T
E
L
KO
M
N
I
KA
T
e
lec
om
m
u
n
icat
ion
,
Com
p
u
t
i
n
g,
E
lec
t
r
on
ics
an
d
Cont
r
ol
Vol.
18
,
No.
3
,
J
une
2020
,
pp.
1671
~
167
7
I
S
S
N:
1693
-
6930,
a
c
c
r
e
dit
e
d
F
ir
s
t
G
r
a
de
by
Ke
me
nr
is
tekdikti
,
De
c
r
e
e
No:
21/
E
/KP
T
/2018
DO
I
:
10.
12928/
T
E
L
KO
M
NI
KA
.
v18i3.
14787
1671
Jou
r
n
al
h
omepage
:
ht
tp:
//
jour
nal.
uad
.
ac
.
id/
index
.
php/T
E
L
K
OM
N
I
K
A
D
e
e
p
h
y
p
e
r
sp
h
e
r
e
e
m
b
e
d
d
in
g
f
or
r
e
al
-
t
im
e
f
a
c
e
r
e
c
ogn
itio
n
Ryann
Alim
u
in
,
E
lm
e
r
Dadi
os
,
Jon
at
h
a
n
Dayao
,
S
h
e
ar
yl
Ar
e
n
as
T
ech
n
o
l
o
g
i
ca
l
In
s
t
i
t
u
t
e
o
f
t
h
e
Ph
i
l
i
p
p
i
n
es
-
Q
u
ez
o
n
Ci
t
y
,
D
e
L
a
Sal
l
e
U
n
i
v
er
s
i
t
y
,
Ph
i
l
i
p
p
i
n
es
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
Aug
31
,
2019
R
e
vis
e
d
De
c
7
,
20
19
Ac
c
e
pted
De
c
12
,
20
19
W
i
t
h
t
h
e
ad
v
an
ceme
n
t
o
f
h
u
man
-
c
o
mp
u
t
er
i
n
t
eract
i
o
n
cap
ab
i
l
i
t
i
es
o
f
r
o
b
o
t
s
,
co
mp
u
t
er
v
i
s
i
o
n
s
u
rv
e
i
l
l
an
ce
s
y
s
t
ems
i
n
v
o
l
v
i
n
g
s
ec
u
ri
t
y
y
i
el
d
s
a
l
arg
e
i
mp
ac
t
i
n
t
h
e
res
earc
h
i
n
d
u
s
t
r
y
b
y
h
el
p
i
n
g
i
n
d
i
g
i
t
a
l
i
za
t
i
o
n
o
f
cert
ai
n
s
ecu
r
i
t
y
p
ro
ce
s
s
e
s
.
Reco
g
n
i
zi
n
g
a
face
i
n
t
h
e
co
m
p
u
t
er
v
i
s
i
o
n
i
n
v
o
l
v
es
i
d
en
t
i
f
i
cat
i
o
n
an
d
cl
as
s
i
fi
cat
i
o
n
o
f
w
h
i
c
h
faces
b
el
o
n
g
s
t
o
t
h
e
s
ame
p
ers
o
n
b
y
mean
s
o
f
co
mp
ar
i
n
g
face
emb
ed
d
i
n
g
v
ec
t
o
r
s
.
In
an
o
rg
an
i
zat
i
o
n
t
h
at
h
as
a
l
arg
e
an
d
d
i
v
ers
e
l
ab
e
l
l
e
d
d
at
a
s
et
o
n
a
l
arg
e
n
u
mb
er
o
f
ep
o
c
h
,
o
ft
en
t
i
mes
,
creat
es
a
t
rai
n
i
n
g
d
i
ff
i
cu
l
t
i
es
i
n
v
o
l
v
i
n
g
i
n
c
o
mp
at
i
b
i
l
i
t
y
i
n
d
i
ffe
re
n
t
v
ers
i
o
n
s
o
f
fa
ce
emb
ed
d
i
n
g
t
h
a
t
l
ead
s
t
o
p
o
o
r
face
reco
g
n
i
t
i
o
n
accu
racy
.
In
t
h
i
s
p
a
p
er,
w
e
w
i
l
l
d
e
s
i
g
n
a
n
d
i
mp
l
emen
t
r
o
b
o
t
i
c
v
i
s
i
o
n
s
ec
u
ri
t
y
s
u
r
v
ei
l
l
a
n
ce
s
y
s
t
em
i
n
c
o
rp
o
rat
i
n
g
h
y
b
ri
d
co
mb
i
n
a
t
i
o
n
o
f
MT
CN
N
fo
r
face
d
et
ect
i
o
n
,
an
d
FaceN
e
t
as
t
h
e
u
n
i
f
i
ed
e
mb
e
d
d
i
n
g
fo
r
face
reco
g
n
i
t
i
o
n
an
d
cl
u
s
t
eri
n
g
.
K
e
y
w
o
r
d
s
:
De
e
p
h
ype
r
s
phe
r
e
e
mbedding
R
e
a
l
-
t
im
e
f
ace
d
e
tec
ti
on
R
e
a
l
-
t
im
e
f
ace
r
e
c
ognit
ion
R
oboti
c
v
is
ions
S
e
c
ur
it
y
s
ur
ve
il
lanc
e
s
ys
tem
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
ya
nn
Alim
uin,
T
e
c
hnologi
c
a
l
I
ns
ti
tut
e
of
the
P
hil
ippi
ne
s
-
Que
z
on
C
it
y,
De
L
a
S
a
ll
e
Unive
r
s
it
y
,
M
a
nil
a
,
P
hil
ippi
ne
s
.
E
mail:
r
ya
nn.
a
li
mui
n@ti
p
.
e
du.
ph
1.
I
NT
RODU
C
T
I
ON
W
i
th
t
he
e
me
r
gi
ng
d
e
v
e
l
op
me
nt
o
f
r
o
bo
ts
c
a
pa
bl
e
o
f
hu
ma
n
-
c
o
mp
ut
e
r
in
te
r
a
c
t
i
on
,
r
e
c
og
ni
z
i
ng
pe
o
pl
e
i
n
c
om
pu
te
r
v
is
io
n
a
n
d
p
a
t
te
r
n
r
e
c
o
gn
it
i
on
h
a
s
a
tt
r
a
c
t
e
d
i
m
me
ns
e
a
tt
e
n
ti
on
a
s
i
t
p
r
o
vi
de
s
hu
ge
a
p
p
li
c
a
t
io
ns
in
t
e
r
ms
o
f
f
i
na
n
c
e
,
m
i
li
ta
r
y
,
p
u
b
li
c
s
e
c
u
r
i
ty
a
nd
da
il
y
l
i
f
e
.
Am
on
g
va
r
io
us
b
io
me
t
r
i
c
s
us
e
d
f
o
r
pe
r
s
o
n
r
e
c
o
g
ni
ti
on
,
t
he
f
a
c
e
is
on
e
o
f
t
he
mos
t
p
op
ul
a
r
,
s
in
c
e
t
h
is
ub
i
qu
it
o
us
b
i
ome
t
r
i
c
c
a
n
be
a
c
q
ui
r
e
d
in
u
nc
ons
t
r
a
in
e
d
e
nv
i
r
o
n
men
ts
w
h
il
e
p
r
ov
id
in
g
s
t
r
o
ng
d
is
c
r
i
m
ina
t
ive
f
e
a
tu
r
e
s
f
o
r
r
e
c
o
gn
it
io
n
[
1]
.
O
ve
r
th
e
y
e
a
r
s
,
th
e
r
e
a
r
e
man
y
b
r
e
a
kt
h
r
o
ug
hs
t
ha
t
c
o
nt
r
ib
ut
e
d
to
t
he
s
uc
c
e
s
s
o
f
f
a
c
e
r
e
c
o
gn
it
i
on
t
e
c
h
no
l
og
y
.
T
h
is
is
w
it
h
th
e
h
e
l
p
o
f
a
d
va
n
c
e
d
n
e
two
r
k
a
r
c
h
i
te
c
t
u
r
e
s
[2
-
5
]
,
d
is
c
r
im
i
na
t
i
ve
a
pp
r
oa
c
h
[
2
]
.
F
a
c
e
r
e
c
o
gn
it
io
n
b
e
g
ins
w
it
h
e
xt
r
a
c
t
in
g
the
c
o
o
r
d
in
a
t
e
s
o
f
f
e
a
t
ur
e
s
s
u
c
h
a
s
w
id
th
o
f
m
ou
th
,
wi
dt
h
o
f
e
ye
s
,
p
up
i
l
,
a
nd
c
o
mpa
r
e
t
he
r
e
s
ul
t
w
i
th
the
me
a
s
u
r
e
me
nts
s
to
r
e
d
in
t
he
da
tab
a
s
e
a
nd
r
e
tu
r
n
th
e
c
l
os
e
s
t
r
e
c
o
r
d
(
f
a
c
ia
l
me
tr
i
c
s
)
[
3
]
.
T
he
r
e
ha
v
e
be
e
n
a
h
ug
e
n
um
be
r
o
f
r
e
s
e
a
r
c
h
o
n
wa
ys
o
f
i
mp
r
ov
in
g
th
e
lo
c
a
l
d
e
s
c
r
ip
to
r
s
,
f
e
a
t
u
r
e
t
r
a
ns
f
o
r
m
a
ti
ons
a
n
d
p
r
e
-
p
r
oc
e
s
s
in
g
in
f
a
c
e
r
e
c
o
gn
it
io
n
s
uc
h
a
s
l
in
e
a
r
s
ubs
pa
c
e
[
4]
,
in
ma
ni
f
o
l
ds
[
5
,
6
]
,
a
n
d
s
p
a
r
s
e
r
e
p
r
e
s
e
nt
a
t
io
n
[
5
]
.
B
u
t
t
he
s
e
a
p
pr
oa
c
he
s
ta
r
g
e
ts
on
ly
a
n
a
s
pe
c
t
o
f
c
ons
t
r
a
in
ts
in
f
a
c
ia
l
f
e
a
t
u
r
e
a
n
d
i
mp
r
o
ve
d
f
a
c
e
r
e
c
o
gn
i
t
i
on
a
c
c
u
r
a
c
y
s
l
ow
ly
[
1
]
.
F
u
r
t
he
r
m
o
r
e
,
c
h
a
l
le
nge
s
i
n
t
e
r
ms
o
f
i
ll
um
in
a
t
io
n
,
e
xp
r
e
s
s
i
on
a
nd
pos
e
a
r
e
th
e
t
h
r
e
e
mo
s
t
kn
ow
n
p
r
o
b
lems
i
n
f
a
c
e
r
e
c
o
gn
i
ti
on
(
F
R
)
.
I
n
t
he
r
e
c
e
n
t
ye
a
r
s
,
r
e
s
e
a
r
c
h
la
nds
c
a
pe
in
f
a
c
e
r
e
c
og
n
it
io
n
s
ig
n
if
ic
a
n
tl
y
r
e
s
ha
pe
d
in
to
t
he
b
r
e
a
k
th
r
ou
gh
o
f
de
e
p
le
a
r
n
in
g
s
u
c
h
a
s
de
e
p
f
a
c
e
m
e
t
ho
d
.
D
e
e
p
lea
r
ni
ng
a
p
pl
ies
m
ul
t
ip
le
p
r
oc
e
s
s
i
ng
la
ye
r
s
t
o
le
a
r
n
r
e
p
r
e
s
e
nt
a
t
io
ns
o
f
d
a
t
a
w
i
th
m
ul
ti
p
le
le
ve
ls
of
f
e
a
tu
r
e
e
x
t
r
a
c
t
io
n
[
7]
.
T
h
e
mos
t
po
pu
la
r
de
e
p
lea
r
ni
ng
a
r
c
h
i
tec
tu
r
e
i
s
t
he
c
o
nv
o
lu
ti
on
a
l
n
e
u
r
a
l
n
e
t
wo
r
k
(
C
NN
)
t
ha
t
c
o
mb
a
ts
s
ig
ni
f
ic
a
n
t
p
r
o
bl
e
ms
in
c
om
pu
te
r
v
is
io
ns
s
uc
h
a
s
im
a
ge
c
las
s
i
f
i
c
a
ti
on
,
s
e
gme
n
ta
ti
on
,
o
bj
e
c
t
d
e
t
e
c
ti
on
,
e
tc
.
[
8
]
.
M
a
n
y
f
a
c
e
r
e
c
o
gn
i
ti
on
a
p
p
li
c
a
t
io
ns
s
e
e
k
a
d
e
s
i
r
a
b
le
l
o
w
-
di
me
ns
io
na
l
r
e
p
r
e
s
e
nta
t
io
n
tha
t
ge
ne
r
a
li
z
e
s
we
ll
t
o
ne
w
f
a
c
e
s
th
a
t
t
he
ne
u
r
a
l
n
e
t
wo
r
k
w
a
s
n
’
t
t
r
a
ined
o
n
bu
t
t
he
r
e
p
r
e
s
e
n
ta
ti
on
is
a
c
o
ns
e
q
ue
nc
e
o
f
t
r
a
i
ni
ng
a
ne
t
w
or
k
f
o
r
hi
gh
-
a
c
c
u
r
a
c
y
c
l
a
s
s
i
f
ica
t
io
n
o
n
t
he
i
r
t
r
a
ini
n
g
da
ta
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
18
,
No
.
3
,
J
une
2020:
1671
-
167
7
1672
O
ne
of
th
e
c
ha
l
le
ng
e
s
o
f
t
h
is
k
in
d
o
f
a
pp
r
o
a
c
h
is
t
ha
t
t
he
r
e
p
r
e
s
e
nt
a
t
io
n
i
s
d
if
f
ic
ul
t
t
o
us
e
be
c
a
us
e
f
a
c
e
s
o
f
t
he
s
a
m
e
pe
r
s
on
a
r
e
n
’
t
n
e
c
e
s
s
a
r
i
ly
c
l
us
t
e
r
e
d
,
th
a
t
the
c
l
a
s
s
i
f
ica
t
io
n
a
l
go
r
i
th
ms
c
a
n
t
a
k
e
a
d
va
nt
a
ge
o
f
[
9
]
.
T
h
is
pa
pe
r
d
is
c
us
s
e
d
a
bo
u
t
th
e
c
om
pu
te
r
v
is
ion
of
r
o
bo
ts
i
nv
ol
vi
ng
f
a
c
e
r
e
c
o
gn
it
i
on
p
r
o
c
e
s
s
i
nc
o
r
p
o
r
a
ti
ng
F
a
c
e
N
e
t
a
s
th
e
un
i
f
ie
d
e
m
be
d
d
in
g
f
or
f
a
c
e
r
e
c
o
gn
it
i
on
a
n
d
c
lus
te
r
in
g
th
a
t
le
a
r
ns
ho
w
to
c
lus
te
r
r
e
p
r
e
s
e
nt
a
t
io
ns
o
f
th
e
s
a
me
pe
r
s
o
n
a
nd
th
a
t
c
a
n
a
l
le
vi
a
t
e
t
r
a
i
ni
ng
di
f
f
icu
l
ti
e
s
t
ha
t
c
a
n
s
i
gn
i
f
ica
nt
ly
i
mp
r
ov
e
F
R
a
c
c
u
r
a
c
y
u
ti
li
z
i
ng
P
yt
ho
n
a
s
t
he
pr
og
r
a
m
mi
ng
la
ng
ua
ge
f
or
t
he
s
u
r
ve
i
ll
a
nc
e
s
ys
te
m
.
I
n
t
his
pa
pe
r
,
w
e
p
r
o
p
os
e
d
a
s
ys
t
e
m
a
nd
a
met
ho
d
f
o
r
ta
r
ge
t
i
de
n
t
i
f
ic
a
t
io
n
us
in
g
a
r
ti
f
ic
ia
l
n
e
u
r
a
l
n
e
t
wo
r
ks
i
nt
e
g
r
a
te
d
in
r
o
bo
t
ic
v
is
io
n
.
T
h
e
c
on
t
r
i
bu
ti
ons
o
f
t
hi
s
pa
pe
r
s
um
ma
r
ize
s
a
s
f
ol
l
ows
:
a.
W
e
pr
e
s
e
nt
a
s
e
c
ur
it
y
s
ur
ve
il
lanc
e
s
ys
tem
that
a
uth
e
nti
c
a
tes
a
pe
r
s
on
in
the
r
obo
ti
c
c
a
mer
a
.
b.
A
method
to
p
r
ovide
a
n
e
quivale
nt
vir
tual
ins
tr
u
ment
that
ha
s
the
s
a
me
c
a
pa
bil
it
y
a
nd
f
unc
ti
ona
li
t
y
that
c
ontains
the
f
oll
owing:
(
1
)
a
digi
tal
f
il
ter
that
is
us
e
d
f
or
im
a
ge
p
r
oc
e
s
s
ing
(
2)
a
mac
hine
lea
r
ning
a
lgo
r
it
hm
that
us
e
s
a
r
ti
f
icia
l
ne
ur
a
l
ne
twor
ks
by
mea
ns
of
f
a
c
e
ve
c
tor
identif
ica
ti
on
f
o
r
tar
ge
t
identif
ica
ti
on
.
c.
W
e
uti
li
z
e
s
a
method
ha
ving
a
un
if
ied
f
a
c
e
im
a
ge
r
e
pr
e
s
e
ntation
ne
c
e
s
s
a
r
y
f
or
be
tt
e
r
r
e
c
ognit
ion
of
f
a
c
e
im
a
ge
s
.
d.
T
he
s
ys
tem
that
c
a
n
be
a
da
pted
to
a
ny
e
xis
ti
ng
s
u
r
ve
il
lanc
e
s
ys
tems
,
pr
ovides
low
c
os
t
memor
y
s
to
r
a
ge
,
ha
s
da
ta
loggi
ng
f
e
a
tur
e
s
a
nd
low
maintena
nc
e
.
2.
RE
S
E
AR
CH
M
E
T
HO
D
T
he
input
of
the
s
ys
tem
will
c
a
me
f
r
om
the
wir
e
le
s
s
c
a
mer
a
e
mbedd
e
d
in
R
obots
f
e
e
ds
that
will
be
pr
oc
e
s
s
e
d
a
nd
e
xa
mi
ne
d
by
the
s
ys
tem.
Onc
e
a
f
a
c
e
im
a
ge
is
de
tec
ted
in
the
c
a
mer
a
f
e
e
ds
,
then
i
t
wil
l
de
c
ide
whe
ther
the
f
a
c
e
de
tec
ted
is
r
e
c
ognize
d
or
not
,
if
the
f
a
c
e
is
r
e
c
ognize
d
then
the
s
ys
tem
logs
t
he
da
te,
ti
me
a
nd
the
c
a
mer
a
number
other
wis
e
the
s
ys
t
e
m
s
ti
ll
logs
the
da
te,
ti
me,
a
nd
a
c
ti
va
tes
the
a
lar
m
a
nd
noti
f
ica
ti
on
s
ys
tem.
Algor
it
hm
1
.
Algo
r
it
hm
o
f
the
s
ys
tem
Algorithm of the
s
ystem
Input: Camera’s real
-
time image data
Step 1:
While
≠
face_image
Step 2: Process Video feeds
Step 3:
If
=
face_image,
then
Step 4:
If
=
face_authorized,
then
Step 5: System logs
date, time and cam
era number
Step 6:
Else
Step 7: System logs
date, time and camera number
Step 8: Alarm and Notification is activated
Step 9:
End if
Step 10:
End if
2.
1.
S
ign
al
c
on
d
it
io
n
in
g
F
igur
e
1
s
hows
the
ge
ne
r
a
l
ove
r
view
of
the
pr
opos
e
d
s
y
s
tem.
T
o
a
na
lyze
,
mea
s
ur
e
a
nd
manipulate
da
ta
f
e
e
ds
f
r
om
c
a
mer
a
f
ootage
,
a
na
log
s
ignals
s
hould
be
c
onve
r
ted
int
o
digi
tal
s
ignal
uti
li
z
ing
the
t
he
or
y
of
digi
tal
s
ignal
p
r
oc
e
s
s
ing.
Ana
log
to
digi
tal
c
onve
r
t
e
r
(
AD
C
)
is
the
one
r
e
s
pons
ibl
e
in
s
a
mpl
ing,
qua
nti
z
ing
a
nd
e
nc
oding
the
c
onti
nuous
-
a
mpl
it
ude
a
na
log
s
ignal
i
nto
dis
c
r
e
te
ti
me
a
nd
a
mpl
it
ude
digi
tal
s
ignal.
(
)
=
∑
[
(
−
)
−
(
−
−
)
]
∞
=
−
∞
(
1)
A
nu
mb
e
r
o
f
va
r
iab
le
bi
t
-
r
a
t
e
da
ta
s
tr
e
a
ms
o
f
i
np
ut
s
i
g
na
ls
f
r
om
d
i
f
f
e
r
e
nt
w
i
r
e
les
s
c
a
me
r
a
s
wi
l
l
be
in
teg
r
a
te
d
i
nt
o
a
c
o
ns
tan
t
c
a
p
a
c
it
y
s
ig
na
l
th
r
o
ug
h
ti
me
d
iv
is
io
n
m
ul
t
ip
le
xe
r
(
T
DM
)
us
e
d
f
or
a
h
ig
he
r
bi
t
-
r
a
t
e
f
l
ow
o
f
da
ta
[
10
]
.
S
u
bs
e
que
nt
l
y
,
t
he
s
ig
na
ls
we
r
e
be
in
g
d
i
gi
ta
ll
y
f
i
lt
e
r
e
d
t
h
r
o
ug
h
d
ig
it
a
l
s
ig
na
l
p
r
oc
e
s
s
i
ng
(
D
S
P
)
t
o
p
r
oc
e
s
s
t
he
i
mag
e
f
o
r
the
in
te
g
r
a
ti
on
o
f
f
a
c
e
de
tec
t
io
n
a
n
d
da
ta
lo
gg
in
g
tec
hn
ol
og
y
us
in
g
a
r
ti
f
ic
ia
l
ne
u
r
a
l
ne
tw
o
r
k
i
ng
.
2.
2.
I
m
age
p
r
oc
e
s
s
in
g
M
ult
it
hr
e
a
d
ing
a
nd
GPU
ba
s
e
d
pr
oc
e
s
s
ing
tec
hnol
ogies
we
r
e
us
e
d
to
pe
r
f
or
m
the
im
a
ge
pr
oc
e
s
s
ing.
Ar
c
hit
e
c
tur
e
of
the
im
a
ge
pr
oc
e
s
s
ing
in
thi
s
r
e
s
e
a
r
c
h
is
s
hown
in
F
igur
e
2.
De
tailed
pr
oc
e
s
s
ing
will
be
e
xplaine
d
in
the
be
low
s
e
c
ti
on.
2.
2.
1.
M
u
lt
i
-
t
as
k
c
as
c
ad
e
d
CN
N
I
n
f
a
c
e
de
tec
ti
on
pha
s
e
,
our
method
is
ba
s
e
d
on
mul
ti
-
tas
k
c
a
s
c
a
de
d
C
NN
u
s
e
d
f
or
joi
nt
f
a
c
e
de
tec
ti
on
a
nd
f
a
c
e
a
li
gnment
[
11]
in
de
tec
ti
ng
f
a
c
e
s
withi
n
t
he
vicinit
y
of
c
a
mer
a
f
ootage
in
r
e
a
l
-
ti
me.
I
t
ini
ti
a
l
ly
r
e
s
ize
s
the
im
a
ge
s
int
o
a
di
f
f
e
r
e
nt
s
c
a
le
buil
ding
a
n
im
a
ge
pyr
a
mi
d.
T
he
pr
oc
e
s
s
c
a
me
with
3
s
tage
s
M
T
C
NN
na
mely:
pr
opos
a
l
ne
twor
k
(
P
Ne
t)
us
e
d
to
obtain
c
a
ndidate
f
a
c
ial
windows
,
a
s
we
ll
a
s
their
bounding
box
r
e
gr
e
s
s
ion
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
De
e
p
hy
pe
r
s
phe
r
e
e
mbe
dding
for
r
e
al
-
ti
me
face
r
e
c
ognit
ion
(
R
y
ann
A
li
muin
)
1673
ve
c
tor
s
.
R
e
f
ine
ne
twor
k
(
R
Ne
t)
that
r
e
f
ines
huge
a
mount
of
f
a
ls
e
c
a
ndidate
a
s
we
ll
a
s
pe
r
f
o
r
ms
c
a
li
br
a
ti
on
with
bounding
box
r
e
gr
e
s
s
ion,
a
nd
c
onduc
ts
NM
S
a
a
nd
las
tl
y,
output
ne
twor
k
(
O
-
Ne
t)
that
is
us
e
d
to
pr
oduc
e
the
f
inal
huge
box
a
nd
f
a
c
ial
landma
r
ks
pos
it
ion
,
r
e
s
pe
c
ti
ve
ly
[
12]
.
T
his
s
tage
a
im
s
to
identif
y
f
a
c
e
r
e
gions
with
mor
e
s
up
e
r
vis
ion
Additi
ona
ll
y
,
M
T
C
NN
us
e
s
a
c
ompl
e
x
a
lgor
it
hm
in
mul
ti
p
le
thr
e
a
ds
whe
r
e
i
n
it
c
a
n
de
tec
t
f
a
c
e
s
e
f
f
e
c
ti
ve
ly
e
ve
n
in
r
a
nge
s
of
dis
tanc
e
f
r
om
the
c
a
mer
a
that
make
s
it
a
good
f
it
f
or
ou
r
a
pp
li
c
a
ti
on.
a.
T
r
a
ini
ng
I
n
tr
a
ini
ng
f
o
r
the
C
NN
de
tec
tor
,
i
t
leve
r
a
g
e
s
thr
e
e
tas
ks
a
s
f
oll
ows
.
-
F
a
c
e
C
las
s
if
ica
ti
on.
I
t
uti
l
ize
s
the
c
r
os
s
-
e
ntr
opy
los
s
in
e
a
c
h
s
a
mpl
e
x
i
.
=
−
(
log
(
)
+
(
1
−
)
(
1
−
(
)
)
)
(
2)
-
B
ounding
box
r
e
gr
e
s
s
ion.
T
he
lea
r
ning
objec
ti
ve
is
f
or
mul
a
ted
a
s
a
r
e
gr
e
s
s
ion
pr
oblem
a
nd
the
method
uti
li
z
e
s
the
E
uc
li
de
a
n
los
s
f
or
e
a
c
h
s
a
mpl
e
x
i
.
=
‖
−
‖
2
2
(
3)
-
F
a
c
ial
landma
r
k
loca
li
z
a
ti
on.
T
he
s
a
me
with
the
bo
unding
box,
the
f
a
c
ial
landma
r
k
de
tec
ti
on
is
f
or
mul
a
ted
a
s
a
r
e
gr
e
s
s
ion
pr
oblem
a
nd
ut
il
ize
s
mi
nim
ize
s
the
E
uc
li
dian
los
s
.
=
‖
−
‖
2
2
(
4)
F
igur
e
1
.
Ge
ne
r
a
l
o
ve
r
view
of
the
s
ys
tem
F
igur
e
2
.
I
mage
p
r
oc
e
s
s
ing
a
r
c
hit
e
c
tur
e
2.
2.
2.
F
ac
e
r
e
c
ogn
it
ion
T
he
a
bove
-
mentioned
pr
oblem
invol
ving
invar
ianc
e
of
f
a
c
e
r
e
pr
e
s
e
ntation
ove
r
a
pe
r
iod
o
f
ti
me
,
c
a
n
be
s
olved
by
the
noti
ons
of
f
indi
ng
pa
c
king
a
s
ympt
oti
c
bounds
,
that
a
r
e
not
ove
r
lapping,
f
or
whic
h
it
c
a
n
be
f
it
withi
n
a
f
a
c
e
r
e
pr
e
s
e
ntation
s
pa
c
e
or
hype
r
s
phe
r
e
.
A
r
e
pr
e
s
e
nta
ti
on
of
the
ge
ometr
ica
l
s
tr
uc
tur
e
c
a
n
b
e
c
a
n
be
de
s
c
r
ibe
whe
r
e
in
the
lowe
r
bound
r
e
pr
e
s
e
nts
the
low
-
dim
e
ns
ional
population
manif
old
e
mbedde
d
in
a
high
dim
e
ns
ional
s
pa
c
e
loca
ted
on
the
uppe
r
bound
hy
pe
r
-
e
ll
ips
oid
that
is
c
lus
ter
e
d
int
o
thei
r
own
c
las
s
s
pe
c
i
f
ic
hype
r
-
e
ll
ips
oids
[
13]
.
T
he
invar
ianc
e
o
f
the
f
a
c
e
r
e
pr
e
s
e
ntation
is
de
ter
mi
ne
d
by
the
number
of
identi
ti
e
s
that
is
pa
c
ke
d
pe
r
hype
r
-
e
ll
ips
oid.
a.
F
a
c
e
Ne
t
I
n
thi
s
pa
pe
r
,
we
int
e
gr
a
ted
a
method
c
a
ll
e
d
F
a
c
e
Ne
t
f
or
the
f
a
c
e
r
e
c
ognit
ion
pha
s
e
.
F
a
c
e
Ne
t
is
a
un
if
ied
e
mbedding
f
o
r
f
a
c
e
r
e
c
ognit
ion
a
nd
c
lu
s
ter
ing
that
dir
e
c
tl
y
lea
r
ns
a
mapping
f
r
om
f
a
c
e
i
mage
s
to
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
18
,
No
.
3
,
J
une
2020:
1671
-
167
7
1674
a
c
ompac
t
E
uc
li
de
a
n
s
pa
c
e
whe
r
e
dis
tanc
e
s
dir
e
c
tl
y
c
or
r
e
s
pond
to
a
mea
s
ur
e
of
f
a
c
e
s
im
il
a
r
it
y
[
1
4]
us
ing
the
tr
ipl
e
t
los
s
f
unc
ti
on
ba
s
e
d
on
L
M
NN
.
b.
T
r
ipl
e
t
los
s
f
unc
ti
on
T
he
s
ys
tem
us
e
d
F
a
c
e
Ne
t
to
map
f
a
c
e
f
e
a
tur
e
s
f
r
o
m
the
input
im
a
ge
s
take
n
f
r
om
the
c
a
mer
a
s
int
o
a
512
-
dim
e
ns
ional
E
uc
li
dian
s
pa
c
e
ve
c
tor
[
15]
.
T
he
e
mbed
ding
is
r
e
pr
e
s
e
nted
by
(
5)
,
ha
ving
a
n
input
im
a
ge
x
e
mbedde
d
int
o
d
-
dim
e
ns
ional
E
uc
li
dia
n
s
pa
c
e
ve
c
tor
ℝ
a
nd
is
c
ons
tr
a
int
s
in
(
6)
[
11
]
.
(
)
∈
ℝ
(
5)
|
|
(
)
|
|
2
=
1
(
6)
F
ur
ther
mor
e
,
the
s
ys
tem
uti
li
z
e
s
tr
ipl
e
t
l
os
s
us
e
d
to
e
f
f
e
c
ti
ve
ly
c
lus
ter
f
a
c
e
ve
c
tor
e
mbedding
be
twe
e
n
t
he
s
a
me
c
las
s
e
s
,
a
nd
is
r
e
pr
e
s
e
nted
by
(
7)
[
16]
.
∑
[
|
|
(
)
−
(
)
|
|
2
2
−
|
|
(
)
−
(
)
|
|
2
2
+
]
(
7)
w
he
r
e
r
e
pr
e
s
e
nts
a
n
a
nc
hor
of
a
s
pe
c
if
ic
pe
r
s
on,
i
ndica
tes
pos
it
ive
r
e
pr
e
s
e
ntation
of
the
s
a
me
pe
r
s
on
.
is
the
ne
ga
ti
ve
r
e
pr
e
s
e
ntation
o
f
a
ny
othe
r
pe
r
s
on
a
nd
be
ing
the
mar
gin
be
twe
e
n
the
pos
it
ive
a
nd
n
e
ga
ti
ve
pa
ir
s
.
T
he
tr
ipl
e
t
los
s
mi
nim
ize
s
t
he
s
qua
r
e
of
the
di
s
tanc
e
be
twe
e
n
the
a
nc
hor
a
nd
a
pos
it
ive
while
max
im
izing
the
s
qua
r
e
of
the
dis
tanc
e
be
twe
e
n
the
a
nc
hor
a
nd
ne
ga
ti
ve
pa
ir
s
[
17]
.
c.
Ha
r
moni
c
e
mbedding
a
nd
t
r
ipl
e
t
los
s
T
he
s
ys
tem
a
ls
o
pr
ovides
a
powe
r
f
ul
f
unc
ti
on
whe
r
e
in
it
ha
s
the
c
a
pa
bil
it
y
to
c
ompar
e
a
ne
w
tr
a
ini
ng
da
tas
e
t
to
the
e
xis
ti
ng
da
tas
e
ts
in
the
ga
ll
e
r
y.
T
his
f
unc
ti
on
is
idea
l
f
o
r
la
r
ge
s
c
a
le
da
tas
e
ts
that
is
di
ve
r
ge
nt,
a
nd
r
e
quir
e
s
r
e
tr
a
ini
ng
the
s
ubjec
t
mul
ti
ple
ti
mes
.
2.
3.
S
e
r
ial
c
om
m
u
n
icat
ion
A
USB
to
s
e
r
ial
a
da
pter
a
ls
o
r
e
f
e
r
r
e
d
to
a
s
a
US
B
s
e
r
ial
c
onve
r
ter
or
R
S
232
a
da
pter
wa
s
us
e
d
f
or
s
e
r
ial
c
omm
unica
ti
on
a
s
the
int
e
r
f
a
c
e
f
r
om
the
c
a
mer
a
int
o
the
c
omput
e
r
[
18
]
.
I
t
is
a
s
mall
e
lec
tr
on
ic
de
vice
whic
h
c
a
n
c
onve
r
t
a
USB
s
ignal
to
s
e
r
ial
R
S
232
da
ta
s
ignals
[
19]
.
I
t
is
t
he
type
o
f
s
ignal
whic
h
is
in
many
older
P
C
s
a
nd
is
r
e
f
e
r
r
e
d
to
a
s
a
s
e
r
ial
C
OM
por
t
.
A
USB
to
s
e
r
ial
a
da
pter
typ
ica
ll
y
c
onve
r
ts
be
twe
e
n
USB
a
nd
e
it
he
r
R
S
232,
R
S
485,
R
S
422
or
T
C
P
s
ignals
,
h
owe
ve
r
s
ome
USB
to
s
e
r
ial
a
da
pter
s
ha
ve
other
s
pe
c
ial
c
onve
r
s
ion
f
e
a
tur
e
s
s
uc
h
a
s
c
us
tom
ba
ud
r
a
tes
,
hig
h
-
s
pe
e
d
or
other
[
20,
21]
.
3.
I
M
P
L
E
M
E
NT
AT
I
ON
T
he
f
oll
owing
a
r
e
the
c
omponents
of
the
whole
s
y
s
tem
a.
W
ir
e
les
s
c
a
mer
a
b.
NVR
c.
USB
s
e
r
ial
c
onve
r
ter
o
r
R
S
232
d.
L
a
ptop
C
omput
e
r
T
his
r
e
s
e
a
r
c
h
us
e
s
c
a
mer
a
s
pe
c
if
ica
ti
ons
a
s
s
hown
in
T
a
b
le
1
,
a
nd
us
e
s
a
c
omput
e
r
a
s
s
hown
in
T
a
ble
2
.
T
a
ble
1.
C
a
mer
a
s
pe
c
if
ica
ti
ons
S
pe
c
if
ic
a
ti
ons
V
a
lu
e
I
ma
ge
S
e
ns
or
½.8’
2.4 M
P
C
M
O
S
E
f
f
e
c
ti
ve
P
ix
e
ls
1984 (
H
)
x 1225 (
V
)
E
le
c
tr
oni
c
S
hut
te
r
1/
3s
–
1/
100,000s
M
in
im
um I
ll
umi
na
ti
on
0.05 lux/
F
1.4, lux I
R
on
T
a
ble
2.
L
a
ptop
c
omput
e
r
s
pe
c
s
S
pe
c
if
ic
a
ti
ons
V
a
lu
e
P
r
oc
e
s
s
or
A
t
le
a
s
t
4G
H
z
O
pe
r
a
ti
ng S
ys
te
m
W
in
dow
s
X
P
or
l
a
te
r
I
nt
e
r
na
l
S
to
r
a
ge
M
in
im
um of
1T
B
R
a
ndom Ac
c
e
s
s
M
e
mor
y (
R
A
M
)
A
t
le
a
s
t
8 G
B
4.
T
E
CHNI
CA
L
RE
S
UL
T
I
n
thi
s
s
e
c
ti
on
we
will
e
va
luate
the
e
f
f
e
c
ti
ve
ne
s
s
a
nd
the
pe
r
f
o
r
manc
e
of
the
pr
opos
e
d
s
ys
tem.
4.
1.
Graph
ic
u
s
e
r
in
t
e
r
f
ac
e
(
GUI
)
T
he
us
e
r
int
e
r
f
a
c
e
of
the
s
ys
tem
indi
c
a
tes
the
po
r
ti
on
whe
r
e
the
4
c
a
mer
a
s
will
be
s
hown.
I
t
a
ls
o
dis
plays
pr
e
view
,
da
taba
s
e
,
c
loud
s
tor
a
ge
,
loca
l
s
t
or
a
ge
,
a
bout
a
nd
las
tl
y,
da
ta
loggi
ng
.
I
n
the
s
e
c
ti
on
of
da
ta
loggi
ng,
it
dis
plays
the
de
tec
ti
on
of
the
c
a
mer
a
s
whe
r
e
the
r
e
c
ognize
d
f
a
c
e
s
a
r
e
a
uthor
ize
d
or
una
uthor
ize
d.
I
t
a
ls
o
indi
c
a
tes
the
identit
y
of
the
de
tec
ted
pe
r
s
on.
S
a
mpl
e
12x12
pixel
of
f
a
c
e
da
tas
e
ts
a
s
s
hown
in
F
igur
e
3.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
De
e
p
hy
pe
r
s
phe
r
e
e
mbe
dding
for
r
e
al
-
ti
me
face
r
e
c
ognit
ion
(
R
y
ann
A
li
muin
)
1675
F
igur
e
3
.
S
a
mpl
e
12x12
pixel
of
f
a
c
e
da
tas
e
ts
4.
2.
T
r
ain
in
g
Dat
a
T
he
tr
a
ini
ng
s
e
t
us
e
d
in
the
e
xpe
r
im
e
nt
a
r
e
f
a
c
e
im
a
ge
s
take
n
f
r
om
s
a
mpl
e
s
pe
c
im
e
ns
,
whe
r
e
f
a
c
ial
f
e
a
tur
e
s
will
be
take
n
mul
ti
ple
ti
mes
.
T
o
m
ini
mi
z
e
f
a
c
e
va
r
iations
,
e
a
c
h
wil
l
be
take
n
without
e
x
pr
e
s
s
ion,
a
nd
then
a
s
ke
d
to
ti
lt
thei
r
f
a
c
e
s
to
the
le
f
t
a
nd
s
lowly
to
the
r
ight
,
a
nd
move
thei
r
f
a
c
e
s
s
lowly
upwa
r
d
a
nd
downw
a
r
d
pos
it
ion.
T
he
r
e
s
ult
of
the
t
r
a
ini
ng
s
e
t
a
c
quis
it
ion
pr
oc
e
s
s
will
pr
oduc
e
200
s
e
ts
of
12x12
pixela
ted
f
a
c
e
im
a
ge
s
f
or
the
tr
a
ini
ng
s
e
t
pe
r
pe
r
s
on
a
nd
a
ddi
ti
ona
l
10
im
a
ge
s
f
o
r
the
s
a
mpl
e
s
ne
c
e
s
s
a
r
y
f
or
the
tes
ti
ng.
4.
3.
De
t
e
c
t
ion
,
E
xt
r
ac
t
io
n
ad
Re
c
ogn
it
ion
t
i
m
e
De
tec
ti
on
is
the
pr
oc
e
s
s
wh
e
r
e
in
the
s
ys
tem
s
e
a
r
c
he
s
f
or
the
f
a
c
e
s
withi
n
a
n
im
a
ge
a
nd
r
e
tur
ns
it
s
c
oor
dinate
s
[
22]
.
E
xtr
a
c
ti
on
,
on
the
other
ha
nd
is
th
e
pr
oc
e
s
s
whe
r
e
the
s
ys
tem
f
il
ter
s
the
de
tec
ted
f
a
c
e
to
f
il
ter
out
the
unne
c
e
s
s
a
r
y
de
tails
of
the
i
mage
[
23]
.
L
a
s
tl
y,
r
e
c
ognit
ion
is
the
p
r
oc
e
s
s
of
the
s
ys
tem
whe
r
e
it
i
de
nti
f
ies
the
de
tec
ted
f
a
c
e
ba
s
e
d
f
r
om
the
da
tas
e
ts
[
24]
.
Af
ter
10
it
e
r
a
ti
ons
,
we
we
r
e
a
ble
to
mea
s
ur
e
the
de
tec
ti
on,
e
xtr
a
c
ti
ons
a
nd
r
e
c
ognit
ion
ti
me.
T
he
a
ve
r
a
ge
d
e
tec
ti
on
ti
me
is
100.
8ms
,
e
xtr
a
c
ti
on
ti
me
is
91.
7
ms
a
nd
r
e
c
ognit
ion
ti
me
is
0/8
ms
.
4.
4.
P
e
r
c
e
n
t
ac
c
u
r
ac
y
p
e
r
p
e
r
s
on
F
igur
e
4
s
hows
the
r
e
s
ult
of
the
a
c
c
ur
a
c
y
of
the
s
y
s
tem
whe
r
e
in
5
dif
f
e
r
e
nt
pe
ople
we
r
e
tes
ted
one
a
t
a
ti
me.
T
he
s
ys
tem
s
hows
highl
y
a
c
c
ur
a
te
c
las
s
if
ica
ti
on
ha
ving
a
n
a
ve
r
a
ge
o
f
86
%
.
F
igur
e
4
.
De
tec
ti
on,
e
xtr
a
c
ti
on
a
nd
r
e
c
ognit
ion
t
im
e
(
ms
)
4.
5.
E
valu
at
io
n
of
t
h
e
s
ys
t
e
m
s
p
e
r
f
or
m
an
c
e
W
e
tes
ted
the
s
ys
tem’
s
pe
r
f
o
r
manc
e
a
nd
li
mi
ta
ti
ons
by
va
r
ying
dis
tanc
e
of
the
s
ubjec
ts
f
r
om
the
c
a
mer
a
that
r
a
nge
s
f
r
om
1
-
7f
t
incr
e
mente
d
wit
h
1f
oot
a
nd
a
t
the
s
a
me
ti
me
va
r
ying
the
number
o
f
pe
ople
be
ing
r
e
c
ogn
ize
d
s
im
ult
a
ne
ous
ly.
T
a
ble
3
s
hows
t
he
numer
ica
l
va
lue
o
f
the
a
c
c
ur
a
c
y.
T
he
e
xpe
r
im
e
nt
s
hows
that
a
dding
the
number
of
s
ubjec
ts
be
ing
r
e
c
ognize
d
by
the
s
ys
tem
s
im
ult
a
ne
ous
ly
c
a
n
g
r
e
a
t
ly
a
f
f
e
c
t
the
pe
r
f
or
manc
e
of
the
s
ys
tem
while
incr
e
a
s
ing
th
e
dis
tanc
e
of
the
s
ubjec
t
f
r
om
the
c
a
mer
a
c
r
e
a
tes
a
mi
nim
a
l
e
f
f
e
c
t
to
the
pe
r
f
or
manc
e
of
the
s
ys
tem.
F
igur
e
5
s
hows
the
a
ve
r
a
ge
a
c
c
ur
a
c
y
of
50%
f
r
o
m
the
mul
ti
ple
f
a
c
e
s
with
va
r
ied
dis
tanc
e
.
F
igur
e
5
.
Ac
c
ur
a
c
y
pe
r
pe
r
s
on
(
ms
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
18
,
No
.
3
,
J
une
2020:
1671
-
167
7
1676
T
a
ble
3.
Ac
c
ur
a
c
y
t
e
s
ti
ng
D
is
ta
nc
e
(
f
t)
1 P
e
r
s
on
2 P
e
r
s
ons
3 P
e
r
s
ons
4 P
e
r
s
ons
5 P
e
r
s
ons
A
ve
r
a
ge
1
100
27
44.44
50
40
52.44
2
90
50
44.44
25
40
49.88
3
90
40
55.55
25
40
50.11
4
80
50
22.22
25
40
43.44
5
90
60
33.33
50
50
56.66
6
80
60
44.44
37.5
40
52.38
7
70
40
22.22
37.5
40
41.94
4.
6.
Conf
u
s
ion
m
a
t
r
ix
T
o
e
va
luate
the
pe
r
f
or
manc
e
of
the
c
las
s
if
ier
,
c
on
f
us
ion
matr
ix
wa
s
us
e
d.
I
t
s
hows
a
vis
ua
li
z
a
ti
on
in
whic
h
the
c
las
s
if
ier
is
c
onf
us
e
d
whe
n
making
a
pr
e
diction
in
de
a
li
ng
with
a
dding
tr
a
ini
ng
s
ubjec
ts
to
the
da
ta
s
e
t.
F
igur
e
6
s
hows
the
r
e
s
ult
of
c
las
s
if
ier
’
s
tr
ue
po
s
it
ive
r
a
te
a
nd
mi
s
c
las
s
if
ica
ti
on
r
a
te
by
mea
ns
of
divi
ding
tempor
a
r
il
y
the
da
tas
e
t
whic
h
is
c
ompos
e
d
of
33
%
tes
t
s
e
ts
a
nd
67%
o
f
t
r
a
in
s
e
ts
[
25]
.
I
t
s
hows
n
or
malize
d
c
onf
us
ion
matr
ix
f
r
om
3
a
nd
4
pe
r
s
ons
known,
r
e
s
pe
c
ti
ve
ly.
(
a
)
(
b)
F
igur
e
6
.
C
onf
us
ion
m
a
tr
ice
s
with
,
(
a
)
3
pe
r
s
ons
,
(
b)
4
pe
r
s
ons
(
n
or
malize
d)
4.
7.
Com
p
ar
is
on
s
of
t
h
e
s
ys
t
e
m
s
p
e
r
f
or
m
an
c
e
wit
h
ot
h
e
r
d
e
e
p
lear
n
in
g
algorit
h
m
Va
r
ying
the
a
lgor
it
hm
on
the
s
ys
tem
pr
ovides
mi
n
im
a
l
va
r
iations
to
the
pe
r
f
or
manc
e
o
f
the
s
ys
tem.
F
igur
e
7
s
hows
the
s
ys
tem’
s
pe
r
f
or
manc
e
in
ter
ms
of
it
s
a
c
c
ur
a
c
y,
s
e
ns
it
ivi
ty
a
nd
s
pe
c
if
ici
ty
us
ing
the
De
e
pF
a
c
e
,
S
phe
r
e
F
a
c
e
a
nd
M
T
C
NN
.
T
he
M
T
C
NN
a
nd
F
a
c
e
Ne
t
a
da
pts
to
the
s
ys
tems
pe
r
f
or
manc
e
by
ha
ving
a
s
tand
out
r
e
s
ult
in
c
ompar
e
d
with
the
two
other
a
lgo
r
it
hm.
F
igur
e
7
.
S
ys
tem's
p
e
r
f
o
r
manc
e
t
e
s
t
with
othe
r
de
e
p
lea
r
ning
a
lgor
it
hm
5.
CONC
L
USI
ON
I
d
e
n
ti
f
yi
ng
a
p
e
r
s
o
n
f
r
o
m
a
s
u
r
v
e
i
l
lan
c
e
s
ys
te
m
e
m
be
dd
e
d
i
n
r
ob
o
ts
o
f
f
e
r
s
s
ig
ni
f
ic
a
n
t
a
d
va
n
ta
ge
s
i
n
t
e
r
ms
o
f
s
e
c
u
r
i
ty
i
n
c
om
pa
r
e
d
wi
th
t
he
t
r
a
di
t
io
na
l
s
u
r
ve
il
la
nc
e
s
ys
te
m
.
I
t
c
a
n
s
a
v
e
hu
ge
a
m
ou
nt
s
t
o
r
a
ge
a
n
d
i
ts
c
o
r
r
e
s
p
on
di
ng
c
os
ts
b
y
o
nl
y
s
t
o
r
i
ng
f
r
a
mes
o
f
f
a
c
e
im
a
g
e
s
t
ha
t
wa
s
de
tec
te
d
by
th
e
s
ys
te
m
.
I
t
o
f
f
e
r
s
mor
e
s
e
c
u
r
e
T
rue
l
a
bel
P
re
di
ct
ed l
a
bel
T
rue
l
a
bel
P
re
di
ct
ed l
a
bel
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
De
e
p
hy
pe
r
s
phe
r
e
e
mbe
dding
for
r
e
al
-
ti
me
face
r
e
c
ognit
ion
(
R
y
ann
A
li
muin
)
1677
e
nv
i
r
o
nm
e
n
t
,
a
s
it
w
i
ll
s
e
n
d
a
l
e
r
ts
r
e
ga
r
d
i
ng
b
u
r
g
lar
y
th
a
t
is
h
a
p
pe
n
in
g
in
r
e
a
l
ti
me
.
I
t
s
e
r
ve
s
a
s
b
io
me
t
r
i
c
s
o
f
a
p
e
r
s
o
n’
s
ide
nt
i
ty
l
og
gi
ng
in
a
n
d
ou
t
f
r
om
a
n
e
s
tab
l
is
h
me
nt
a
nd
c
a
n
be
ve
r
y
us
e
f
u
l
i
n
lo
c
a
ti
ng
a
n
d
i
de
nt
i
f
y
in
g
c
r
im
in
a
ls
a
r
o
un
d
th
e
c
it
y
.
T
h
e
e
xp
e
r
i
me
nts
un
c
o
ve
r
t
h
e
s
ys
te
m
’
s
li
m
it
a
t
io
n
in
de
tec
t
in
g
a
nd
i
de
n
t
if
yi
ng
m
u
lt
ip
le
p
e
r
s
o
n
a
t
a
s
pe
c
i
f
ie
d
d
is
t
a
n
c
e
s
im
ul
ta
ne
ous
ly
.
T
he
s
ys
t
e
m
r
e
s
u
lt
e
d
to
o
nl
y
5
0
%
o
f
th
e
a
ve
r
a
ge
a
c
c
u
r
a
c
y
w
he
n
d
e
a
li
ng
wi
th
mu
l
ti
pl
e
pe
r
s
on
in
c
o
mpa
r
e
d
w
i
th
8
6
%
i
n
a
c
c
u
r
a
te
l
y
i
de
n
t
if
y
in
g
d
if
f
e
r
e
n
t
f
a
c
e
s
a
t
a
t
im
e
.
F
o
r
f
ut
u
r
e
d
e
s
i
gn
e
r
s
w
ho
wi
s
he
s
t
o
v
e
n
tu
r
e
i
nt
o
th
is
s
tu
dy
.
W
e
h
ig
h
ly
r
e
c
o
mm
e
n
d
to
f
ur
t
he
r
im
p
r
o
ve
the
pe
r
f
o
r
ma
nc
e
o
f
t
he
s
ys
te
m’
s
a
c
c
u
r
a
c
y
b
y
tr
y
in
g
o
th
e
r
ty
pe
s
o
f
a
lg
o
r
i
th
m
kn
ow
in
g
th
a
t
t
he
r
e
a
r
e
a
lo
t
o
f
o
p
ti
ons
.
W
e
a
ls
o
r
e
c
o
mm
e
nd
t
o
us
e
be
tt
e
r
s
pe
c
i
f
ica
t
io
ns
o
f
c
a
m
e
r
a
a
n
d
c
om
pu
te
r
me
nt
io
ne
d
i
n
s
e
c
t
io
n
3
.
AC
KNOWL
E
DGM
E
N
T
T
he
pr
opone
nts
would
li
ke
to
thank
De
L
a
S
a
ll
e
Unive
r
s
it
y
-
M
a
nil
a
a
nd
T
e
c
hnologi
c
a
l
I
ns
ti
tut
e
of
the
P
hil
ippi
ne
s
-
Que
z
on
C
it
y
f
or
the
r
e
s
e
a
r
c
h
c
ol
labor
a
ti
on
a
nd
to
the
f
oll
owing
s
tudents
na
mely
R
.
E
nor
me,
L
.
E
s
tr
a
da
,
C
.
M
ontea
lt
o,
J
.
Or
e
ta
.
RE
F
E
RE
NC
E
S
[1
]
I.
Mas
i
,
et
a
l
.
,
“
D
ee
p
Face
Rec
o
g
n
i
t
i
o
n
:
a
Su
r
v
ey
,
”
2
0
1
8
3
1
s
t
S
IB
G
R
A
P
I
C
o
n
f
er
en
ce
o
n
G
r
a
p
h
i
c
s
,
P
a
t
t
er
n
s
a
n
d
Im
a
g
es
(S
I
B
G
R
A
P
I)
,
p
p
.
4
7
1
-
4
7
8
,
2
0
1
8
.
[2
]
Y
.
Su
n
,
et
al
.
,
“
D
eep
l
earn
i
n
g
face
rep
res
e
n
t
a
t
i
o
n
b
y
j
o
i
n
t
i
d
e
n
t
i
fi
ca
t
i
o
n
-
v
eri
f
i
ca
t
i
o
n
,
”
A
d
v
a
n
ce
s
i
n
Ne
u
r
a
l
In
f
o
r
m
a
t
i
o
n
P
r
o
ce
s
s
i
n
g
S
y
s
t
e
m
2
7
(
NI
P
S
2
0
1
4
)
,
p
p
.
1
9
8
8
-
1
9
9
6
,
2
0
1
4
.
[3
]
A
.
R.
S.
Si
s
w
an
t
o
,
et
al
.
,
“
Imp
l
emen
t
at
i
o
n
o
f
face
re
co
g
n
i
t
i
o
n
al
g
o
ri
t
h
m
fo
r
b
i
o
met
r
i
cs
-
b
as
e
d
t
i
me
at
t
en
d
a
n
ce
s
y
s
t
em,
”
2
0
1
4
In
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
e
r
en
ce
o
n
ICT
f
o
r
S
m
a
r
t
S
o
ci
e
t
y
(ICIS
S
)
,
p
p
.
1
4
9
-
1
5
4
,
2
0
1
4
.
[4
]
W
.
D
en
g
,
et
a
l
.
,
“
T
ran
s
fo
rm
-
I
n
v
ar
i
a
n
t
PC
A
:
A
U
n
i
f
i
ed
A
p
p
r
o
ach
t
o
F
u
l
l
y
A
u
t
o
ma
t
i
c
Face
A
l
i
g
n
men
t
,
R
ep
re
s
en
t
at
i
o
n
,
an
d
Reco
g
n
i
t
i
o
n
,
”
IE
E
E
Tr
a
n
s
a
c
t
i
o
n
s
o
n
P
a
t
t
e
r
n
A
n
a
l
y
s
i
s
a
n
d
M
a
ch
i
n
e
In
t
el
l
i
g
en
ce
,
v
o
l
.
3
6
,
n
o
.
6
,
p
p
.
1
2
7
5
-
1
2
8
4
,
Jun
e
2
0
1
4
.
[5
]
J
.
W
ri
g
h
t
,
et
a
l
.
,
“
Ro
b
u
s
t
Face
v
i
a
S
p
ar
s
e
Rep
re
s
en
t
at
i
o
n
,
”
IE
E
E
T
r
a
n
s
a
ct
i
o
n
o
n
P
a
t
t
er
n
A
n
a
l
ys
i
s
a
n
d
M
a
ch
i
n
e
In
t
e
l
l
i
g
e
n
t
,
v
o
l
.
3
1
,
n
o
.
2
,
p
p
.
2
1
0
-
2
2
7
,
Feb
r
u
ary
2
0
0
9
.
[6
]
E
.
K
l
arre
i
ch
,
“
S
p
h
ere
Pac
k
i
n
g
S
o
l
v
ed
i
n
H
i
g
h
e
r
D
i
me
n
s
i
o
n
s
,”
Q
u
a
n
t
a
M
a
g
a
z
i
n
e
,
Mar
ch
20
1
6.
[O
n
l
i
n
e].
A
v
ai
l
a
b
l
e
:
h
t
t
p
s
:
/
/
w
w
w
.
q
u
a
n
t
ama
g
azi
n
e.
o
r
g
/
s
p
h
ere
-
p
ac
k
i
n
g
-
s
o
l
v
e
d
-
in
-
h
i
g
h
er
-
d
i
me
n
s
i
o
n
s
-
2
0
1
6
0
3
3
0
/
.
[7
]
M.
W
an
g
an
d
W
.
D
e
n
g
,
“
D
ee
p
Face
Reco
g
n
i
t
i
o
n
:
A
Su
r
v
ey
,
”
arX
i
v
:
1
8
0
4
.
0
6
6
5
5
,
2
0
1
9
.
[8
]
W
.
L
i
u
,
et
al
.
,
“
D
eep
H
y
p
er
s
p
h
eri
cal
L
earn
i
n
g
,
”
arX
i
v
:
1
7
1
1
.
0
3
1
8
9
,
2
0
1
8
.
[9
]
B.
A
mo
s
,
et
al
.
,
“
O
p
en
Face:
A
g
en
eral
-
p
u
r
p
o
s
e
face
reco
g
n
i
t
i
o
n
l
i
b
rar
y
w
i
t
h
mo
b
i
l
e
ap
p
l
i
cat
i
o
n
s
,
”
CMU
-
CS
-
16
-
1
1
8
,
2
0
1
6
.
A
v
ai
l
ab
l
e:
h
t
t
p
:
/
/
e
l
i
j
ah
.
c
s
.
cmu
.
ed
u
/
D
O
CS
/
CMU
-
CS
-
16
-
1
1
8
.
p
d
f.
[1
0
]
M.
Pl
o
n
u
s
,
“
CH
A
PT
E
R
9
-
D
i
g
i
t
al
S
y
s
t
ems
,
”
E
l
ect
r
o
n
i
c
s
a
n
d
Co
m
m
u
n
i
ca
t
i
o
n
s
f
o
r
S
ci
e
n
t
i
s
t
s
a
n
d
E
n
g
i
n
e
er
s
,
A
cad
em
i
c
Pres
s
,
p
p
.
3
2
7
-
4
0
3
,
2
0
0
1
.
[1
1
]
K
.
Z
h
an
g
,
e
t
al
.
,
“
J
o
i
n
t
Fa
ce
D
e
t
ect
i
o
n
an
d
A
l
i
g
n
me
n
t
U
s
i
n
g
Mu
l
t
i
t
a
s
k
C
a
s
cad
e
d
Co
n
v
o
l
u
t
i
o
n
al
N
et
w
o
r
k
s
,
”
I
E
E
E
S
i
g
n
a
l
P
r
o
ce
s
s
i
n
g
Let
t
er
s
,
v
o
l
.
2
3
,
n
o
.
1
0
,
p
p
.
1
4
9
9
-
1
5
0
3
,
O
ct
2
0
1
6
.
[1
2
]
Pro
g
rammer
So
u
g
h
t
,
“
Face
k
ey
p
o
i
n
t
d
et
ec
t
i
o
n
al
g
o
r
i
t
h
m
-
MT
C
N
N
,
”
2
0
1
8
.
[O
n
l
i
n
e].
A
v
a
i
l
a
b
l
e
:
h
t
t
p
:
/
/
w
w
w
.
p
r
o
g
rammer
s
o
u
g
h
t
.
c
o
m/
ar
t
i
c
l
e/
9
0
0
0
1
1
8
8
4
2
8
/
.
[1
3
]
S.
G
o
n
g
,
et
al
.
,
“
O
n
t
h
e
C
ap
ac
i
t
y
o
f
Face
Rep
res
en
t
at
i
o
n
,
”
arX
i
v
:
1
7
0
9
.
1
0
4
3
3
,
2
0
1
9
.
[1
4
]
F.
Sch
ro
ff,
et
al
.
,
“
FaceN
et
:
A
u
n
i
fi
e
d
emb
ed
d
i
n
g
fo
r
f
ace
reco
g
n
i
t
i
o
n
an
d
cl
u
s
t
er
i
n
g
,
"
2
0
1
5
IE
E
E
Co
n
f
e
r
en
ce
o
n
Co
m
p
u
t
er
V
i
s
i
o
n
a
n
d
P
a
t
t
er
n
R
ec
o
g
n
i
t
i
o
n
(CV
P
R
)
,
p
p
.
8
1
5
-
8
2
3
,
2
0
1
5
.
[1
5
]
M.
W
an
g
an
d
W
.
D
en
g
,
“
D
ee
p
Face
Reco
g
n
i
t
i
o
n
:
A
Su
r
v
ey
,
”
2
0
1
8
.
[O
n
l
i
n
e].
A
v
a
i
l
a
b
l
e
:
h
t
t
p
s
:
/
/
w
w
w
.
r
es
earch
g
at
e.
n
et
/
p
u
b
l
i
ca
t
i
o
n
/
3
2
4
6
0
0
0
0
3
_
D
ee
p
_
F
ace_
Reco
g
n
i
t
i
o
n
_
A
_
S
u
rv
e
y
.
[1
6
]
F.
Sch
ro
ff,
et
al
.
,
“
FaceN
et
:
A
U
n
i
f
i
ed
E
m
b
ed
d
i
n
g
f
o
r
F
ace
Reco
g
n
i
t
i
o
n
an
d
Cl
u
s
t
er
i
n
g
,
”
arX
i
v
:
1
5
0
3
.
0
3
8
3
2
,
2
0
1
5
.
[1
7
]
T.
T
.
D
o
,
et
al
.
,
“
A
T
h
eo
ret
i
cal
l
y
So
u
n
d
U
p
p
er
Bo
u
n
d
o
n
t
h
e
T
ri
p
l
e
t
L
o
s
s
fo
r
Imp
r
o
v
i
n
g
t
h
e
E
ff
i
ci
e
n
c
y
o
f
D
eep
D
i
s
t
a
n
ce
Met
ri
c
L
earn
i
n
g
,
”
CV
P
R
2
0
1
9
,
A
p
r
il
2
0
1
9
.
[1
8
]
R.
O
Bri
en
,
“
H
o
w
D
o
es
a
U
SB
t
o
Seri
al
A
d
ap
t
er
W
o
r
k
?
”
2
0
1
9
.
[O
n
l
i
n
e].
A
v
ai
l
ab
l
e:
h
t
t
p
s
:
/
/
i
t
s
t
i
l
l
w
o
r
k
s
.
c
o
m/
u
s
b
-
s
eri
a
l
-
ad
a
p
t
er
-
w
o
r
k
-
4
9
6
9
1
6
2
.
h
t
ml
.
[1
9
]
W
i
k
i
p
ed
i
a,
“
RS
-
2
3
2
,
”
A
v
a
i
l
a
b
l
e
:
h
t
t
p
s
:
/
/
e
n
.
w
i
k
i
p
e
d
i
a.
o
r
g
/
w
i
k
i
/
RS
-
2
3
2
.
[2
0
]
R.
A
l
i
mu
i
n
,
et
a
l
.
,
“
D
es
i
g
n
o
f
a
4
-
Ch
a
n
n
e
l
V
i
r
t
u
a
l
In
s
t
r
u
men
t
V
i
d
eo
A
d
a
p
t
er
fo
r
D
i
g
i
t
al
D
a
t
a
Mu
l
t
i
p
l
ex
i
n
g
,
”
2
0
1
8
IE
E
E
1
0
t
h
In
t
er
n
a
t
i
o
n
a
l
Co
n
f
e
r
en
ce
o
n
H
u
m
a
n
o
i
d
,
Na
n
o
t
ec
h
n
o
l
o
g
y,
In
f
o
r
m
a
t
i
o
n
Tech
n
o
l
o
g
y,
Co
m
m
u
n
i
c
a
t
i
o
n
a
n
d
Co
n
t
r
o
l
,
E
n
vi
r
o
n
m
e
n
t
a
n
d
M
a
n
a
g
em
e
n
t
(H
NIC
E
M
)
,
p
p
.
1
-
5
,
2
0
1
8
.
[2
1
]
T
rei
c
h
l
er
J.
an
d
L
ari
mo
re
,
“
T
h
eo
r
y
an
d
d
es
i
g
n
o
f
ad
a
p
t
i
v
e
fi
l
t
ers
,
”
Pren
t
i
ce
-
H
al
l
o
f
In
d
i
a,
2
0
0
7
.
[2
2
]
A
.
Si
n
h
a
,
“
H
o
w
T
o
D
et
ec
t
a
n
d
E
x
t
ract
Face
s
fr
o
m
a
n
Imag
e
w
i
t
h
O
p
e
n
CV
a
n
d
Py
t
h
o
n
,
”
D
i
g
i
t
a
l
O
ce
an
,
2
0
1
9
.
[O
n
l
i
n
e].
A
v
a
i
l
ab
l
e:
h
t
t
p
s
:
/
/
w
w
w
.
d
i
g
i
t
al
o
cean
.
c
o
m/
c
o
mmu
n
i
t
y
/
t
u
t
o
ri
al
s
/
h
o
w
-
to
-
d
e
t
ect
-
a
n
d
-
e
x
t
rac
t
-
faces
-
fr
o
m
-
an
-
i
mag
e
-
w
i
t
h
-
o
p
e
n
cv
-
a
n
d
-
p
y
t
h
o
n
.
[2
3
]
N
.
J
.
Py
u
n
,
“
E
x
t
ract
i
o
n
o
f
an
i
ma
g
e
i
n
o
r
d
er
t
o
a
p
p
l
y
fa
ce
reco
g
n
i
t
i
o
n
met
h
o
ds
,
”
A
r
t
i
f
i
c
i
a
l
In
t
e
l
l
i
g
e
n
ce
,
U
n
i
v
er
s
i
t
é
So
rb
o
n
n
e
Pari
s
Ci
t
é
,
2
0
1
5
.
[2
4
]
J
.
Bro
w
n
l
ee,
“
A
G
en
t
l
e
In
t
ro
d
u
c
t
i
o
n
t
o
D
eep
L
earn
i
n
g
fo
r
Face
Reco
g
n
i
t
i
o
n
,
”
Mach
i
n
e
L
earn
i
n
g
Mas
t
er
y
,
2
0
1
9
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
s
:
/
/
mac
h
i
n
el
ear
n
i
n
g
ma
s
t
er
y
.
co
m
/
i
n
t
r
o
d
u
c
t
i
o
n
-
to
-
d
eep
-
l
earn
i
n
g
-
fo
r
-
face
-
rec
o
g
n
i
t
i
o
n
/
.
[2
5
]
K
.
V
.
A
ry
a
an
d
A
.
A
d
ars
h
,
"
A
n
E
ff
i
ci
e
n
t
Face
D
e
t
ect
i
o
n
an
d
Rec
o
g
n
i
t
i
o
n
Met
h
o
d
fo
r
Su
r
v
ei
l
l
a
n
ce,
"
2
0
1
5
In
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
er
e
n
ce
o
n
Co
m
p
u
t
a
t
i
o
n
a
l
I
n
t
e
l
l
i
g
e
n
ce
a
n
d
Co
m
m
u
n
i
c
a
t
i
o
n
Net
wo
r
ks
(CICN)
,
p
p
.
2
6
2
-
2
6
7
,
2
0
1
5
.
Evaluation Warning : The document was created with Spire.PDF for Python.