Int
ern
at
i
onal
Journ
al of E
le
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
8
, No
.
6
,
Decem
ber
201
8
, p
p.
4258
~
4264
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v8
i
6
.
pp4258
-
42
64
4258
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Strate
gy for Fo
reground
Moveme
nt Identi
fication
Adaptiv
e to
Bac
k
grou
nd
Va
riations
K.
An
ur
adha
,
N
.
R.
Raajan
School
of El
ec
tr
i
ca
l
and
Elec
t
roni
cs
Engi
n
ee
rin
g,
SA
STRA
Dee
me
d
Univer
si
t
y
,
In
dia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
J
a
n
30
, 2
01
8
Re
vised
Ju
l
18
,
201
8
Accepte
d
Aug
9
, 2
01
8
Video
proc
essing
has
gai
n
ed
a
l
ot
of
significan
c
e
because
of
it
s
appl
i
ca
t
ions
in
var
ious
a
rea
s
of
rese
ar
ch.
This
inc
lude
s
m
onitoring
m
ovements
in
public
pla
c
es
for
survei
ll
an
ce
.
Video
se
quenc
es
from
va
rious
standa
rd
d
at
ase
ts
such
as
I2R,
CAV
IAR
and
UCS
D
are
of
te
n
r
eferred
for
vid
eo
proc
essing
appl
i
ca
t
ions
and
rese
arc
h
.
Ide
n
tification
of
a
ct
or
s
as
well
as
the
m
ovemen
ts
in
vide
o
sequ
en
ce
s
should
be
a
cc
om
pli
shed
wi
th
the
st
at
i
c
an
d
d
y
nami
c
bac
kground.
Th
e
signifi
c
ance
of
rese
arc
h
in
vide
o
proc
ess
ing
li
es
in
ide
nti
f
y
i
ng
the
fore
ground
move
m
ent
of
actors
and
obje
cts
in
vide
o
seque
nce
s.
Fore
ground
ide
n
ti
fi
c
at
ion
ca
n
be
do
ne
with
a
static
or
d
y
namic
bac
kground.
This
t
y
pe
of
ide
nt
ifica
t
ion
bec
om
es
complex
while
d
et
e
ct
ing
the
m
ovements
in
vide
o
sequ
e
nce
s
with
a
d
y
nami
c
ba
ckg
round.
For
ide
nti
f
icati
on
of
fore
ground
m
ovement
in
v
ideo
seque
nc
es
wi
th
d
y
namic
bac
kground,
two
al
go
rit
hm
s
are
proposed
in
thi
s
art
ic
l
e.
Th
e
al
g
orit
hm
s
are
te
rm
ed
as
Fram
e
Diffe
ren
ce
bet
wee
n
Ne
igh
boring
Fram
es
using
Hue,
Satura
ti
on
and
Value
(FD
N
F
-
HS
V)
and
Fram
e
Diffe
ren
ce
bet
we
en
Neighbor
ing
Fra
m
es
using
Gre
y
s
ca
l
e
(FD
NF
-
G).
W
it
h
reg
ard
to
F
-
m
ea
sure
,
re
call
and
pr
ec
is
ion,
the
propose
d
al
gorit
hm
s
are
eva
luated
with
stat
e
-
of
-
ar
t
te
chn
ique
s.
R
esult
s
of
ev
al
ua
ti
o
n
show
tha
t,
the
proposed
al
gor
i
thms
have
show
n
enha
nc
ed
per
form
ance.
Ke
yw
or
d:
Ad
a
ptive t
hr
es
ho
l
d
Ba
ckgrou
nd s
ubtract
io
n
Fr
am
e d
iffer
e
nc
e
Moti
on
detect
ion
Track
i
ng
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
K.
A
nuradha
,
School
of Elec
tric
al
an
d El
ect
ronics E
nginee
rin
g,
SA
STR
A Dee
m
ed
U
niv
e
rsity
,
Tirum
al
ai
sa
m
u
dr
am
, Th
a
njav
ur,
I
ndia
.
E
m
a
il
:
kan
ukal
ya
n79@gm
ai
l.
com
1.
INTROD
U
CTION
In
vid
e
o
proc
essing,
f
or
e
groun
d
identific
at
ion
can
be
accom
plished
with
a
sta
ti
c
or
dynam
ic
backg
rou
nd.
T
his
proce
ss
is
diff
ic
ult,
w
he
n
identify
in
g
th
e
m
ov
em
ents
in
vi
deo
seq
ue
nces
with
a
dynam
i
c
backg
rou
nd.
N
um
ero
us
al
go
r
it
h
m
s
hav
e
pe
rfor
m
ed
the
f
oreg
rou
nd
i
den
t
ific
at
ion
in
vid
eo
seq
ue
nces
.
It
is
fou
nd
that,
t
he
se
al
gorithm
s
hav
e
done
th
e
foregrou
nd
i
den
ti
ficat
io
n
with
le
ss
i
m
po
rta
nce
to
cha
ng
e
s
in
backg
rou
nd
a
nd
il
lu
m
inati
on
.
This
wor
k
proposes
tw
o
al
gorithm
s
FD
NF
-
HSV
an
d
F
D
NF
-
G.
T
he
pr
opos
e
d
al
gorithm
s
hav
e
ide
ntifie
d
t
he
foregr
ound
m
ov
e
m
ents
with
si
gn
ific
a
nce
to
var
ia
ti
ons
i
n
il
lum
i
na
ti
on
a
nd
backg
rou
nd.
The
pr
opos
e
d
al
gorithm
s
hav
e
addresse
d
th
e
issue,
by
co
m
pu
ti
ng
the
ad
aptive
th
reshol
d
of
th
e
changes
am
on
g fr
am
es.
This
sect
ion
pr
e
sent
s the lit
eratu
re
of the
releva
nt
work stu
died
.
A
r
obus
t
te
ch
ni
qu
e
f
or
ta
r
get
trackin
g
wa
s
present
ed
in
[
1]
.
This
te
chn
i
que
m
ines
the
ta
rg
et
s
from
a
vid
e
o.
T
he
n,
t
he
ta
r
gets
we
r
e
cat
egorized
per
ti
ne
nt
to
prop
e
rtie
s
of
im
a
ges.
T
his
m
et
ho
d
has
trac
ke
d
obj
ect
s
with v
a
riat
ion
s
in
ap
pear
a
nce.
A
bac
kgr
ound
m
od
el
f
or s
urveil
la
nce w
it
h t
he
Pantil
t
-
Zo
om
(
PTZ)
ca
m
e
ra
wa
s
pro
po
se
d
i
n
[2]
.
This
m
od
el
has
pr
ese
nte
d
a
te
xture
de
scr
iptor
f
or
e
nc
oding
s
patio
-
te
m
poral
data
.
T
his
m
od
el
was
li
nked
wi
th
a
set
of
te
chn
i
qu
e
s
f
or
t
rack
i
ng
obj
ect
s.
I
n
[
3]
,
a
te
xtu
al
descr
i
ptor
re
ferred
a
s
Ce
nter
sy
m
m
e
tric
sca
le
inv
ariant
lo
cal
te
rn
ary
pa
tt
ern
s
(CS
-
SIL
TP
)
wa
s
co
ntr
ibu
te
d
for
bui
lding
a
backg
rou
nd
m
od
el
.
This
m
od
el
was
ap
plied
for
ide
ntif
yi
ng
f
or
e
gro
und
obj
ect
s
i
n
eac
h
of
the
capt
ur
e
d
fr
am
es.
To
i
den
ti
fy
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Stra
te
gy
for
F
ore
groun
d
M
ov
emen
t
I
den
ti
fi
cation A
daptive
to B
ackg
r
ound
Va
ri
ations (
K.
Anur
adha)
4259
the
obj
ect
s
in
m
otion
,
a
no
n
-
pa
noram
ic
a
lgorit
hm
relyi
ng
on
s
patio
-
t
e
m
po
ral
bac
kg
rou
nd
was
pre
sented
in
[
4]
.
T
his
a
lgorit
hm
has
so
lve
d
t
he
p
r
ob
le
m
s
with
pano
ram
ic
m
e
thods
s
uc
h
as
slo
w
init
ia
li
zat
ion,
backg
rou
nd
ad
aptat
ion
an
d
e
r
ror
acc
um
ulatio
n.
Re
ly
ing
on
Ba
ye
s
decisi
on
fr
am
ewo
r
k,
a
te
ch
nique
to
e
xcerpt
foregr
ound
ob
je
ct
s
from
a
vi
deo
was
c
on
t
ribu
te
d
in
[
5]
.
I
n
this
w
ork
,
a
backg
rou
nd
wi
th
m
ov
ing
a
nd
sta
ti
c
obj
ect
s
was
m
od
el
e
d
.
T
he
n,
t
he
obj
ect
s
in
the
f
oreg
rou
nd
wer
e
i
den
ti
fie
d
by
com
bin
in
g
the
outc
om
e
of
sta
ti
c
po
i
nts a
nd poi
nts in
m
ov
em
e
nt.
By
ap
plyi
ng
a m
ixtur
e m
od
el
, th
e w
ork
pro
pose
d
in
[6]
ide
ntifie
s sh
ad
ow
s in
m
otion
. This t
echn
iq
ue
has
re
du
c
ed
over
hea
d
by
det
ect
ing
sh
a
dow
s
on
ly
on
f
or
e
gro
und
pi
xels.
By
sa
m
pling
values
,
the
co
debo
ok
backg
rou
nd
s
ubtract
io
n
m
et
h
od
pr
e
sente
d
in
[
7]
captu
res
structu
ral
bac
kgr
ound
m
ov
em
ents
fo
r
a
lo
ng
ti
m
e.
This
al
gorithm
has
trac
ke
d
m
ov
ing
obj
ect
s
with
var
yi
ng
backg
rou
nd
a
nd
c
ha
nges
in
il
lu
m
inati
on
.
T
o
deal
with
c
om
plica
t
ed
dynam
ic
back
gr
ound
,
a
m
et
hod
was
pro
po
s
ed
in
[
8]
.
This
m
et
ho
d
ha
s
pro
pose
d
a
scal
e
-
inv
a
riant
patte
rn
operat
or
an
d
a
te
c
hn
i
que
for
est
im
ation
of
ke
rn
el
de
ns
i
ty
in
patte
r
ns
.
By
com
bin
ing
these,
the
m
et
ho
d
ha
s
ha
nd
le
d
dynam
ic
back
gr
ound.
An
al
gorithm
fo
r
det
ect
ing
ob
j
ect
s
in
m
ot
ion
was
offer
e
d
i
n
[9]
.
T
his
m
et
ho
d
store
s
the
values
of
pi
xels
in
the
past.
The
n,
it
relat
es
the
prese
nt
a
nd
past
values
of
t
he
pix
e
l.
Fi
nally
,
the
m
e
tho
d
i
de
ntifie
s
the
pa
rtic
ular
bac
kgr
ound
to
wh
ic
h
the
pi
xel
belo
ngs.
By
com
bini
ng
the
colo
r
in
form
at
i
on
an
d
S
ILT
P,
a
blo
c
k
-
wise
backg
rou
nd
m
od
el
was
c
ontr
ibu
te
d
in
[10]
.
This
m
od
el
ha
s
dealt
with
m
ultim
o
dal
an
d
dyna
m
ic
back
gr
ound.
F
or
ef
fici
ent
ide
ntific
at
ion
of
obj
e
ct
s
in
m
otion
,
t
he
c
ol
or
inf
or
m
at
ion
and SILT
P
wer
e
com
bin
ed
.
An
ef
fici
ent
m
et
hod
base
d
on
bac
kgrou
nd
su
bt
racti
on
wa
s
pr
ese
nted
in
[11]
.
By
red
uc
ing
the
data
dim
ens
ion
al
it
y
of
im
age
fr
am
e
and
a
pp
ly
in
g
the
sp
ars
e
re
presentat
io
n,
th
e
m
et
ho
d
ha
s
extracte
d
fore
gro
und
obj
ect
s.
A
ne
w
integrate
d
m
e
t
hod
w
as
pro
po
sed
in
[
12]
.
By
com
bin
ing
the
reg
io
n
grow
i
ng
with
th
res
ho
l
ding,
the
m
e
tho
d
fin
ds
the
reg
i
on
of
interest
(R
OI)
in
an
im
age.
An
al
gorithm
t
erm
ed
as
DTGLM
M
-
K
was
offe
re
d
in
[
13]
.
T
he
a
lgorit
hm
has
f
ocused
on
im
age
se
gm
e
ntati
on
in
a
n
en
ha
nced
m
ann
er.
This
al
gorith
m
was
appr
opriat
e
f
or
var
i
ous
ty
pes
of
data
a
nd
ap
plica
ti
on
s.
The
m
e
tho
d
de
vel
op
e
d
i
n
[
14]
presents
a
f
ram
ew
or
k
for
detect
ion
a
nd
rec
ogniti
on
of
hum
an
act
ion
s.
The
m
et
ho
d
se
gm
ents
the
obj
e
ct
s
i
n
m
otion
,
exce
r
pts
a
set
of
featur
e
s
an
d
c
hoos
es
t
he
fea
tures.
For
rec
ogniti
on
of
ac
ti
on
s,
t
he
ch
ose
n
feat
ur
es
w
ere
cl
assifi
ed
with
a
m
ul
ti
cl
ass
SV
M.
A
m
et
ho
d
f
or
i
den
ti
fyi
ng
m
ov
ing
ob
j
ect
s
was
propose
d
in
[
15
]
.
T
he
m
et
hod
assesse
s
the
bi
-
directi
onal
op
t
ic
al
flow
betw
een
fr
am
es.
Th
en
it
is
i
m
pr
oved
an
d
norm
al
i
zed.
Finall
y,
th
e
m
et
ho
d
ide
ntifie
s
the
m
ov
in
g
obj
ect
s
by
ve
rify
ing
t
he
t
hr
es
ho
ld
of
t
he
op
ti
c
al
flo
w
of
eac
h
blo
c
k
a
nd
th
e
opti
cal
flo
w
of
the
obj
ect
und
e
r
c
on
si
der
at
io
n.
An
al
gorithm
te
rm
ed
as
‘
Goo
d
Feat
ur
es
to
T
rack’ w
as p
r
es
ented
i
n
[16]
.
First,
the
al
gorithm
extr
act
s
fea
ture
from
the
fr
am
es
.
Then,
the
fea
tures
of
m
ov
ing
obj
ect
s
we
re
excerp
te
d
fro
m
the
su
bse
que
nt
fr
a
m
e.
Ba
sed
on
the
m
otion
inform
ation
an
d
locat
ion
,
the
ob
je
ct
s
in
m
otion
for
each
f
ram
e
wer
e
identifie
d
. T
his
work
has
al
s
o form
ed
cl
us
te
r
s of
obj
ect
s i
n
m
ot
ion
.
Dep
e
ndin
g
on
the
fixe
d
s
pati
al
associat
ion
a
m
on
g
the
c
urren
t
a
nd
ra
ndom
ly
cho
sen
pi
xels,
a
novel
fr
am
ewo
r
k
wa
s
pr
e
sente
d
in
[
17
]
.
T
he
m
et
ho
d
f
or
m
s
a
sam
ple
set
of
sp
a
ti
al
info
rm
at
ion
f
or
e
ach
pix
e
l.
The
n,
the
S
patia
l
Sam
ple
Diff
ere
nc
e
Consens
us
(
SSD
C
)
was
c
om
pu
te
d
f
or
i
de
ntific
at
ion
of
f
or
e
gro
und
obj
e
ct
s.
A
n
appr
oach
for
detect
ion
of
m
ot
ion
an
d
se
gm
entat
ion
w
as
dev
el
oped
in
[18]
.
T
he
appr
oach
c
ontrols
the
un
ce
rtai
nties
i
n
the
m
ov
em
e
nt
of
t
he
cam
e
ra
a
nd
com
pu
t
at
ion
al
dis
pa
riti
es.
By
com
bin
in
g
the
col
or,
m
otion
pro
bab
il
it
y
and
de
pth
c
ues
,
the
ap
proac
h
ha
s
segm
ented
the
m
ov
ing
obje
ct
s.
W
it
h
im
portance
t
o
sta
ti
sti
cal
le
arn
in
g,
a
n
effi
ci
ent
m
et
ho
d
f
or
im
age
segm
entat
ion
w
as
pr
esented
in
[19]
.
The
m
et
ho
d
app
li
es
the
Ra
yl
ei
gh
distrib
ution
to
com
pu
te
the
pro
ba
bili
ty
den
s
it
y
of
bac
kground
pi
xel.
T
hi
s
m
et
ho
d
was
app
li
ed
to
im
a
ges
of
var
i
ou
s
c
olors.
An
al
gorithm
base
d
on
the
s
cal
e
inv
aria
nt
featur
e
t
ran
s
f
orm
(S
IF
T)
was
con
t
rib
uted
in
[20]
.
The
m
et
ho
d
w
as
ap
plied
f
or
i
den
ti
fyi
ng
gra
d
ual
t
ransi
ti
ons
an
d
s
udde
n
va
riat
ion
s
with
out
a
nee
d
f
or
tr
ai
ning
of
t
he
vi
deo.
A
m
et
ho
d
f
or
detect
ion
of
m
ov
i
ng
obje
ct
s
in
GPU
was
pro
posed
in
[21]
.
This
m
et
ho
d
ha
s
enh
a
nce
d
the
qu
al
it
y
of
outc
om
e
in
scenario
s
w
he
re
the
ba
ckgr
ound
an
d
the
obj
ect
s
i
n
m
ot
ion
ap
pea
r
to
be
sam
e. Th
is m
e
t
hod has
au
t
oma
ti
cal
ly
ch
os
en
the
reg
i
on
s
of
interest
(
R
OI).
2.
RESEA
R
CH M
ET
HO
D
Tw
o
al
gorithm
s
FD
N
F
-
HSV
and
F
D
NF
-
G
are
pro
po
se
d
f
or
f
or
e
gro
und
m
ov
e
m
ent
ide
ntific
at
ion.
The
F
D
NF
-
H
SV
al
gorithm
identifie
s
t
he
fore
groun
d
m
ov
e
m
ent
us
ing
H
SV
col
or
m
od
el
.
Sect
ion
2.1
.
1
pr
ese
nts
the
al
gorithm
fo
r
F
DNF
-
HSV.
Se
ct
ion
2.2
.1
pr
e
sents
t
he
al
gor
it
h
m
fo
r
F
DNF
-
G.
T
he
al
go
rithm
s
com
m
on
fo
r
FDNF
-
HSV
a
nd
F
D
NF
-
G
(
i.e.
)
cal
culat
io
n
of
thres
hold
an
d
drawi
ng
boun
ding
box
a
re
pr
ese
nted
in se
ct
ion
s
2.1.2 a
nd
2.1.3
resp
ect
i
vely
.
2.1.
Motio
n d
e
tect
ion u
sin
g HS
V
ch
annel
The
propose
d
m
et
ho
d
ref
e
r
red
as
FDNF
-
HS
V
(Fram
e
Diff
e
re
nce
bet
ween
Neig
hbori
ng
Fr
am
es
us
in
g
H
SV)
f
ocuses
on
the
foregr
ound
det
ect
ion
.
T
he
pr
opos
e
d
al
gorit
hm
wo
rks
as
f
ollows:
Using
HS
V
(Hue,
Sat
ur
at
i
on,
Value
)
col
or
m
od
el
,
a
pix
el
-
based
dif
fe
ren
ce
am
on
g
t
he
fr
am
es
are
com
pu
te
d
.
F
or
each
fr
am
e
ta
ken
i
nto
acc
ount,
t
he
three
c
om
po
ne
nts
of
t
he
fr
am
e
(
i.e.
)
HSV
a
r
e
obta
ined
.
Ba
sed
on
the
H
S
V,
t
he
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4258
-
4264
4260
diff
e
re
nce
am
ong
t
he
su
cces
sive
fr
am
es
(
i.
e.
)
del
h
,
del
s
a
nd
de
l
v
are
c
om
pu
te
d.
S
ubse
qu
e
ntly
,
th
ree
bin
a
ry
i
m
ages
are
obta
ined
by
ap
pl
yi
ng
the
pro
po
s
ed
a
da
ptiv
e
threshold
al
gorithm
pr
esented
in
2.1
.2.
Using
equ
at
io
n
(
1)
,
t
he
res
ultant
im
age
is
ob
ta
in
ed
by
cal
culat
ing
t
he
m
axi
m
um
value
from
each
pi
xel
of
bin
a
ry
i
m
ages.
12
m
a
x(
_
,
_
,
_
)
f
re
su
lt
th
im
th
im
th
im
(1)
Af
te
r
filt
erin
g
the
sm
al
l
con
necte
d
r
egi
ons
in
the
res
ult
f
,
the
pro
pose
d
al
gorith
m
detect
s
th
e
foregr
ound m
ov
em
ent b
y ap
pl
yi
ng
the
boun
ding
box
m
et
ho
d.
2.1.
1.
Algori
thm
for
Motio
n
Det
ec
tion usin
g
H
SV
Step
1
:
I
n
fram
e
R
G
B
toH
S
V
b
b
//
chan
ges
the t
ru
e
co
l
or
im
age RGB to
the
HSV
of
bac
kgr
ound im
age
b
Step
2
: for
f
= 2
t
o NF
-
1
//
N
F
d
e
no
te
s
the
num
ber
of
fr
am
es
Step i
:
I
n
fra
m
e
R
G
B
toH
S
V
f
f
(i.e
.)
,
|
1
,
1
I
n
I
n
i
j
i
r
o
w
j
c
o
l
f
//
cha
ng
e
s the
tru
e
colo
r
im
age RGB to
HSV c
ol
or cha
nnel
s
Step ii
: Si
m
il
arly
1
1
I
n
frame
R
G
B
toHSV
f
f
Step ii
i
: Find t
he
a
bso
lute dif
fer
e
nce
betwee
n
H
c
ha
nn
el
1
f
In
an
d
f
In
:
1
|
(
:
,
:
,
1
)
(
:
,
:
,
1
)
|
h
f
f
d
e
l
I
n
I
n
Step iv
: Find t
he
a
bso
lute dif
fer
e
nce
betwee
n
S
ch
a
nn
el
1
f
In
an
d
f
In
:
1
|
(
:
,
:
,
2
)
(
:
,
:
,
2
)
|
s
f
f
d
e
l
I
n
I
n
Step
v
: Find t
he
a
bso
lute dif
fer
e
nce
betwee
n
V
c
ha
nn
el
1
f
In
an
d
f
In
:
1
|
(
:
,
:
,
3
)
(
:
,
:
,
3
)
|
v
f
f
d
e
l
I
n
I
n
Step
vi
: Cal
l
th_
i
m
=
T
hr
es
hold(
33
,
d
e
l
d
e
l
)
Step
vi
: Cal
l
th_
i
m1
= T
hr
es
hold(
11
,
d
e
l
d
e
l
)
Step
vii
: Cal
l
th_
i
m2
= T
hr
es
hold(
22
,
d
e
l
d
e
l
)
Step
viii
: Fin
d
the
r
es
ul
t by
fin
ding
pi
xel
-
wise
the m
axim
a
l value
12
m
a
x(
_
,
_
,
_
)
f
re
su
lt
th
im
th
im
th
im
Step
ix
: C
onver
t i
m
age
f
r
esu
l
t
to B
oo
le
a
n
i
m
age b
y a
pp
ly
ing
:
,
1
,
:
,
0
1
,
1
f
f
T
re
su
lt
i
j
b
o
o
l
i
j
F
re
su
lt
i
j
i
ro
w
j
c
o
l
Step
x
: Dr
a
w b
ounding b
ox to
d
et
ec
t t
he
f
or
e
gro
un
d
m
ov
em
ent.
2.1.2.
Algori
thm
for Thres
h
old c
alcula
tio
n:
Thres
ho
ld
(
12
,
im
im
)
Step
1
: Cal
culat
e thre
shold
us
in
g
11
1
m
a
x
m
a
x
T
im
im
row
c
ol
row
c
ol
row
c
ol
Step
2
: If
0
the
n
bin
a
r
y im
age o
btai
ne
d
is
,
|
1
,
1
f
t
i
m
a
g
e
p
i
x
i
j
i
r
o
w
j
c
o
l
, whe
re
1
1
1
0
,
,
:
,
,
1,
i
m
i
j
i
m
i
m
i
j
p
i
x
i
j
o
t
h
e
r
w
i
s
e
Step
3
: If
0
the
n
bin
a
r
y im
age
ob
ta
i
ne
d
is
,
|
1
,
1
t
i
m
a
g
e
p
i
x
i
j
i
r
o
w
j
c
o
l
f
, whe
re
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Stra
te
gy
for
F
ore
groun
d
M
ov
emen
t
I
den
ti
fi
cation A
daptive
to B
ackg
r
ound
Va
ri
ations (
K.
Anur
adha)
4261
,
,
|
1
,
1
2
p
ix
i
j
im
i
j
i
ro
w
j
c
o
l
2.1.3.
Algori
thm
for
Dra
w
in
g
B
ou
ndin
g
B
ox
Step
1
: Find t
he
c
onne
ct
ed
re
gions i
n
the
gi
ven Bo
olean
im
age.
Step
2
: Fil
te
rs
out sm
al
le
r
co
nn
ect
e
d re
gions
Step
3
: Id
e
ntify t
he
four c
orne
rs for
each re
gion,
i.
e.
bl
obs
Step
4
: M
ark
blobs
foun
d
in
the
fr
a
m
e w
it
h
recta
ngle
s.
2.2.
M
oti
on D
etectio
n usi
ng
Grey sc
ale
The
pro
po
s
e
d
al
gorithm
te
r
m
ed
as
FDNF
-
G
(
F
ram
e
Diff
er
ence
betwee
n
Neig
hbori
ng
F
ram
es
us
in
g
Gr
ey
scal
e
)
c
om
bin
es
the
bac
kgr
ound
subtra
ct
ion
wit
h
the p
ixel
-
base
d
va
riat
ion
b
et
wee
n
nei
ghbori
ng
f
ram
es.
The
al
gorithm
sta
rts b
y
fin
ding the
grey
scal
e for
t
he
f
ram
e
s in
the
v
i
deo s
equ
e
nce
unde
r c
onsiderati
on
. T
hr
ee
diff
e
re
nt
i
m
ages
del
1
(a
bs
ol
ut
e
diff
ere
nce
be
tween
the
c
urren
t
f
ram
e
and
it
s
pr
evio
us
f
r
a
m
e),
del
2
(
abso
lute
diff
e
re
nce
bet
ween
t
he
cu
rrent
fr
am
e
and
it
s
nex
t
fr
am
e)
and
del
3
(
absol
ute
diff
e
re
nce
betwee
n
the
c
urren
t
fr
am
e
and
t
he
backg
rou
nd
f
ra
m
e)
are
ob
ta
in
ed
.
By
a
pp
ly
in
g
the
pro
pose
d
ada
ptive
th
res
ho
l
d
al
go
rithm
2.1.2,
the
bin
a
ry
i
m
a
ges
f
or
del
1
,
de
l
2
and
del
3
ar
e
ob
ta
ine
d
.
A
f
te
r
filt
ering
th
e
no
ise
,
the
al
gorithm
identifie
s
t
he
foregr
ound m
otion
.
Algo
rith
m
f
or
m
otion
det
ect
ion
us
in
g
gr
ey
scal
e
:
Step
1
:
I
n
fram
e
R
G
B
toG
re
y
b
b
//
changes
the
tru
e
col
or
i
m
age
RGB
to
the
gr
ay
scal
e
intensit
y
of
backg
rou
nd im
age
b
Step
2
:
Fin
d bina
ry im
age for
b
In
as
b
b
i
n
Step
2
:
f
or
f
= 2 t
o NF
-
1
//
N
F
d
e
no
te
s
the
num
ber
of
fr
am
es
Step i
:
I
n
frame
RG
BtoG
re
y
f
f
(i.e.)
,
|
1
,
1
I
n
I
n
i
j
i
r
o
w
j
c
o
l
f
//
chan
ge
s
the
tr
ue
colo
r
im
age RGB to
the
gr
ay
scal
e intensit
y
of fram
es.
Step ii
: Si
m
il
arly
1
1
I
n
frame
R
G
B
toGre
y
f
f
,
1
1
I
n
fram
e
R
G
B
toG
re
y
f
f
Step ii
i
: Find t
he
a
bso
lute dif
fer
e
nce
of
1
f
In
and
f
In
:
||
1
1
de
l
In
In
ff
Step iv
: Find t
he
a
bso
lute dif
fer
e
nce
of
1
f
In
and
f
In
:
||
2
1
d
e
l
I
n
I
n
ff
Step
v
: Find t
he
a
bso
lute dif
fer
e
nce
of
b
In
and
f
In
:
||
3
d
e
l
I
n
I
n
bf
Step
vi
: C
al
l
th_
im
=
Thres
ho
l
d(
3
,
b
d
e
l
b
i
n
)
Step
vii
: C
al
l
th_
im
1
= T
hr
es
hold(
1
,_
d
e
l
t
h
i
m
)
Step
viii
: Cal
l
th_
i
m2
= T
hr
es
hold(
2
,_
d
e
l
t
h
i
m
)
Step
ix
: Fin
d
the
r
es
ul
t by findin
g pi
xel
-
wise lo
gica
l or o
per
at
io
n
12
_
|
_
|
_
f
re
su
lt
th
im
th
im
th
im
Step
x
: C
onver
t i
m
age
f
r
esu
l
t
to B
oo
le
a
n
i
m
age b
y a
pp
ly
ing
:
,
1
,
:
,
0
1
,
1
f
f
T
re
su
lt
i
j
b
o
o
l
i
j
F
re
su
lt
i
j
i
ro
w
j
c
o
l
Step
xi
: Dr
a
w b
ounding b
ox to
d
et
ec
t act
or
3.
RESU
LT
S
A
ND D
I
SCUS
S
ION
The
pro
posed
al
gorithm
s
FD
N
F
-
HSV
a
nd
F
D
NF
-
G
ha
ve
been
asses
sed
with
t
he
dataset
s
viz
.
UCSD,
I
2R
an
d
CA
VIAR.
T
he
outc
om
e
of
the
assessm
ents
is
sh
own
in
Figure
1.
The
pro
po
se
d
al
gor
it
h
m
was
asse
ssed
with
var
i
ou
s
m
et
rics
viz.
re
cal
l,
pr
eci
sio
n,
F
-
m
easur
e
a
nd
per
ce
ntage
of
w
r
ong
cl
assi
ficat
ion
(
pwc
)
. T
he foll
ow
i
ng p
a
ram
eter
s
wer
e
consi
der
e
d
for
c
om
pu
ti
ng these
m
etr
ic
s.
a.
Tru
e
posit
ive
(
TP)
-
pi
xels in
a
for
e
gro
und o
bject
cate
gorize
d
as
p
i
xels in
t
he fo
regr
ound
.
b.
Tru
e
n
e
gative
(TN)
-
pix
el
s in
a
b
ac
kgr
ound
obj
ect
c
at
e
goriz
ed
as
p
i
xels in t
he
bac
kgr
ound
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4258
-
4264
4262
c.
False
posit
ive
(
FP)
-
pi
xels in
a
backg
rou
nd cate
gorized
as
pixe
ls i
n
the
for
e
gro
und
.
d.
False
n
e
gative
(F
N
)
-
pixe
ls i
n a f
or
e
gro
und
c
at
egorized a
s
pi
xels in
t
he bac
kgr
ound
.
Figure
1. Sam
ple o
ut
pu
t
f
or
t
he
U
CS
D
datase
t, I
2R d
at
aset
a
nd CA
VIAR
da
ta
set
The
cal
c
ulati
on
of
pwc
,
preci
sion, recall
a
nd
F
-
m
easur
e
(si
m
il
arity
m
easur
e)
a
re
giv
e
n
he
re.
100%
FN
FP
pw
c
T
P
FN
FP
T
N
TP
pre
T
P
FP
TP
re
c
a
ll
TP
F
N
2
r
e
c
a
ll
p
r
e
F
m
e
a
s
u
r
e
r
e
c
a
ll
p
r
e
W
it
h
ref
e
re
nc
e
to
I2
R,
CA
VIAR
an
d
U
CSD
dataset
s,
the
pr
op
os
ed
al
go
rithm
s
FD
N
F
-
HSV
an
d
FDNF
-
G
detect
act
ivit
ie
s
with
com
plex
di
ff
ere
nces
in
t
he
bac
kgr
ound.
T
his
enc
om
passes
il
lu
m
inati
on
var
ia
ti
ons
an
d
acqu
i
rin
g
dy
na
m
ic
activiti
es.
The
I
2R
da
ta
set
com
pr
ise
s
vi
deo
se
quences
for
ide
ntific
at
ion
of
foregr
ound
obj
ect
s
with
int
ricat
e
bac
kgr
ound.
Pe
rtinent
to
F
-
m
easure
m
e
tric
for
the
I
2R
datas
et
,
an
assessm
ent
of
the
pro
pose
d
a
ppr
oach
(FDN
F
-
H
SV)
with
oth
e
r
te
ch
niqu
es
is
pr
e
sente
d
in
Ta
ble
1.
R
esults
exh
i
bit
that,
the
pro
posed
a
ppr
oach
ha
s
a
ccom
plished
m
axi
m
u
m
F
-
m
easur
e
t
han
the
oth
er
te
ch
niques
.
Figure
2
pro
vi
des
a
gr
a
ph
ic
a
l
rep
re
sentat
io
n
of
the
res
ults
at
ta
ined
f
or
dep
ic
ti
ng
the
m
et
rics
viz.,
preci
sio
n,
recall
, F
-
m
easur
e
f
or
I2
R
d
a
t
aset
. F
ig
ur
e
3
pr
ese
nts a
grap
hical
r
e
pr
ese
nt
at
ion
of
t
he ou
t
com
e o
btained
for
the
m
et
rics v
iz
., pre
ci
sion
,
r
ecal
l,
F
-
m
easur
e f
or
CAV
IA
R
datas
et
.
Table
1
.
Asses
sm
ent o
f
F
-
m
e
asur
e
w
it
h va
ri
ou
s
tech
niques
and the
pr
opose
d
m
et
ho
d (F
D
NF
-
HSV
)
for
the
I2R
d
at
aset
Bo
o
tstrap
Escalator
Fo
u
n
tain
s
Lob
b
y
W
ate
r
su
rf
ace
Fr
m
D
if
f
[1]
3
5
.36
2
1
.72
2
5
.43
1
6
.33
2
4
.26
MoG
[
6
]
5
6
.29
4
1
.17
7
6
.91
4
7
.92
7
9
.7
ACMM
M03
[
5
]
6
0
.43
3
2
.6
5
6
.51
3
0
.31
6
3
.66
Co
d
e Bo
o
k
[
7
]
6
3
.66
4
9
.82
6
1
.36
2
5
.51
7
3
.09
SIL
T
P
[
8
]
7
3
.32
6
5
.88
8
6
.23
7
8
.57
8
4
.36
ViBe
[
9
]
7
8
.26
6
4
.72
6
0
.96
2
6
.55
8
6
.82
CS
-
SIL
TP
[
3
]
7
6
.35
7
0
.72
8
7
.46
8
0
.23
8
7
.38
BITC
[
1
0
]
6
4
.86
6
3
.37
9
5
.24
7
6
.67
9
3
.02
DFB
–
A
[
4
]
7
1
.86
6
6
.37
7
7
.43
1
3
.24
9
3
.81
No
n
-
Para
m
etric
[
2
1
]
6
4
.10
-
7
0
.49
-
9
0
.11
Prop
o
sed
App
roach
(
FDN
F
-
HS
V)
8
6
.48
7
8
.04
8
6
.69
9
7
.43
9
8
.98
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Stra
te
gy
for
F
ore
groun
d
M
ov
emen
t
I
den
ti
fi
cation A
daptive
to B
ackg
r
ound
Va
ri
ations (
K.
Anur
adha)
4263
Figure
2
.
G
raph
dep
ic
ti
ng m
et
rics v
iz
., p
reci
sion, recall
, F
-
m
easur
e
for I2
R dataset
Figure
3. G
raph
represe
nting
m
et
rics v
iz
., pre
ci
sion
,
r
ecal
l,
F
-
m
easur
e f
or
CAV
IA
R
datas
et
The
pro
posed
m
e
tho
d
(
FDNF
-
G
)
ai
m
s
at
identify
in
g
f
or
e
gro
und
m
ov
em
ent
in
vi
de
o
se
quences
avail
able
in
U
CSD
bac
kgr
ound
s
ubtract
io
n
dataset
.
The
re
su
lt
s
of
v
al
idati
on
are p
rese
nt
ed
in
Table
2. Re
su
lt
s
il
lustrate
that,
t
he
pro
posed
m
et
hod
has
at
ta
ined
a
n
i
m
pr
o
ve
d
F
-
m
easur
e
wh
e
n
com
par
e
d
with
ot
her
m
et
hods
ta
ken
i
nto
acco
un
t.
Table
2
.
C
om
par
iso
n of t
he pr
opos
e
d
te
c
hn
i
que
(F
D
NF
-
G)
with
var
i
ou
s
m
et
hods
on
UCS
D back
gro
und
su
bt
racti
on dat
aset
with
reg
a
r
d
to
ave
ra
ge
F
-
m
easur
e
Seq
u
en
ces
Fr
m
D
if
f
[
1
]
MoG
[
6
]
ACMM
M03
[
5
]
Co
d
e
Bo
o
k
[
7
]
SIL
T
P
[
8
]
ViBe
[
9
]
CS
-
SIL
T
P
[
3
]
Prop
o
sed
W
o
rk
(FDNF
-
G)
Bird
s
1
9
.63
2
8
.35
2
8
.82
2
1
.21
3
0
.11
3
1
.92
2
9
.76
4
4
.44
Bo
ttle
1
9
.31
5
1
.12
2
3
.44
2
3
.09
5
5
.23
5
7
.5
6
3
.41
100
Freewa
y
2
6
.81
5
3
.06
3
1
.88
3
8
.48
5
1
.94
5
4
.62
5
5
.54
8
6
.41
Ocean
1
0
.07
2
8
.95
1
9
.03
22
5
7
.87
2
6
.33
6
0
.63
9
4
.11
Ped
estrian
s
2
4
.28
7
9
.63
2
8
.89
6
4
.3
8
1
.09
8
0
.1
8
5
.7
9
4
.73
Rain
4
1
.88
7
4
.81
8
1
.28
4
3
.28
8
5
.24
9
3
.7
8
9
.74
9
4
.34
Av
g
F
-
m
e
asu
re
2
3
.66
5
2
.65
3
5
.56
3
5
.39
6
0
.25
5
7
.36
6
4
.13
8
5
.67
4.
CONCL
US
I
O
N
Re
search
on
va
rio
us
m
et
ho
ds
an
d
al
gorith
m
s
in
vid
eo
pr
ocessin
g
is
co
m
m
end
able.
R
el
evan
t
to
the
identific
at
ion
of
f
or
e
gro
und
in
vid
e
o
seq
ue
nc
es,
lot
of
al
gori
thm
s
are
avail
a
ble
in
the
li
te
ra
ture.
M
os
t
of
these
al
gorithm
s
hav
e
done
the
f
or
e
gro
und
i
de
ntific
at
ion
with
le
ss
at
te
ntio
n
to
va
riat
ions
in
il
lum
inatio
n
an
d
backg
rou
nd.
But,
the
pro
pose
d
al
gorith
m
s
FD
NF
-
HSV
a
nd
F
DN
F
-
G
ha
ve
id
en
ti
fied
the
for
egro
un
d
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4258
-
4264
4264
m
ov
e
m
ent
with
im
po
rtance
t
o
c
hanges
i
n
il
lum
inati
on
a
nd
bac
kgrou
nd.
The
pro
po
se
d
work
has
pe
rfor
m
ed
foregr
ound
m
ov
em
ent
identif
ic
at
ion
by
cal
c
ulati
ng
the
a
da
ptive
t
hr
es
hold
of
the
var
ia
ti
ons
betwee
n
f
ra
m
es.
Re
su
lt
s
of
the
pro
posed
al
gorithm
s
are
be
tt
er
than
the
oth
e
r
sta
nd
a
r
d
al
go
rithm
s
i
n
te
rm
s
of
sim
il
arity
m
et
rics.
REFERE
NCE
S
[1]
A.
J.
Li
pton,
e
t
al.
,
“
Moving
ta
rge
t
c
la
ss
ifi
c
ati
on
and
tra
cki
ng
from
rea
l
-
ti
m
e
vide
o,
”
4th
IE
EE
Workshop
on
Appl
ic
a
ti
ons o
f Com
pute
r V
ision
,
vol
/i
ss
ue:
98
(
2
)
,
pp
.
8
–
14
,
1998
.
[2]
N.
Li
u,
et
al
.
,
“
Hier
arc
h
ical
ens
emble
of
bac
kgr
ound
m
odel
s
for
PTZ
-
base
d
video
surveil
la
n
ce,”
IEE
E
transact
io
ns
on
cy
b
erne
tics
,
v
ol
/i
ss
ue:
45
(
1
)
,
p
p.
89
–
102
,
2015
.
[3]
H.
W
u,
et
al
.
,
“
Rea
l
-
ti
m
e
ba
ckg
round
subtrac
t
io
n
-
base
d
vide
o
s
urve
il
l
ance
of
p
eopl
e
b
y
int
eg
ra
ti
ng
lo
ca
l
te
xtur
e
pat
t
ern
s,”
Signal,
Image
and
Vi
d
eo
Proc
essing
,
v
ol
/i
ss
ue:
8
(
4
)
,
pp
.
665
–
676
,
2014
.
[4]
S.
W
.
Kim
,
e
t
al
.
,
“
Det
ec
t
ion
of
m
oving
obje
c
ts
with
a
m
oving
c
amera
using
non
-
panor
amic
b
ac
k
ground
m
odel,
”
Mac
hine V
ision
and
Applications
,
vol
/i
ss
ue:
24
(
5
)
,
pp
.
10
15
–
1028
,
2013.
[5]
L.
L
i,
et
a
l.
,
“
Foreground
obj
ec
t
d
etec
t
ion
fr
om
vide
os
con
t
ai
ning
complex
bac
kground
,
”
Proce
ed
ings
of
the
el
e
ve
nth
ACM
in
te
rnational
con
f
ere
nce on
Mu
lt
i
media
,
pp
.
2
–
10
,
2003
.
[6]
P.
Kae
wtra
kulpo
ng
and
R.
Bowd
en,
“
An
Im
prove
d
Adapti
v
e
Ba
c
kgroun
d
Mixture
Model
for
Re
a
l
-
ti
m
e
Tr
ac
king
with
Shadow De
te
c
ti
on,
”
Adv
an
c
ed
V
ide
o
Based
Surve
il
lan
ce Sy
s
te
ms
,
pp
.
1
–
5
,
2
001.
[7]
K.
Kim
,
e
t
al
.
,
“
Rea
l
-
ti
m
e
fo
re
ground
-
bac
kgr
ound
segm
ent
ati
on
using
cod
eb
ook
m
odel
,
”
R
e
al
-
Time
Imagin
g
,
vol
/i
ss
ue:
11
(
3
)
,
pp.
172
–
185
,
20
05.
[8]
S.
Li
ao
,
et
al
.
,
“
Modeli
ng
pixel
proc
ess
with
sca
le
inva
r
ia
nt
loc
a
l
pat
t
ern
s
for
ba
ckgr
ound
sub
tracti
on
in
compl
e
x
sce
nes,
”
Comput
er
Vi
sion
and
Pa
tt
ern
Re
cogn
it
io
n
(
CVP
R)
,
2010
IEE
E
Confe
ren
c
e
on
,
pp.
1301
–
1
306
,
2010
.
[9]
O.
Barnich
and
M.
Van
Droo
ge
nbroe
ck
,
“
Vi
Be:
A
un
ive
rsa
l
ba
ckgr
ound
s
ubtra
c
ti
on
al
gor
it
hm
for
vid
eo
seque
nce
s,
”
I
EEE
Tr
ansacti
ons
on
Image Proce
s
sing
,
vol
/i
ss
ue:
20
(
6
)
,
pp
.
1709
–
1
724,
2011
.
[10]
H.
Han,
et
al
.
,
“
Moving
obje
ct
det
ection
rev
isi
te
d:
Speed
and
robustness,”
IE
EE
Tr
ansacti
on
s
on
Circui
ts
and
Syste
ms
for Vide
o
Technol
og
y
,
v
ol
/i
ss
ue:
25
(
6
)
,
p
p.
910
–
921
,
201
5.
[11]
Y.
W
ang,
et
al.
,
“
Com
pre
ss
ive
bac
kground
mode
li
ng
for
fore
ground
ext
ra
ct
i
on,
”
Journal
of
El
ec
tri
cal
an
d
Computer
Engi
n
ee
ring
,
vol
.
201
5,
pp
.
1
–
8
,
2015
.
[12]
S.
Hore
,
et
al
.
,
“
An
i
nte
gra
te
d
int
eract
iv
e
t
ec
h
nique
for
image
segm
ent
at
ion
u
sing
stac
k
base
d
see
ded
reg
io
n
growing
and
thr
esholdi
ng,
”
Int
e
rnational
Journ
al
of
E
le
c
tric
a
l
and
Computer
Engi
ne
ering
,
vo
l
/i
ss
ue:
6
(
6
)
,
pp
.
2773
–
2780,
201
6.
[13]
T.
J
y
o
thi
rm
a
y
i
,
et
al.
,
“
Im
age
segm
ent
atio
n
bas
ed
on
doubl
y
tr
unca
t
ed
gene
r
ali
ze
d
l
apl
a
ce
m
ixture
m
odel
and
K
m
ea
ns
cl
usteri
ng
,
”
Inte
rnat
ional
Journal
of
El
ect
rical
and
Comp
ute
r
Engi
nee
rin
g
,
vol
/i
ss
ue:
6
(
5
)
,
pp.
2188
–
2196,
2016.
[14]
M.
Sharif
,
e
t
al
.
,
“
A
fra
m
ework
of
hum
an
d
et
e
c
ti
on
and
a
ct
ion
rec
ogni
ti
on
b
ase
d
on
uniform
se
gm
ent
at
ion
and
combinat
ion
of
Euc
li
d
ea
n
d
istance
and
joi
nt
e
ntrop
y
-
b
ase
d
fe
at
ure
s
sel
ec
t
ion,”
Eurasip
Journal
on
Image
and
Vi
deo
Proce
ss
in
g
,
vol
/i
ss
ue:
201
7
(
1
)
,
2017
.
[15]
S.
S.
Sengar
an
d
S.
Mukhopad
h
y
a
y
,
“
Motion
det
e
ct
ion
using
bloc
k
b
ase
d
bi
-
dire
c
ti
ona
l
opt
ical
flow
m
e
thod,”
Journal
of
Vi
sua
l
Comm
unic
at
io
n
and
Image
Re
p
resentat
ion
,
vol
.
49,
pp
.
89
–
103
,
2017.
[16]
A.
Keiva
ni
,
et
al.
,
“
Motion
-
bas
ed
m
oving
obje
ct
detec
t
ion
and
tra
ck
ing
using
aut
om
at
i
c
K
-
m
ea
ns,”
A
FR
ICON
,
2017
IEEE
,
pp.
32
–
37
,
2
017
.
[17]
C.
Zha
ng,
e
t
al.
,
“
Moving
obje
ct
detec
ti
on
algorithm
base
d
on
pixe
l
spati
a
l
sam
ple
diffe
ren
ce
conse
nsus
,
”
Mult
imedi
a
Tool
s and
Applicatio
ns
,
vol
/
issue:
76
(
21
)
,
pp
.
22077
–
22093,
2017
.
[18]
D.
Zhou,
et
al
.
,
“
Moving
obje
c
t
d
et
e
ct
ion and
seg
m
ent
at
ion in
urb
an
envi
ronm
ent
s
from
a
m
oving
pla
tform,”
Image
and
Vi
sion
Computing
,
vol
.
68
,
pp.
76
–
87
,
2017
.
[19]
M.
Mous
sa,
e
t
a
l.
,
“
Com
par
at
ive
stud
y
of
st
at
ist
i
ca
l
bac
kg
round
m
odel
ing
and
subtraction,”
Indo
nesian
Journal
o
f
El
e
ct
rica
l
Eng
in
ee
ring a
nd
Computer
Sc
ie
nc
e
,
v
ol
/i
ss
ue:
8
(
2
)
,
pp
.
287
–
295
,
2017
.
[20]
Y.
Ta
bi
i,
e
t
al.
,
“
Video
Shot
B
oundar
y
D
et
e
ct
i
on
using
the
Scal
e
Inv
ari
an
t
Fe
at
ure
Tra
nsform
and
RGB
Color
Channe
ls,
”
In
te
r
nati
onal
Journal
of
El
e
ct
ri
cal
an
d
Computer
Eng
i
nee
ring
(
IJE
CE
)
,
vol
/i
ss
ue:
7
(
5
)
,
pp.
2565
–
2673,
2017.
[21]
D.
Berj
ón,
et
a
l.
,
“
Re
al
-
t
ime
nonpar
ametr
i
c
ba
ckgr
ound
subtra
ct
ion
with
track
ing
-
base
d
fore
g
round
updat
e
,
”
Pat
te
rn
Recogni
t
ion
,
vo
l. 74, pp.
156
–
170,
2018
.
Evaluation Warning : The document was created with Spire.PDF for Python.