
to
move
f
r
om
one
plac
e
to
a
nother
.
T
he
s
e
moveme
nts
c
a
n
a
ls
o
c
ha
r
a
c
ter
i
z
e
s
pe
c
if
ic
a
c
ti
vit
ies
,
f
a
c
il
it
a
ti
ng
the
identif
ica
t
ion
a
nd
unde
r
s
tanding
of
be
ha
vior
s
.
Onc
e
the
moveme
nts
we
r
e
de
f
ined,
the
number
of
r
e
pe
ti
ti
ons
f
or
da
ta
c
oll
e
c
ti
on
a
nd
s
ubs
e
que
nt
da
tas
e
t
c
r
e
a
ti
on
wa
s
s
pe
c
if
ied.
T
a
ble
2
de
s
c
r
ibes
the
moveme
nts
a
nd
the
r
e
pe
ti
ti
on
f
r
e
que
nc
ies
a
t
whic
h
mea
s
ur
e
ments
we
r
e
take
n.
T
a
ble
2.
M
ove
ments
with
their
f
r
e
que
nc
ies
a
nd
dis
tanc
e
s
M
ove
me
nt
s
# of
r
e
pe
ti
ti
ons
D
is
ta
nc
e
(
m)
N
umbe
r
of
t
e
s
ts
W
a
lk
1
9.4
3
J
ump
6
-
3
R
un
1
9.4
3
R
a
is
e
A
r
ms
6
-
3
T
he
s
c
e
na
r
io
wa
s
im
pleme
nted
us
ing
the
e
quipm
e
nt
a
nd
mate
r
ials
li
s
ted
in
T
a
ble
1
,
a
nd
it
s
diagr
a
m
is
de
picte
d
in
F
igu
r
e
2
.
T
he
c
onti
nuous
-
wa
ve
r
a
da
r
wa
s
im
pleme
nted
by
e
mpl
oying
a
r
a
dio
f
r
e
que
nc
y
ge
ne
r
a
tor
a
s
the
t
r
a
ns
mi
tt
ing
e
quipm
e
nt
c
onf
igur
e
d
to
t
r
a
ns
mi
t
a
s
ine
wa
ve
a
t
a
s
pe
c
if
ic
f
r
e
que
nc
y
a
nd
powe
r
leve
l.
T
he
r
e
c
e
iver
is
a
s
pe
c
tr
um
a
na
lyze
r
with
s
pe
c
tr
ogr
a
m
f
unc
ti
ona
li
ty
a
nd
e
ther
ne
t
c
om
muni
c
a
ti
on
with
a
pe
r
s
ona
l
c
omput
e
r
(
P
C
)
.
S
ync
h
r
oniza
ti
on
be
twe
e
n
the
tr
a
ns
mi
tt
e
r
a
nd
r
e
c
e
iver
is
a
c
hieve
d
us
ing
a
c
oa
xial
c
a
ble
c
onne
c
ted
to
the
B
a
yone
t
Ne
il
l
-
C
onc
e
lm
a
n
(
B
NC
)
c
onne
c
tor
s
de
s
ignate
d
f
or
thi
s
pu
r
pos
e
on
e
a
c
h
e
quipm
e
nt.
T
he
c
us
tom
-
made
c
ontr
oll
e
r
a
r
e
a
ne
twor
k
(
C
AN
-
type)
a
ntenna
s
we
r
e
tuned
us
ing
a
ve
c
tor
ne
twor
k
a
na
lyze
r
.
T
he
mas
ter
too
ls
s
of
twa
r
e
on
th
e
P
C
wa
s
us
e
d
f
or
s
pe
c
tr
ogr
a
m
a
c
quis
it
ion.
F
igur
e
2.
P
r
opos
e
d
s
c
e
na
r
io
f
or
da
ta
c
oll
e
c
ti
on
2.
2.
P
h
as
e
B
:
s
am
p
le
ac
q
u
is
it
ion
an
d
s
ign
al
p
r
o
c
e
s
s
in
g
I
n
thi
s
pha
s
e
,
s
a
mpl
e
a
c
quis
it
ion
a
nd
s
ignal
pr
oc
e
s
s
ing
we
r
e
c
onduc
ted
f
or
a
na
lys
is
.
A
c
onti
nuous
-
wa
ve
s
ignal
f
r
e
que
nc
y
of
2
,
395
M
Hz
wa
s
s
e
lec
t
e
d
with
a
tr
a
ns
mi
s
s
ion
powe
r
of
-
10
dB
m.
S
pe
c
t
r
ogr
a
ms
we
r
e
c
onf
igur
e
d
to
be
r
e
c
e
ived
with
a
r
e
f
e
r
e
nc
e
l
e
ve
l
of
-
4
dB
m,
a
s
pa
n
o
f
201
Hz
,
a
r
e
s
olut
ion
b
a
ndwidth
(
R
B
W
)
of
10
Hz
,
a
nd
a
n
a
c
quis
it
ion
ti
me
of
45
s
e
c
onds
.
T
he
r
e
c
e
ived
powe
r
is
a
ppr
oxim
a
tely
-
49
dB
m.
I
t
is
im
por
tant
to
c
ons
ider
the
p
r
oxim
it
y
be
twe
e
n
the
tr
a
ns
mi
tt
ing
a
nd
r
e
c
e
ivi
ng
a
ntenna
s
a
nd
the
a
b
s
e
nc
e
of
is
olation
be
twe
e
n
them,
a
pa
r
t
f
r
om
the
dis
tanc
e
.
F
igu
r
e
3
a
ll
ows
us
to
obs
e
r
ve
one
of
the
obtaine
d
s
pe
c
tr
ogr
a
ms
.
T
his
s
pe
c
tr
ogr
a
m
is
a
vis
ua
l
r
e
pr
e
s
e
ntation
that
il
lus
tr
a
tes
the
e
ne
r
gy
va
r
iation
of
dif
f
e
r
e
nt
f
r
e
que
nc
ies
in
a
r
a
dio
s
ignal
ove
r
ti
me
[
18
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
Hum
an
motion
c
las
s
if
ication
by
mic
r
o
-
doppler
r
adar
us
ing
…
(
A
ndr
e
s
F
e
li
pe
A
r
ias
B
all
e
n
)
459
T
he
us
e
o
f
c
olo
r
s
r
e
pr
e
s
e
nts
thi
s
f
r
e
que
nc
y
of
e
ne
r
gy
or
int
e
ns
it
y
a
t
a
given
mom
e
nt.
T
he
da
r
ke
r
c
olor
s
typi
c
a
ll
y
ind
ica
te
lowe
r
powe
r
,
while
wa
r
mer
c
olor
s
s
uc
h
a
s
gr
e
e
n
o
r
ye
ll
ow
indi
c
a
te
high
e
r
powe
r
.
T
he
f
r
e
que
nc
y
s
hif
ts
to
the
r
ight
o
f
the
c
e
ntr
a
l
f
r
e
q
ue
nc
y
in
the
s
pe
c
tr
ogr
a
m
f
r
e
que
nc
y
incr
e
a
s
e
r
e
pr
e
s
e
nts
the
moveme
nt
of
a
pe
r
s
on
a
ppr
oa
c
hing
the
r
a
da
r
.
T
he
gr
e
a
ter
the
s
hif
t
,
the
h
igher
the
pe
r
s
on's
ve
lo
c
it
y.
T
he
f
r
e
que
nc
y
s
hif
ts
to
the
lef
t
f
r
e
que
nc
y
de
c
r
e
a
s
e
r
e
p
r
e
s
e
nts
a
pe
r
s
on
movi
ng
a
wa
y
f
r
o
m
the
r
a
da
r
.
T
he
s
e
s
hif
ts
a
r
e
c
a
us
e
d
by
the
dopp
ler
e
f
f
e
c
t,
whic
h
oc
c
ur
s
w
he
n
ther
e
is
r
e
lative
mot
ion
be
twe
e
n
the
s
ignal
s
our
c
e
,
in
thi
s
c
a
s
e
,
the
pe
r
s
on,
a
nd
the
s
ignal
r
e
c
e
iver
that
is
r
e
c
or
ding
the
moveme
nt
[
19]
.
T
he
Dopple
r
e
f
f
e
c
t
manif
e
s
ts
a
s
a
c
ha
nge
in
s
ignal
f
r
e
que
nc
ies
a
s
a
p
e
r
s
on
moves
.
W
he
n
a
pe
r
s
on
is
c
los
e
r
to
the
a
nte
nna
s
,
the
s
ignal
will
ha
ve
mor
e
powe
r
a
nd
tend
towa
r
ds
wa
r
mer
c
olor
s
,
a
s
obs
e
r
ve
d
in
the
pr
e
vious
f
igur
e
.
C
onve
r
s
e
ly,
if
the
pe
r
s
on
moves
a
wa
y
f
r
om
the
a
ntenna
s
,
the
s
ignal
will
e
xhibi
t
c
ooler
c
olo
r
s
.
F
igur
e
4
de
picts
the
im
pleme
nted
s
c
e
na
r
io,
whe
r
e
the
a
ntenna
s
a
r
e
s
pa
c
e
d
30
c
m
a
pa
r
t.
W
it
h
thi
s
c
onf
igur
a
ti
on,
we
c
onduc
ted
the
a
c
quis
it
ion
of
500
s
a
mpl
e
im
a
ge
s
c
a
ptur
ing
va
r
ious
movem
e
nts
.
T
he
s
pe
c
tr
ogr
a
m
im
a
ge
s
pr
ovided
by
the
s
of
twa
r
e
we
r
e
r
e
s
ize
d
to
700×
274
pixels
with
a
de
pth
o
f
8
b
it
s
,
r
e
s
ult
ing
in
a
f
il
e
s
ize
of
11
k
B.
F
igur
e
3.
S
pe
c
tr
og
r
a
m
c
ha
r
a
c
ter
is
ti
c
s
F
igur
e
4.
P
r
opos
e
d
s
c
e
na
r
io
with
30
c
m
s
pa
c
ing
be
twe
e
n
a
ntenna
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
455
-
466
460
T
o
mi
ti
ga
te
ove
r
f
it
ti
ng
a
nd
e
ns
ur
e
r
obus
t
model
e
va
luation,
the
da
tas
e
t
wa
s
divi
de
d
int
o
80
%
of
im
a
ge
s
f
or
tr
a
ini
ng
a
nd
20
%
f
o
r
va
li
da
ti
on
.
T
hi
s
a
ppr
oa
c
h
r
e
s
e
r
ve
s
a
por
ti
on
of
the
da
ta
f
o
r
va
li
da
ti
on,
e
na
bli
ng
a
s
s
e
s
s
ment
of
the
model'
s
pe
r
f
or
manc
e
with
da
ta
not
uti
li
z
e
d
dur
ing
tr
a
ini
ng.
B
y
divi
ding
the
da
ta
int
o
tr
a
ini
ng
a
nd
va
li
da
ti
on
s
e
ts
,
thi
s
s
tr
a
tegy
e
ns
ur
e
s
im
pr
ove
d
ge
ne
r
a
li
z
a
ti
on
of
the
model
a
nd
be
tt
e
r
a
da
ptation
to
ne
w
da
ta,
thus
e
nha
nc
ing
it
s
a
bil
it
y
to
a
ddr
e
s
s
diver
s
e
s
c
e
na
r
ios
in
the
f
utur
e
.
T
he
d
a
tas
e
t
of
500
im
a
ge
s
wa
s
labe
led
a
s
W
AR
J
M
AX
W
E
L
L
s
a
mpl
e
c
oll
e
c
ti
on
took
plac
e
in
the
labo
r
a
tor
ies
of
Unive
r
s
idad
M
il
it
a
r
Nue
va
Gr
a
na
da
,
invol
ving
a
to
tal
of
60
indi
viduals
a
ge
d
be
twe
e
n
18
a
nd
50
ye
a
r
s
.
F
igur
e
5
s
hows
the
s
c
e
na
r
io
with
the
e
quipm
e
nt
us
e
d
a
nd
the
wa
y
s
a
mpl
e
s
a
r
e
c
a
ptur
e
d
f
or
two
of
the
f
our
s
e
lec
ted
moveme
nts
.
F
igur
e
5(
a
)
s
hows
the
e
quipm
e
nt
us
e
d
in
the
im
pleme
nted
s
c
e
na
r
io.
F
igur
e
5(
b)
de
mons
tr
a
tes
how
s
a
mpl
e
s
a
r
e
a
c
quir
e
d
while
the
tar
ge
t
pe
r
s
on
r
a
is
e
s
their
a
r
ms
,
a
nd
F
igur
e
5
(
c
)
s
hows
the
a
c
quis
it
ion
pr
oc
e
s
s
whe
n
the
pe
r
s
on
is
wa
lki
ng.
T
he
pe
r
s
on
mus
t
s
tand
dir
e
c
tl
y
in
f
r
ont
o
f
the
p
r
o
tot
ype
to
pe
r
f
or
m
the
moveme
nt
be
c
a
us
e
the
a
ntenna
s
a
r
e
dir
e
c
ti
ona
l
a
nd
ha
ve
a
na
r
r
ow
be
a
mwidt
h.
F
igur
e
6
s
hows
the
f
ou
r
s
pe
c
tr
ogr
a
ms
f
o
r
e
a
c
h
of
the
f
ou
r
mo
ve
ments
to
be
c
las
s
if
ied:
r
unning
F
igur
e
6
(
a
)
,
wa
lki
ng
F
igur
e
6(
b)
,
jum
ping
F
igur
e
6(
c
)
,
a
nd
r
a
is
ing
a
r
ms
F
igur
e
6(
d)
.
T
he
c
e
ntr
a
l
pa
r
t
of
the
s
pe
c
tr
ogr
a
m
bl
ue
c
olor
c
or
r
e
s
ponds
to
the
c
a
r
r
ier
s
ignal,
whic
h
is
vis
ua
li
z
e
d
with
h
igher
powe
r
.
T
he
late
r
a
l
c
omponents
ye
l
low
a
nd
r
e
d
c
olor
s
c
or
r
e
s
pond
to
the
dopple
r
s
hif
t
in
f
r
e
que
nc
ies
due
to
the
moveme
nt
.
(
a
)
(
b)
(
c
)
F
igur
e
5.
P
r
opos
e
d
s
c
e
na
r
io
with
e
quipm
e
nt
us
e
d
f
or
de
tec
ti
ng
a
r
m
-
r
a
is
ing
a
nd
wa
lki
ng
moveme
nts
(
a
)
s
c
e
na
r
io
pr
opos
e
d
with
us
e
d
e
quipm
e
nt,
(
b)
s
pe
c
tr
ogr
a
m
a
c
quis
it
ion
of
r
a
is
e
d
a
r
ms
,
a
nd
(
c
)
s
pe
c
tr
ogr
a
m
a
c
quis
it
ion
while
wa
lki
ng
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
Hum
an
motion
c
las
s
if
ication
by
mic
r
o
-
doppler
r
adar
us
ing
…
(
A
ndr
e
s
F
e
li
pe
A
r
ias
B
all
e
n
)
461
(
a
)
(
b)
(
c
)
(
d)
F
igur
e
6.
S
pe
c
tr
og
r
a
m
(
a
)
r
unning
,
(
b)
wa
lki
ng,
(
c
)
jum
ping,
a
nd
(
d
)
r
a
is
ing
a
r
ms
2.
3.
P
h
as
e
C:
n
e
t
wor
k
ar
c
h
it
e
c
t
u
r
e
s
e
lec
t
ion
an
d
t
r
ain
in
g
F
or
the
c
las
s
if
ica
ti
on
of
moveme
nts
,
6
ne
ur
a
l
ne
twor
k
models
we
r
e
s
e
lec
ted:
VG
G
-
16,
VG
G
-
19,
M
obil
e
Ne
t,
M
obil
e
Ne
t
V2
,
Xc
e
pti
on
,
a
nd
I
nc
e
pti
on
V3.
S
ome
o
f
the
c
ha
r
a
c
ter
is
ti
c
s
of
thes
e
mo
de
ls
a
r
e
s
hown
in
T
a
ble
3
.
I
n
thi
s
c
a
s
e
,
“
S
ize
(
M
B
)
”
r
e
pr
e
s
e
nts
the
s
ize
in
mega
bytes
(
M
B
)
of
the
model
a
f
t
e
r
be
ing
tr
a
ined,
“
pa
r
a
mete
r
s
”
indi
c
a
tes
the
number
of
pa
r
a
mete
r
s
or
c
oe
f
f
icie
nts
the
model
ha
s
lea
r
ne
d
d
ur
ing
the
tr
a
ini
ng
pr
oc
e
s
s
,
a
nd
“
de
pth”
r
e
pr
e
s
e
nts
the
de
pt
h
or
number
of
laye
r
s
the
model
ha
s
.
T
he
mor
e
laye
r
s
a
model
ha
s
,
the
de
e
pe
r
it
is
.
T
a
ble
3.
T
e
c
hnica
l
s
pe
c
if
ica
ti
ons
of
the
us
e
d
mode
ls
M
ode
l
S
iz
e
(
M
B
)
P
a
r
a
me
te
r
s
D
e
s
c
r
ip
ti
on
D
e
pt
h
VGG
-
16
528
138.4 M
T
he
V
G
G
-
16 a
r
c
hi
te
c
tu
r
e
c
ons
is
ts
of
16 l
a
ye
r
s
, i
nc
lu
di
ng 13 c
o
nvol
ut
io
na
l
la
ye
r
s
w
it
h 3
×
3 f
il
te
r
s
a
nd z
e
r
o
pa
ddi
ng, f
ol
lo
w
e
d by 2
×
2
M
a
x
pool
in
g l
a
ye
r
s
to
r
e
duc
e
di
me
ns
io
na
li
ty
. A
f
te
r
w
a
r
d, t
he
r
e
a
r
e
3 f
ul
ly
c
onne
c
te
d l
a
ye
r
s
r
e
s
pons
ib
le
f
or
c
la
s
s
if
ic
a
ti
on t
a
s
ks
.
T
hi
s
s
tr
uc
tu
r
e
e
n
a
bl
e
s
V
G
G
-
16 t
o
pr
ogr
e
s
s
iv
e
ly
l
e
a
r
n f
r
om s
im
pl
e
f
e
a
tu
r
e
s
l
ik
e
e
dge
s
t
o mor
e
c
o
mpl
e
x f
e
a
tu
r
e
s
in
i
ma
ge
s
, ma
ki
ng i
t
a
n e
f
f
e
c
ti
ve
de
e
p ne
twor
k f
or
c
omput
e
r
vi
s
io
n
[
20]
, [
21]
.
16
VGG
-
19
549
143.7 M
VGG
-
19 i
s
a
n e
xt
e
nde
d ve
r
s
io
n of
t
he
V
G
G
-
16 a
r
c
hi
te
c
tu
r
e
, c
ha
r
a
c
te
r
iz
e
d by
it
s
de
pt
h a
nd unif
or
mi
ty
i
n t
he
a
r
r
a
nge
me
nt
of
19 l
a
ye
r
s
, i
nc
lu
di
ng both
c
onvolut
io
na
l
a
nd f
ul
ly
c
onne
c
te
d l
a
ye
r
s
. L
ik
e
V
G
G
-
16, i
t
e
mp
lo
ys
s
ma
ll
f
il
te
r
s
a
nd M
a
x
-
pool
in
g l
a
ye
r
s
t
o e
xt
r
a
c
t
f
e
a
tu
r
e
s
f
r
om i
ma
ge
s
.
U
s
e
d i
n t
a
s
k
s
s
uc
h a
s
c
la
s
s
if
ic
a
ti
on, obje
c
t
de
te
c
ti
on, a
nd mor
e
, V
G
G
-
19 ha
s
be
e
n
s
ig
ni
f
ic
a
nt
i
n c
omput
e
r
vi
s
io
n, de
mons
tr
a
ti
ng s
ol
id
pe
r
f
or
ma
nc
e
, a
lt
hough
a
c
c
ur
a
c
y ma
y va
r
y de
pe
ndi
ng on the
d
a
ta
s
e
t
a
nd s
pe
c
if
ic
t
a
s
k
[
22]
, [
23]
.
19
M
obi
le
N
e
t
16
4.3 M
T
he
e
xa
c
t
numbe
r
of
l
a
ye
r
s
i
n a
M
obi
le
N
e
t
c
a
n va
r
y de
pe
ndi
ng
on t
he
s
pe
c
if
ic
ve
r
s
io
n a
nd modi
f
ic
a
ti
ons
ma
de
. H
ow
e
v
e
r
, i
n ge
ne
r
a
l,
M
obi
le
N
e
t
a
r
c
hi
te
c
tu
r
e
s
ty
pi
c
a
ll
y c
ons
is
t
of
mul
ti
pl
e
c
onvolut
io
na
l
a
nd pooli
ng l
a
ye
r
s
, a
s
w
e
ll
a
s
f
ul
ly
c
onne
c
te
d l
a
ye
r
s
a
t
th
e
e
nd f
or
c
la
s
s
if
ic
a
ti
on or
a
s
pe
c
if
ic
t
a
s
k.
F
or
e
xa
mpl
e
,
M
obi
le
N
e
tV1 c
ons
i
s
ts
of
a
ppr
oxi
ma
te
ly
55 l
a
ye
r
s
.
T
he
s
e
ne
two
r
ks
a
r
e
de
s
ig
ne
d f
or
a
ppl
ic
a
ti
ons
on mobi
le
de
vi
c
e
s
a
nd e
mbe
dd
e
d s
y
s
te
ms
, s
tr
ik
in
g a
ba
la
nc
e
be
twe
e
n c
ompl
e
xi
ty
a
nd c
omput
a
ti
ona
l
e
f
f
ic
ie
nc
y
[
24]
.
55
M
obi
le
N
e
t
V2
14
3.5 M
I
t
ut
il
iz
e
s
bui
ld
in
g bl
oc
ks
c
a
ll
e
d “
in
ve
r
te
d r
e
s
id
ua
l
s
”
t
ha
t
opt
im
iz
e
f
e
a
tu
r
e
r
e
pr
e
s
e
nt
a
ti
on i
n de
e
p ne
twor
ks
w
hi
le
h
a
vi
ng a
r
e
duc
e
d numbe
r
of
pa
r
a
me
te
r
s
.
T
hi
s
a
ll
ow
s
f
or
a
ba
la
n
c
e
be
twe
e
n a
c
c
ur
a
c
y a
nd s
p
e
e
d. I
n t
hi
s
p
r
oj
e
c
t,
i
t
c
a
n be
us
e
d i
n t
a
s
k
s
s
uc
h a
s
obj
e
c
t
de
t
e
c
ti
on, i
ma
ge
c
la
s
s
if
ic
a
ti
on, a
nd
ot
he
r
vi
s
io
n
-
r
e
la
te
d t
a
s
ks
, ma
ki
ng i
t
s
ui
ta
bl
e
f
or
c
la
s
s
if
yi
ng move
me
nt
s
[
25]
.
105
X
c
e
pt
io
n
88
22.9 M
I
t
ut
il
iz
e
s
bui
ld
in
g bl
oc
ks
c
a
ll
e
d “
in
ve
r
te
d r
e
s
id
ua
l
s
”
t
ha
t
opt
im
iz
e
f
e
a
tu
r
e
r
e
pr
e
s
e
nt
a
ti
on i
n de
e
p ne
twor
ks
w
hi
le
h
a
vi
ng a
r
e
duc
e
d numbe
r
of
pa
r
a
me
te
r
s
.
T
hi
s
a
ll
ow
s
f
or
a
ba
la
n
c
e
be
twe
e
n a
c
c
ur
a
c
y a
nd s
p
e
e
d. I
n t
hi
s
p
r
oj
e
c
t,
i
t
c
a
n be
us
e
d i
n t
a
s
k
s
s
uc
h a
s
obj
e
c
t
de
t
e
c
ti
on, i
ma
ge
c
la
s
s
if
ic
a
ti
on, a
nd
ot
he
r
vi
s
io
n
-
r
e
la
te
d t
a
s
ks
, ma
ki
ng i
t
s
ui
ta
bl
e
f
or
c
la
s
s
if
yi
ng move
me
nt
s
[
25]
.
81
I
nc
e
pt
io
n
V3
92
23.9 M
I
t
s
ta
nds
out
f
or
t
he
i
mpl
e
me
nt
a
ti
on of
f
a
c
to
r
iz
e
d
c
onvolut
io
ns
, i
n w
hi
c
h
s
ta
nda
r
d 3
×
3 c
onvolut
io
ns
a
r
e
s
e
pa
r
a
te
d i
nt
o t
w
o
s
ma
ll
e
r
c
onvo
lu
ti
ons
(
1
×
3
a
nd 3
×
1)
, w
it
h a
t
ot
a
l
of
189 hidden la
ye
r
s
. T
hi
s
a
ll
ow
s
f
or
c
a
pt
ur
in
g s
pa
ti
a
l
pa
tt
e
r
ns
a
nd r
e
duc
in
g c
omput
a
ti
ona
l
lo
a
d. A
ddi
ti
ona
ll
y, I
nc
e
pt
i
on
-
v3 e
mpl
oys
s
ma
ll
e
r
-
s
iz
e
d f
il
te
r
s
t
o i
mpr
ove
e
f
f
ic
ie
nc
y a
nd r
e
duc
e
t
he
numb
e
r
of
pa
r
a
me
te
r
s
. I
ts
di
s
ti
nc
ti
ve
f
e
a
tu
r
e
i
s
t
he
“
I
nc
e
pt
io
n”
modul
e
s
, w
hi
c
h pe
r
f
or
m
pa
r
a
ll
e
l
c
onvolut
io
ns
w
it
h di
f
f
e
r
e
nt
f
il
te
r
s
iz
e
s
a
nd t
he
n c
onc
a
te
na
te
t
he
ir
r
e
s
ul
ts
, e
na
bl
in
g t
he
ne
twor
k t
o c
a
pt
ur
e
i
nf
or
ma
ti
on a
t
mul
ti
pl
e
s
c
a
le
s
[
23]
.
189
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
455
-
466
462
T
he
e
xpe
r
im
e
ntation
with
the
s
e
lec
ted
models
wa
s
c
onduc
ted
us
ing
T
e
ns
or
F
low
Ke
r
a
s
,
a
longs
ide
the
r
e
quis
it
e
P
ython
li
br
a
r
ies
a
nd
modul
e
s
f
or
im
a
ge
pr
oc
e
s
s
ing
a
nd
d
e
e
p
lea
r
ning
model
c
ons
tr
uc
ti
on.
T
o
opti
mi
z
e
the
p
r
oc
e
s
s
ing
s
pe
e
d
of
tr
a
ini
ng
a
nd
va
li
da
ti
on
da
ta,
c
a
c
he
a
nd
pr
e
f
e
tch
methods
we
r
e
e
mp
loyed.
I
n
or
de
r
to
p
r
e
ve
nt
ove
r
f
it
ti
ng
of
the
model
to
the
or
i
ginal
da
ta,
a
da
ta
a
ugmenta
ti
on
model
wa
s
im
pleme
nted
to
ge
ne
r
a
te
ne
w
im
a
ge
s
f
r
om
the
tr
a
ini
ng
s
e
t
thr
ough
r
a
ndom
tr
a
ns
f
o
r
mations
.
A
f
lexible
f
unc
t
ion
wa
s
de
ve
loped
to
a
c
c
e
pt
va
r
ious
input
pa
r
a
mete
r
s
,
de
f
i
ning
a
c
onvolut
ional
ne
ur
a
l
ne
twor
k
model
a
nd
tr
a
ini
ng
it
on
both
tr
a
ini
ng
a
nd
va
li
da
ti
on
da
tas
e
ts
.
T
he
s
e
input
pa
r
a
mete
r
s
e
nc
ompas
s
c
r
it
ica
l
c
onf
igur
a
ti
ons
pivot
a
l
in
c
ons
tr
uc
ti
ng
a
nd
tr
a
ini
ng
a
de
e
p
ne
ur
a
l
ne
twor
k.
T
he
c
hoice
of
model
a
r
c
hit
e
c
tur
e
,
be
i
t
VG
G16,
V
GG
19,
or
M
obil
e
Ne
t,
de
ter
mi
ne
s
the
o
r
ga
niza
ti
on
of
laye
r
s
withi
n
the
ne
twor
k.
T
he
de
c
is
ion
of
whe
ther
the
laye
r
s
s
hould
be
tr
a
inable
c
ondit
ions
their
a
da
ptabili
ty
to
ne
w
da
ta
or
r
e
tention
of
s
tatic
pr
op
e
r
ti
e
s
.
Dur
ing
thi
s
pr
oc
e
s
s
,
the
opti
m
ize
r
s
uc
h
a
s
Ada
m
or
s
tocha
s
ti
c
gr
a
dient
de
s
c
e
nt
(
S
GD
)
f
a
c
il
it
a
tes
pa
r
a
mete
r
a
djus
tm
e
nts
,
uti
li
z
ing
the
lea
r
ning
r
a
te
to
r
e
gulate
the
magnitude
of
thes
e
a
djus
tm
e
nts
in
e
a
c
h
it
e
r
a
ti
on
.
I
n
a
ddit
ion
to
the
las
t
c
onvolut
ional
block
,
the
incl
us
ion
of
e
xtr
a
laye
r
s
a
nd
tec
hniques
li
ke
dr
opout
inf
luenc
e
the
de
pth
a
nd
r
obus
tnes
s
of
the
ne
twor
k
.
F
u
ll
y
-
c
onne
c
ted
laye
r
s
r
e
f
ine
f
iner
de
tails
,
a
nd
the
number
of
laye
r
s
a
nd
ne
ur
ons
c
a
n
va
r
y
a
c
c
or
dingl
y.
P
r
e
s
e
r
ving
th
e
model
a
nd
it
s
we
ight
s
r
e
tains
the
knowle
dge
a
c
quir
e
d
du
r
ing
the
p
r
oc
e
s
s
,
while
the
number
of
e
poc
hs
s
pe
c
if
ies
the
f
r
e
que
nc
y
a
t
whic
h
the
tr
a
ini
ng
da
ta
will
be
t
r
a
ve
r
s
e
d.
P
r
ope
r
c
a
li
br
a
ti
on
of
thes
e
pa
r
a
mete
r
s
is
c
r
uc
ial
f
or
a
c
hieving
opti
mal
pe
r
f
or
manc
e
in
the
de
s
ir
e
d
tas
k
by
the
ne
ur
a
l
ne
twor
k
.
T
he
input
pa
r
a
mete
r
s
e
n
c
ompas
s
the
model
a
r
c
hit
e
c
tur
e
,
a
nd
the
f
unc
ti
on
r
e
tur
ns
a
his
tor
y
objec
t
that
r
e
c
or
ds
tr
a
ini
ng
a
nd
va
li
da
ti
on
a
c
c
ur
a
c
y
a
nd
los
s
thr
oughout
the
t
r
a
ini
ng
p
r
oc
e
s
s
.
F
igur
e
7
e
lucida
tes
thi
s
de
s
c
r
ipt
ion,
while
T
a
ble
4
de
li
n
e
a
tes
the
s
e
lec
ti
on
of
hype
r
pa
r
a
mete
r
s
uti
li
z
e
d
f
or
a
ll
model
s
.
F
igur
e
7.
VG
G
-
16
a
r
c
hit
e
c
tur
e
T
a
ble
4.
Hype
r
pa
r
a
mete
r
s
us
e
d
M
e
tr
ic
V
a
lu
e
L
e
a
r
ni
ng r
a
te
0.01
D
r
op r
a
te
0.01
E
poc
hs
10
M
a
x pooli
ng
2
3.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
T
a
ble
5
s
umm
a
r
ize
s
the
r
e
s
ult
s
obtaine
d
f
r
om
tr
a
ini
ng
the
models
.
T
he
VG
G
-
16
model
with
the
Ada
m
opti
mi
z
e
r
a
c
hieve
s
a
tr
a
ini
ng
a
c
c
ur
a
c
y
of
a
r
ound
100%
a
nd
a
va
li
da
ti
on
a
c
c
ur
a
c
y
of
a
ppr
o
xim
a
tely
96%
.
Addit
ionally,
both
t
r
a
ini
ng
a
nd
va
li
da
ti
on
los
s
e
s
a
r
e
lowe
r
c
ompar
e
d
to
o
ther
models
.
T
his
s
ugg
e
s
ts
that
the
VG
G
-
16
model
is
c
a
pa
ble
of
s
uc
c
e
s
s
f
ull
y
c
las
s
if
ying
im
a
ge
s
int
o
the
two
tar
ge
t
c
las
s
e
s
.
VG
G
-
16
is
a
r
obus
t
a
nd
wide
ly
us
e
d
model,
e
s
pe
c
ially
in
im
a
g
e
c
las
s
if
ica
ti
on
tas
ks
.
I
t
is
known
f
or
it
s
de
e
p
a
nd
unif
or
m
a
r
c
hit
e
c
tur
e
,
making
it
e
f
f
e
c
ti
ve
a
t
e
xt
r
a
c
ti
ng
f
e
a
tur
e
s
f
r
om
im
a
ge
s
of
di
f
f
e
r
e
nt
s
c
a
les
a
nd
c
ompl
e
xit
ies
.
How
e
ve
r
,
due
to
it
s
de
pth,
it
ha
s
a
r
e
latively
lar
ge
number
of
pa
r
a
mete
r
s
,
whic
h
c
a
n
r
e
s
ult
in
a
la
r
g
e
r
model
s
ize
a
nd
r
e
qui
r
e
mo
r
e
c
omput
a
ti
ona
l
r
e
s
our
c
e
s
f
or
t
r
a
ini
ng
a
nd
e
xe
c
uti
on.
T
ha
nks
to
the
-
10
dB
c
onf
igur
a
ti
on,
the
da
tas
e
t
e
f
f
e
c
ti
ve
ly
c
a
ptur
e
s
hu
man
mot
ion
without
r
e
f
lec
ti
ons
or
int
e
r
f
e
r
e
nc
e
f
r
om
the
e
nvir
onment.
R
e
ga
r
ding
the
t
im
e
r
e
quir
e
d
to
tr
a
in
the
model,
ther
e
is
a
notable
s
im
il
a
r
it
y
in
the
r
e
s
ult
s
.
T
he
s
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
Hum
an
motion
c
las
s
if
ication
by
mic
r
o
-
doppler
r
adar
us
ing
…
(
A
ndr
e
s
F
e
li
pe
A
r
ias
B
all
e
n
)
463
ti
mes
we
r
e
obtaine
d
us
ing
a
gr
a
phics
pr
oc
e
s
s
ing
unit
(
GPU)
r
unt
im
e
,
with
a
tot
a
l
ti
me
o
f
2
mi
n
utes
.
T
he
a
ve
r
a
ge
tr
a
ini
ng
ti
me
f
or
the
6
models
wa
s
a
ppr
o
xim
a
tely
1.
16
mi
nu
tes
.
All
models
a
r
e
c
a
pa
ble
of
c
or
r
e
c
tl
y
c
las
s
if
ying
the
4
types
of
moveme
nts
with
high
a
c
c
ur
a
c
y
in
a
s
hor
t
tr
a
ini
ng
ti
me,
a
s
s
hown
in
T
a
ble
5
.
T
a
ble
6
r
e
s
ume
a
c
c
ur
a
c
y,
r
e
c
a
ll
,
a
nd
F
1
s
c
or
e
me
tr
ics
r
e
s
ult
s
.
T
he
be
s
t
tr
a
in
a
c
c
ur
a
c
y
r
e
s
ult
s
we
r
e
obtaine
d
by
the
VG
G
-
16,
VG
G
-
19,
M
obil
e
N
e
t,
a
nd
M
obil
e
N
e
tV2
models
.
How
e
ve
r
,
a
c
c
or
ding
to
va
li
da
ti
on
a
c
c
ur
a
c
y,
the
model
that
be
s
t
ge
ne
r
a
li
z
e
s
the
va
li
da
ti
on
da
ta
is
VG
G
-
19,
c
ons
ider
ing
that
the
other
va
li
da
ti
on
a
c
c
ur
a
c
y
r
e
s
ult
s
a
r
e
a
bove
0.
900.
De
s
pit
e
VG
G
-
19
s
howin
g
the
lowe
s
t
tr
a
in
los
s
,
it
is
e
vident
that
the
mod
e
l
is
not
ove
r
f
it
ti
ng
be
c
a
us
e
it
ge
ne
r
a
li
z
e
s
the
va
li
da
ti
on
da
ta
c
or
r
e
c
tl
y
with
a
va
li
da
ti
on
a
c
c
ur
a
c
y
o
f
0.
970
a
n
d
a
tr
a
in
a
c
c
ur
a
c
y
of
1.
T
a
ble
5.
R
e
s
ult
s
obtaine
d;
lea
r
ning
r
a
te:
0.
01;
de
ns
e
laye
r
s
:
1024;
number
of
e
poc
hs
:
10
M
ode
l
O
pt
im
iz
e
r
V
a
l
a
c
c
ur
a
c
y
T
r
a
in
a
c
c
ur
a
c
y
T
r
a
in
l
os
s
V
a
l
lo
s
s
T
im
e
(
M
)
VGG
-
16
S
G
D
0.849
0.943
0.232
0.421
2
A
da
m
0.961
1.000
4.191E
-
04
0.430
2
VGG
-
19
A
da
m
0.970
1.000
1.608E
-
04
0.099
3
S
G
D
0.207
0.7037
14.606
112.455
3
M
obi
le
N
e
t
A
da
m
0.934
0.997
0.022
0.251
0
S
G
D
0.910
1000
0.015
0.261
0
M
obi
le
N
e
t
V2
A
da
m
0.930
0.990
0.029
0.363
0
S
G
D
0.920
1.000
0.021
0.256
0
X
c
e
pt
io
n
A
da
m
0.860
0.820
0.383
0.703
1
S
G
D
0.840
0.972
0.184
0.382
1
I
nc
e
pt
io
n
V3
A
da
m
0.640
0.815
0.516
1.152
1
S
G
D
0.880
0.905
0.331
0.347
1
T
a
ble
6.
R
e
s
ult
s
of
e
va
luate
d
metr
ics
f
o
r
the
6
mo
de
ls
us
e
d
M
ode
l
M
ove
me
nt
s
R
un
J
ump
R
a
is
e
a
r
ms
w
a
lk
VGG
-
16
P
r
e
c
is
io
n=
1
R
e
c
a
ll
=
0.98
F
1 S
c
or
e
=
0.98
P
r
e
c
is
io
n=
0.92
R
e
c
a
ll
=
0.98
F
1 S
c
or
e
=
0.96
P
r
e
c
is
io
n=
0.98
R
e
c
a
ll
=
0.94
F
1 S
c
or
e
=
0.96
P
r
e
c
is
io
n=
0.96
R
e
c
a
ll
=
1.00
F
1 S
c
or
e
=
0.98
VGG
-
19
P
r
e
c
is
io
n=
0.92
R
e
c
a
ll
=
0.94
F
1 S
c
or
e
=
0.95
P
r
e
c
is
io
n=
0.94
R
e
c
a
ll
=
1
F
1 S
c
or
e
=
0.96
P
r
e
c
is
io
n=
0.96
R
e
c
a
ll
=
0.94
F
1 S
c
or
e
=
0.94
P
r
e
c
is
io
n=
0.94
R
e
c
a
ll
=
0.96
F
1 S
c
or
e
=
0.98
M
obi
le
N
e
t
P
r
e
c
is
io
n=
0.95
R
e
c
a
ll
=
0.94
F
1 S
c
or
e
=
0.94
P
r
e
c
is
io
n=
0.94
R
e
c
a
ll
=
0.92
F
1 S
c
or
e
=
0.94
P
r
e
c
is
io
n=
0.96
R
e
c
a
ll
=
0.94
F
1 S
c
or
e
=
0.98
P
r
e
c
is
io
n=
0.94
R
e
c
a
ll
=
0.92
F
1 S
c
or
e
=
0.92
M
obi
le
N
e
t
V
2
P
r
e
c
is
io
n=
0.96
R
e
c
a
ll
=
0.94
F
1 S
c
or
e
=
0.94
P
r
e
c
is
io
n=
0.86
R
e
c
a
ll
=
0.90
F
1 S
c
or
e
=
0.90
P
r
e
c
is
io
n=
0.92
R
e
c
a
ll
=
0.94
F
1
S
c
or
e
=
0.91
P
r
e
c
is
io
n=
0.92
R
e
c
a
ll
=
0.92
F
1 S
c
or
e
=
0.96
X
c
e
pt
io
n
P
r
e
c
is
io
n=
0.90
R
e
c
a
ll
=
0.92
F
1 S
c
or
e
=
0.90
P
r
e
c
is
io
n=
0.86
R
e
c
a
ll
=
0.88
F
1 S
c
or
e
=
0.92
P
r
e
c
is
io
n=
0.90
R
e
c
a
ll
=
0.84
F
1 S
c
or
e
=
0.88
P
r
e
c
is
io
n=
0.90
R
e
c
a
ll
=
0.94
F
1 S
c
or
e
=
0.96
I
nc
e
pt
io
n
P
r
e
c
is
io
n=
0.90
R
e
c
a
ll
=
0.94
F
1 S
c
or
e
=
0.86
P
r
e
c
is
io
n=
0.86
R
e
c
a
ll
=
0.86
F
1 S
c
or
e
=
0.84
P
r
e
c
is
io
n=
0.90
R
e
c
a
ll
=
0.84
F
1 S
c
or
e
=
0.92
P
r
e
c
is
io
n=
0.90
R
e
c
a
ll
=
0.86
F
1 S
c
or
e
=
0.84
T
he
model
f
a
c
e
d
c
ha
ll
e
nge
s
in
a
c
c
ur
a
tely
c
la
s
s
if
ying
the
“
jum
ping
”
moveme
nt,
e
xhibi
ti
ng
the
lowe
s
t
pr
e
c
is
ion
a
mong
a
ll
c
a
tegor
ize
d
moveme
n
ts
.
C
onve
r
s
e
ly,
the
moveme
nt
c
las
s
if
ied
with
the
highes
t
pr
e
c
is
ion
by
the
models
is
“
r
unning.
”
T
h
is
dis
ti
nc
ti
on
c
a
n
be
a
tt
r
ibut
e
d
to
the
s
pe
c
tr
ogr
a
m's
typi
c
a
ll
y
mor
e
pr
onounc
e
d
pr
e
s
e
nc
e
,
f
a
c
il
it
a
ti
ng
it
s
di
f
f
e
r
e
nti
a
ti
on
f
r
om
other
moveme
nts
.
F
igur
e
8
s
howc
a
s
e
s
the
pe
r
f
or
manc
e
r
e
s
ult
s
of
the
VG
G
-
16
model
th
r
oug
h
a
c
onf
us
ion
matr
ix
.
W
it
h
a
n
a
c
c
ur
a
c
y
e
xc
e
e
ding
92%
in
r
e
c
ognizing
a
r
m
moveme
nts
,
the
model
de
mo
ns
tr
a
tes
c
omm
e
nda
ble
ove
r
a
ll
a
r
c
hit
e
c
tu
r
a
l
p
r
o
f
icie
nc
y.
None
thele
s
s
,
it
is
notew
or
thy
that
“
a
r
m
moveme
nt
”
oc
c
a
s
ionally
incur
s
c
onf
us
ion
with
“
jum
ping,
”
given
their
s
ha
r
e
d
c
ha
r
a
c
ter
is
ti
c
s
.
De
s
pit
e
thi
s
,
“
jum
p
ing
”
maintains
a
pr
e
c
is
ion
of
100%
,
unde
r
s
c
or
ing
the
a
lgor
it
hm's
a
de
ptnes
s
in
im
a
ge
c
las
s
if
ica
ti
on
a
nd
t
he
VG
G
-
16
a
r
c
hit
e
c
tur
e
's
s
uc
c
e
s
s
f
ul
pe
r
f
or
manc
e
.
I
n
F
i
gu
r
e
9
,
t
he
pr
e
d
ic
ti
on
o
f
a
n
im
a
ge
c
las
s
i
f
ica
ti
on
mo
de
l
w
it
h
i
np
ut
di
me
ns
i
ons
o
f
15
0×
15
0
p
ixe
ls
a
nd
3
c
o
lo
r
c
h
a
n
ne
ls
a
r
e
p
r
e
s
e
n
te
d
.
T
he
m
a
i
n
ob
jec
t
ive
is
to
r
e
c
o
gn
ize
hu
ma
n
a
c
t
i
ons
,
a
nd
s
p
e
c
if
ic
a
l
ly
,
in
th
is
i
ns
ta
nc
e
,
t
he
a
c
t
io
n
o
f
r
a
is
e
a
r
ms
.
T
he
m
od
e
l
's
p
r
e
di
c
ti
on
y
ie
ld
e
d
a
h
ig
hl
y
a
c
c
u
r
a
te
r
e
s
u
lt
,
w
he
r
e
the
c
las
s
R
a
is
e
A
r
ms
ha
s
a
p
r
o
ba
bi
li
ty
o
f
a
p
p
r
o
xi
ma
te
ly
6
8
%
,
c
on
f
i
r
m
i
ng
t
he
r
ob
us
t
ne
s
s
a
n
d
e
f
f
e
c
ti
ve
ne
s
s
o
f
th
e
m
o
de
l
in
i
de
n
ti
f
y
i
ng
t
his
s
p
e
c
if
ic
a
c
t
io
n
wi
th
c
e
r
ta
in
ty
.
T
h
is
p
r
e
c
is
i
on
s
u
pp
o
r
ts
t
he
s
u
it
a
b
il
i
ty
o
f
t
he
p
r
o
pos
e
d
a
p
p
r
o
a
c
h
,
h
i
gh
li
gh
t
in
g
i
ts
a
p
p
li
c
a
b
i
li
ty
in
e
n
vi
r
o
n
men
ts
w
he
r
e
a
c
c
u
r
a
te
id
e
n
ti
f
ic
a
t
io
n
o
f
hu
man
a
c
ti
ons
is
e
s
s
e
n
ti
a
l,
s
u
c
h
a
s
i
n
s
e
c
u
r
it
y
m
on
i
to
r
in
g
s
ys
te
ms
o
r
me
dic
a
l
a
p
pl
ic
a
t
io
ns
f
o
r
th
e
a
n
a
l
ys
is
o
f
s
p
e
c
i
f
ic
m
ove
me
nts
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
455
-
466
464
F
igur
e
8.
C
onf
us
ion
matr
ix
in
pe
r
c
e
ntage
F
igur
e
9.
P
r
e
diction
4.
CONC
L
USI
ON
T
he
VG
G
-
16
model
a
c
hieve
s
outs
tanding
p
r
e
c
is
ion
r
e
s
ult
s
f
or
moveme
nt
identif
ica
ti
on
,
with
va
lues
a
s
f
oll
ows
:
wa
lki
ng
96%
,
r
unning
100%
,
a
nd
a
r
m
r
a
is
ing
98%
.
F
o
r
the
“
jum
ping”
moveme
nt
,
both
VG
G
-
19
a
nd
M
obil
e
Ne
t
s
ur
pa
s
s
e
d
the
VG
G
-
16
model,
a
c
hieving
a
pr
e
c
is
ion
of
94%
.
How
e
ve
r
,
Xc
e
pt
ion
a
nd
I
nc
e
pti
on
models
de
li
ve
r
e
d
the
lea
s
t
f
a
vor
a
ble
pr
e
c
is
ion
va
lues
f
or
identi
f
ying
the
“
jum
ping”
moveme
nt,
with
both
models
s
c
or
ing
86%
.
I
n
ter
ms
of
r
e
c
a
ll
,
VG
G
-
16
s
tands
out
with
a
n
a
ve
r
a
ge
va
lue
of
0.
975.
Not
a
bly,
the
VGG
-
19
model
a
c
hieve
d
a
pe
r
f
e
c
t
r
e
c
a
ll
s
c
or
e
of
1
f
o
r
the
“
jum
ping”
moveme
nt,
the
highes
t
a
mong
a
ll
models
e
va
luate
d.
R
e
ga
r
ding
the
F
1
s
c
or
e
metr
ic,
a
n
a
ve
r
a
ge
va
lu
e
of
0
.
97
wa
s
obtaine
d
f
or
identif
ying
the
f
our
moveme
nts
.
T
he
I
nc
e
pti
on
model
obtaine
d
the
lo
we
s
t
s
c
or
e
in
thi
s
metr
ic,
with
a
va
lue
of
0.
865
.
I
n
ter
ms
of
tr
a
ini
ng
los
s
,
VG
G
-
16
a
c
hieve
d
the
lowe
s
t
va
lue
a
t
4.
191
E
-
04.
C
onve
r
s
e
ly,
f
or
va
li
da
ti
on
l
os
s
,
the
M
obil
e
Ne
t
model
yielde
d
the
mos
t
f
a
vo
r
a
ble
r
e
s
ul
t,
wi
th
a
s
c
or
e
o
f
0
.
251.
T
he
I
nc
e
pti
on
model
r
e
c
or
de
d
the
lea
s
t
f
a
vor
a
ble
r
e
s
ult
s
f
o
r
both
tr
a
ini
ng
a
nd
va
li
da
ti
on
los
s
,
with
a
va
lue
o
f
1.
152
.
Ove
r
a
ll
,
the
VG
G
-
16
model
e
mer
ge
s
a
s
one
of
the
mos
t
e
f
f
e
c
ti
ve
tool
s
f
o
r
s
pe
c
tr
ogr
a
m
c
las
s
if
ica
ti
on.
F
ur
the
r
m
or
e
,
ou
r
im
pleme
nted
s
c
e
na
r
io
unde
r
s
c
or
e
s
the
viabili
ty
of
r
a
da
r
tec
hniques
f
o
r
de
tec
ti
ng
s
low
-
s
pe
e
d
moveme
nts
of
pe
ople
or
objec
ts
whe
n
c
ombi
ne
d
wi
th
int
e
ll
igent
a
lgor
it
hms
a
nd
tec
hnologi
e
s
.
AC
KNOWL
E
DGE
M
E
NT
S
T
his
wor
k
wa
s
de
ve
loped
withi
n
the
GI
S
S
I
C
r
e
s
e
a
r
c
h
gr
oup
,
a
s
a
pr
oduc
t
de
r
ived
f
r
om
the
M
a
xwe
ll
r
e
s
e
a
r
c
h
s
e
e
dbe
d,
e
ndor
s
e
d
by
the
vice
-
r
e
c
tor
a
te
f
o
r
r
e
s
e
a
r
c
h
a
t
the
Unive
r
s
idad
M
il
it
a
r
Nue
va
G
r
a
na
da
.
Evaluation Warning : The document was created with Spire.PDF for Python.