Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
24
,
No.
2
,
N
ov
em
ber
20
21
,
pp.
1063
~
1073
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v
2
4
.i
2
.
pp
106
3
-
107
3
1063
Journ
al h
om
e
page
:
http:
//
ij
eecs.i
aesc
or
e.c
om
Anomal
y event d
etection
and lo
calization
of vide
o clips usi
ng
global
an
d local
outlie
rs
Sa
w
sen
Abdul
ha
di
M
ah
m
oo
d
1
,
Az
al
M
on
s
hed
Ab
id
2
,
Sadeq H
.
L
afta
3
1
,
2
Depa
rtment
of
Com
pute
r
Sci
en
ce
,
Mus
ta
nsir
i
y
a
h
Univer
sit
y
,
Ba
ghdad,
I
raq
3
Depa
rtment of
Applie
d
Sc
ie
c
e,
Univer
sit
y
of
T
e
chnol
og
y
,
Bagh
dad,
I
raq
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
pr
30
,
2021
Re
vised
Sep
7
,
2021
Accepte
d
Se
p
16
,
2021
The
aut
om
atic
d
et
e
ct
ion
of
ano
m
aly
ev
ent
s
in
vide
o
seque
nc
e
has
bec
om
e
a
cri
tica
l
issue
and
essenti
al
deman
d
for
the
ext
ensi
ve
depl
o
y
m
en
t
of
computer
vision
s
y
stems
such
as
vide
o
sur
vei
llanc
e
applic
a
ti
ons.
An
anoma
l
y
eve
n
t
in
vide
o
ca
n
be
denot
ed
as
out
li
e
r
beh
avi
or
withi
n
vide
o
fr
ames
which
form
ula
te
d
b
y
a
devi
at
ion
from
the
stabl
e
sc
ene
.
In
thi
s
pape
r,
a
n
anomal
y
eve
nt
de
tecti
on
and
loc
a
li
z
atio
n
m
et
hod
in
vide
o
seque
nc
e
i
s
pre
sente
d
inc
ludi
ng
m
ult
i
l
eve
l
stra
te
g
y
a
s
te
m
pora
l
fra
m
es
diffe
ren
ce
s
esti
m
at
ion,
m
odel
li
ng
of
n
orm
al
and
abno
rm
al
beha
v
ior
u
sing
reg
ression
m
odel
and
fina
lly
d
ensity
–
base
d
cl
uster
ing
to
det
ect
the
o
utl
ie
rs
(ab
norm
a
l
eve
nt)
at
cl
ips
le
ve
l.
Henc
e,
ou
tl
i
er
scor
e
i
s
obta
in
ed
at
the
segm
ent
or
clip
le
ve
l
a
long
vide
o
fra
m
es
seq
uenc
es.
The
pr
o
posed
m
et
hod
sepli
ts
vide
o
fr
ame
s
int
o
non
-
over
la
pp
ed
c
li
ps
using
globa
l
o
utl
ie
r
detec
t
ion
proc
ess.
After
w
ard
,
a
t
e
ach
cl
ip
,
th
e
lo
ca
l
outl
ie
rs
are
d
eterm
ine
d
base
d
on
density
of
ea
ch
clip.
Ext
ensiv
e
exp
erim
ent
s
were
cond
uct
ed
upon
two
publi
c
vide
o
datase
ts
whic
h
inc
lud
e
dense
a
nd
sca
tt
er
ed
outliers
al
ong
vid
eo
seque
nce.
The
expe
riments
were
per
fo
rm
ed
on
two
comm
on
publi
c
da
ta
sets
(Avenue
)
and
Univer
sit
y
o
f
Cal
ifornia,
San
Diego
(UCS
D)
.
The
exp
er
iment
al
result
s
exh
ibited
that
the
proposed
m
et
ho
d
detec
ts
we
ll
outl
i
er
fr
ames
at
clip
l
eve
l
with
lower
computat
ion
al
c
om
ple
xity
compari
ng
to
th
e
st
ate
-
of
-
the
-
ar
t
m
et
h
ods.
Ke
yw
or
d
s
:
Anom
al
y event
Den
s
e cluste
ring
Local
ou
tl
ie
rs
Ou
tl
ie
rs
Tem
po
ral
dif
fe
ren
ces
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
B
Y
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
Sawse
n Abd
ul
had
i M
a
hm
oo
d
Dep
a
rtm
ent o
f C
om
pu
te
r
Scie
nce
Mustansiriy
ah
Un
i
ver
sit
y
Ba
ghda
d,
Ir
a
q
Em
a
il
:
sawsenhad
i
@uom
us
ta
ns
iriy
ah.
e
du.iq
1.
INTROD
U
CTION
Ou
tl
ie
rs,
al
s
o
known
as
an
om
al
ie
s,
abn
orm
al
i
ti
es
or
rare
even
ts,
are
da
ta
sa
m
ples
or
obj
ect
s
in
ra
w
data
that
do
no
t
a
dap
t
to
a
con
ce
pt
of
norm
al
beh
aviour
[
1].
O
wing
to
the
gr
owing
re
quirem
e
nt
a
nd
app
li
cat
io
ns
in
wide
dom
ai
ns
,
su
c
h
as
vi
deo
su
r
veill
ance,
s
ecur
it
y,
he
al
th
care
an
d
m
edical
risk
[2
]
as
w
el
l
as
intru
si
on
de
te
ct
ion
[3
]
,
an
oma
ly
and
outl
ie
r
detect
io
n
ta
s
k
play
s
si
gn
ific
ant
r
ole
in
vari
ant
fiel
ds
incl
ud
i
ng
com
pu
te
r
vision,
sta
ti
sti
cal
analy
sis
an
d
m
achine
le
ar
ni
ng
.
O
utli
er
de
te
ct
ion
is
a
t
ask
of
d
et
erm
ining
a
su
b
-
r
egi
on
in
a
giv
en
data
sa
m
ples
wh
ic
h
are
co
ns
ide
red
as
abnor
m
al
su
bse
t
and
ha
ve
unusual
be
hav
i
ou
r
from
oth
er
sa
m
ples.
It
is
on
e
of
the
esse
ntial
data
m
ining
obj
ect
ives
a
nd
ba
sic
fiel
d
in
m
any
app
li
cat
ion
s,
su
c
h
as
vid
e
o s
urveil
la
nce sys
te
m
s w
hich
are
u
se
d
t
o
rec
ogn
iz
e the pote
ntial
thr
eat
s
[2
]
-
[
4
].
Ou
tl
ie
rs
or
a
nom
aly
even
ts
can
dev
ia
te
or
distor
t
t
he
sta
ti
sti
cal
m
easur
em
ents
an
d
data
distrib
uti
on
du
e
to
it
s
rar
el
y
app
ea
ra
nce
i
n
the
real
li
fe
e
ven
ts
,
al
lo
wing
a
c
onfu
s
ed
re
presentat
io
n
of
t
he
im
plici
t
dat
a
an
d
relat
ion
s
hip
s
[
5].
P
racti
cal
ly
,
the
var
ia
nt
distribu
ti
on
of
norm
al
and
a
no
m
al
y
even
ts
w
it
hin
vid
e
o
f
ra
m
es
is
un
a
nnou
nced
with
ra
re
pres
entat
ion
of
outl
ie
rs
.
In
s
urve
il
la
nce
vid
eo
s,
the
pr
e
dom
in
ant
eve
nts
occ
urri
ng
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
2
4
, N
o.
2
,
N
ove
m
ber
20
21
:
106
3
-
107
3
1064
fr
e
qu
e
ntly
are
denote
d
as
no
r
m
al
beh
avi
our,
wh
il
e
a
bnor
m
al
even
ts
a
re
r
efer
red
t
o
the
even
ts
occ
urre
d
with
low
pro
ba
bili
ty
[6
]
.
The
aim
of
outl
ie
r
detect
ion
an
d
loc
al
iz
at
ion
is
to
determ
ine
the
sp
at
ia
l
and
te
m
po
ral
su
b
-
r
egi
on
s
i
nvol
ved
t
he
an
om
al
ou
s
eve
nts
with
aut
om
a
ti
c
m
ann
er
[
7]
.
O
utli
er
detec
ti
on
m
et
ho
ds
can
be
cl
assifi
ed
in
te
rm
s
of
the
dat
a
ty
pe
re
quire
d
to
trai
n
the
m
od
el
su
ch
as
;
po
i
nt,
c
on
te
xt
ual
an
d
c
ollec
ti
ve
ano
m
al
ie
s.
Po
int
ou
tl
ie
r
is
re
ferred
to
a
sin
gu
la
r
data
inst
ance
de
viate
d
from
the
whol
e
data
sa
m
ples.
A
con
te
xtu
al
outl
ie
r
represe
nts
an
in
div
id
ual
da
ta
sa
m
ple
devi
at
ed
within
a
con
te
xt
w
hich
requires
a
no
ti
on
of
con
te
xt
as
we
ll
as
con
diti
on
al
ano
m
al
ie
s.
Coll
ect
ive
anom
al
ie
s
are
den
ote
d
by
a
se
t
of
co
rr
el
at
e
d
data
instances
a
nd
needs
a
relat
ion
s
hi
p
am
on
g
da
ta
instance
s
represe
nted
by
seq
uen
ti
al
and
s
patia
l
da
ta
.
The
singular
patte
r
ns
within
a
c
ol
le
ct
ive
ano
m
a
ly
are
co
ns
ide
red
not
an
om
a
lou
s
by
them
sel
ves.
Ba
se
d
on
the
ano
m
al
y
def
init
ion
pro
vid
e
d
in
[8
]
,
“Vi
de
o
an
om
al
ie
s
c
an
be
th
ought
of
as
the
oc
currence
of
unusual
app
ea
ra
nce o
r
m
ot
ion
p
at
te
rns
or
the
occ
urr
ence o
f
usual
a
pp
ea
ra
nce o
r
m
ot
ion
at
trib
ut
es
in u
nus
ual
l
ocati
ons
or
ti
m
es”
we
form
ulate
the
pro
blem
of
an
om
aly
even
t
de
te
ct
ion
as
re
gr
essi
on
m
od
el
fitt
ing
as
we
ll
as
cl
us
te
rin
g base
d
-
l
ocal outl
ie
r dete
ct
ion t
ask.
In
t
his
sect
io
n,
a
discuss
i
on
of
rece
nt
resea
rch
es
rele
van
t
ou
tl
ie
r
a
nd
a
nom
aly
even
t
de
te
ct
ion
in
vid
e
o
seq
ue
nc
e
will
be
pr
ese
nted.
Re
ce
ntly
,
vid
eo
a
no
m
al
y
detect
ion
s
ba
sed
on
dee
p
le
arn
i
ng
m
et
ho
ds
are
introd
uced
incl
ud
i
ng
diff
e
re
nt
scenarios
s
uc
h
as;
sal
ie
ncy
vid
e
o
detect
io
n
m
e
tho
d
base
d
on
sp
at
ia
l
-
te
m
po
ral
featur
e
s
a
nd
3D
co
nvol
utio
n
netw
ork
[9
]
,
te
m
po
ral
se
gm
ent
networ
k
[
10]
,
a
uto
e
nc
od
e
r
netw
ork
-
base
d
m
ot
ion
patte
r
n
le
arn
in
g
[
11
]
,
end
to
en
d
an
om
al
y
detect
ion
of
vid
e
o
base
d
on
pre
-
trai
ned
dee
p
netw
orks
[
12
]
,
deep
a
nom
aly
detect
ion
with
dev
ia
ti
on
net
w
orks
[
13]
.
H
oweve
r,
dee
p
le
arn
i
ng
m
et
ho
ds
are
re
quired
a
huge
dataset
sa
m
ple
to
ob
ta
in
an
a
ccur
at
e
pe
rform
ance
m
easur
e
fo
r
detect
in
g
ano
m
al
y
even
ts
within
vid
e
o
fr
am
es.
Othe
r
resea
rchers
f
or
m
ulate
d
the
pro
blem
of
vi
deo
a
nom
a
ly
detect
ion
as
a
regressio
n
m
od
el
to
prov
i
de
an
ano
m
al
y
scor
e
by
fr
am
e
or
cl
ip
le
vel
[14
]
,
[
15
]
.
Sim
ultaneousl
y,
ot
her
researc
he
r
s
hav
e
i
nteres
te
d
by
trajecto
ry
-
base
d
vid
e
o
a
no
m
al
y detec
ti
on
m
et
hods
[16
]
,
[
17
].
T
r
a
j
e
c
t
o
r
y
b
a
s
e
d
m
e
t
h
o
d
s
i
n
c
l
u
d
e
l
e
a
r
n
i
n
g
t
h
e
n
o
r
m
a
l
t
r
a
j
e
c
t
o
r
i
e
s
o
f
p
e
d
e
s
t
r
i
a
n
w
i
t
h
a
s
s
i
s
t
a
n
c
e
o
f
t
r
a
c
k
i
n
g
a
l
g
o
r
i
t
hm
s
t
o
d
e
c
i
d
e
w
h
e
t
h
e
r
t
h
e
v
i
d
e
o
n
o
r
m
a
l
o
r
a
n
o
m
a
l
y
.
T
h
e
m
a
i
n
o
b
s
t
a
c
l
e
o
f
t
r
a
j
e
c
t
o
r
y
-
b
a
s
e
d
m
e
t
h
o
d
s
i
s
t
h
e
i
n
f
l
u
e
n
c
e
o
f
o
c
c
l
u
s
i
o
n
e
s
p
e
c
i
a
l
l
y
i
n
c
r
o
w
d
e
d
s
c
e
n
e
s
a
s
w
e
l
l
a
s
t
h
e
l
i
k
e
l
i
h
o
o
d
o
f
e
x
i
s
t
i
n
g
d
i
f
f
e
r
e
n
t
t
r
a
j
e
c
t
o
r
i
e
s
f
o
r
s
i
n
g
l
e
s
c
e
n
e
[
1
8
]
.
L
o
c
a
l
a
nd
g
l
o
b
a
l
a
n
o
m
a
l
y
d
e
t
e
c
t
i
o
n
m
e
t
h
o
d
u
s
i
n
g
h
i
e
r
a
r
c
h
i
c
a
l
f
e
a
t
u
r
e
r
e
p
r
e
s
e
n
t
a
t
i
o
n
a
n
d
G
a
u
s
s
i
a
n
p
r
o
c
e
s
s
r
e
g
r
e
s
s
i
o
n
w
a
s
i
nt
r
o
d
u
c
e
d
i
n
[
1
9
]
,
w
h
e
r
e
g
l
o
b
a
l
a
n
o
m
a
l
i
e
s
r
e
f
e
r
t
o
t
h
e
a
n
om
a
l
i
e
s
a
m
on
g
s
e
q
u
e
n
c
e
f
r
a
m
e
s
a
n
d
l
o
c
a
l
a
n
o
m
a
l
i
e
s
d
e
n
o
t
e
t
h
e
a
n
om
a
l
o
u
s
r
e
g
i
o
n
s
w
i
t
h
i
n
a
f
r
a
m
e
.
L
o
c
a
l
a
n
om
a
l
i
e
s
a
r
e
c
om
m
o
nl
y
d
e
t
e
c
t
e
d
i
n
v
i
d
e
o
s
b
a
s
e
d
o
n
l
o
c
a
l
s
p
a
t
i
o
-
t
e
m
p
o
r
a
l
f
e
a
t
u
r
e
s
,
w
h
e
r
e
m
o
t
i
o
n
i
s
o
c
c
u
r
r
i
n
g
a
n
d
g
e
n
e
r
a
t
i
n
g
d
u
e
t
o
m
u
l
t
i
p
l
e
o
b
j
e
c
t
s
m
o
vi
n
g
w
i
t
h
i
n
s
i
n
g
l
e
s
c
e
n
e
[
2
0
]
.
V
i
d
e
o
a
n
o
m
a
l
y
de
t
e
c
t
i
o
n
m
e
t
h
o
d
s
a
r
e
i
n
t
e
r
e
s
t
e
d
i
n
d
e
t
e
r
m
i
ni
n
g
w
h
e
t
h
e
r
t
h
e
c
u
r
r
e
n
t
f
r
a
m
e
o
f
a
g
i
v
e
n
v
i
d
e
o
d
e
m
o
n
s
t
r
a
t
e
s
a
n
a
n
om
a
l
y
o
r
n
o
t
[
2
0
]
,
[
2
1
]
.
Anothe
r
ty
pe
of
a
nom
alies
nam
ed
con
te
xtu
al
an
om
al
y
wh
ic
h
c
orresponds
t
o
the
s
a
m
ples
hav
in
g
sign
ific
a
nt
va
r
ia
ti
on
causi
ng
ano
m
al
ie
s
relat
ed
to
certa
in
co
ntext
[
20]
.
Co
ntextual
ano
m
al
ie
s
c
an
be
cat
egorized
ba
sed
on
sp
at
ia
l
and
te
m
po
ral
f
eat
ur
es
of
vi
de
o
fr
am
es
[22
]
,
[
23]
.
O
utli
ers
detect
ion
m
e
thod
s
base
d
on
cl
us
t
erin
g
al
go
rith
m
s
are
insp
ire
d
by
m
any
ear
ly
researc
her
s
to
ef
fici
ently
de
te
ct
ou
tl
ie
rs
ba
sed
on
the
entire
data
set
[24].
The
pro
posed
w
ork
in
[
25
]
ha
ndle
d
the
ano
m
al
y
detect
ion
pro
ble
m
as
low
li
kelihoo
d
patte
rn
detect
ion
due
to
li
m
i
t
ed
avail
a
bili
ty
of
a
de
qu
at
e
in
s
ta
nces
of
a
no
m
al
y
even
ts
with
assist
ance
of
ne
ur
al
netw
ork
f
or
norm
al
even
ts
le
arn
i
ng.
L
u
et
al.
[
26]
,
sug
ge
ste
d
a
dicti
on
ary
-
ba
sed
a
ppr
oach
t
o
le
ar
n
norm
al
beh
a
viors
an
d
detect
ano
m
al
y
even
ts
us
ing
r
eco
ns
tr
uct
ed
erro
r.
H
oweve
r,
this
m
et
hod
is
no
t
rob
us
t
adequate
ly
to
disti
nguish
bet
ween
norm
al
and
ab
norm
al
eve
nts
bas
ed
on
the
rec
onstructio
n
e
rror.
Th
e
pro
po
se
d
work
in
[27]
prese
nt
ed
a
le
arn
i
ng
m
od
el
of
norm
al
scenes
us
i
ng
two
-
stream
recu
r
ren
t
a
uto
e
nc
od
e
r
in
a
sem
i
su
pervised
le
a
rn
i
ng
m
ann
er
a
nd
tra
j
ect
ory
-
ba
sed
s
patio
-
te
m
po
ral
feature
s.
Anot
her
w
ork
[
28]
base
d
on
a
ut
oen
c
ode
r
net
wor
k
an
d
sp
at
ia
l
featu
res
for
le
ar
ning
norm
al
even
ts.
The
m
ai
n
lim
itati
on
of
t
his
m
et
hod
represe
nted
by
it
s
dep
e
nd
e
nc
y
on
s
uccee
ding
the
tra
j
ect
or
y
-
base
d
featu
res
extracti
on
ph
a
se
w
hich
is
a
ffec
t
ed
in cro
wd
e
d
sce
nes
.
The
m
ai
n
obj
e
ct
ive
of
t
his
re
search
is
t
o
de
sign
a
nd
im
ple
m
ent
a
com
pu
te
r
visio
n
syst
e
m
to
address
the
pr
ob
le
m
of
a
no
m
al
y
even
t
detect
io
n
i
n
vid
e
o.
The
Anom
al
y
even
ts
(outie
rs
)
ar
e
detect
ed
in
vide
o
seq
uen
ce
b
ase
d on cl
ip le
vel.
W
e
ca
n sum
m
arize
the e
ssent
ia
l con
tri
bu
ti
on
s of the
pr
opose
d
m
et
ho
d by
;
Determ
in
ing
a
uto
m
at
ic
ally
th
e
global
an
d
lo
cal
ou
tl
ie
rs
in
vid
e
o
ba
sed
on
tem
po
ral
re
dund
a
ncy
of
f
ra
m
es
and d
e
ns
e
cli
ps cl
us
te
rin
g.
Fu
rt
her,
trai
ni
ng
a
re
gr
essi
on
m
od
el
can
e
stim
at
e
the
g
lo
bal
an
om
al
y
scor
e
for
the
e
nt
ire
vid
e
o
f
ra
m
es,
hen
ce
d
et
ect
s t
he
a
no
m
al
ou
s
po
te
ntial
f
ram
es w
it
hi
n
eac
h
c
li
p
in term
o
f
local o
utli
er
detect
ion
.
In
a
c
on
ti
nu
ou
s
m
ann
er,
the
t
raine
d
regressi
on
m
od
el
is
updated
with
no
r
m
al
segm
ents
wh
ic
h
detect
ed
in
the te
sti
ng m
od
e in
or
der
t
o
t
ake a
dv
a
nta
ge of
nor
m
al
b
eha
viour
v
a
riat
ion
.
T
h
e
r
e
s
t
o
f
t
h
i
s
p
a
p
e
r
i
s
o
r
g
a
n
i
z
e
d
a
s
f
o
l
l
ow
s
:
s
e
c
t
i
o
n
s
2
a
n
d
3
p
r
e
s
e
n
t
t
h
e
d
e
t
a
i
l
d
e
s
c
r
i
p
t
i
o
n
o
f
a
n
om
a
l
y
e
v
e
n
t
d
e
t
e
c
t
i
o
n
a
n
d
l
o
c
a
l
i
z
a
t
i
o
n
m
e
t
h
o
d
o
f
v
i
d
e
o
c
l
i
p
s
u
s
i
n
g
gl
o
b
a
l
a
n
d
l
o
c
a
l
o
u
t
l
i
e
r
s
.
T
h
e
r
e
s
ul
t
s
a
n
d
d
i
s
c
u
s
s
i
o
n
a
r
e
i
l
l
u
s
t
r
a
t
e
d
i
n
s
e
c
t
i
o
n
4
.
S
e
c
t
i
o
n
5
c
o
n
c
l
u
d
e
s
t
h
i
s
p
a
p
e
r
.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
An
omaly eve
nt d
et
ect
io
n and l
oca
li
za
ti
on
of
vi
deo
cli
ps
us
i
ng
… (
Sawsen
Ab
du
l
hadi M
ahm
ood
)
1065
2.
THE
PROPO
SED
METHO
D
In
t
his
sect
io
n,
we
will
dem
on
strat
e
th
e
pro
posed
m
et
hod
of
an
om
al
y
even
t
de
te
ct
ion
a
nd
local
iz
at
ion
al
ong
vid
e
o
f
ra
m
es
based
on
glob
al
an
d
loc
al
ou
tl
ie
r
pre
di
ct
ion
as
s
how
n
in
Fi
g
ure
1.
In
t
his
pap
e
r,
a
m
ult
ilevel
fr
am
ewo
r
k
is
adopted
to
determ
ine
the
ano
m
al
y
behavio
ur
(
outl
ie
rs)
in
a
giv
en
vid
e
o
sam
ple.
First
le
vel
is
re
pr
e
se
nted
by
est
im
a
ti
on
the
te
m
po
ral
dif
fer
e
nces
betwee
n
se
que
nced
f
ram
es
based
on
visu
al
featu
res suc
h
as m
otion a
nd
histo
gr
am
sim
il
arit
y featu
res
.
Seco
nd
le
vel
m
i
m
ic
s
the
det
ect
ion
of
gl
ob
a
l
ou
tl
ie
rs
us
i
ng
tem
po
ral
dif
fe
ren
ce
sco
re
a
nd
po
ly
nom
ia
l
fitt
ing
m
od
el
.
To
this
e
nd,
a
n
ass
um
ption
of
t
he
re
gr
e
ssion
m
od
el
is
presente
d
a
nd
t
raine
d
to
s
peci
fy
the
po
ly
nom
ia
l
fi
tt
ing
wit
h
the
tem
po
ral
di
ff
e
r
ence
ra
w
data
of
the
e
ntire
vid
e
o
seq
ue
nc
e.
Con
se
que
ntly
,
the
sam
ples
that
dev
ia
te
f
ro
m
the
re
gr
essi
on
m
od
el
are
c
and
i
date
to
be
global
outl
ie
rs.
Accor
dingly
,
the
su
ggest
e
d
global
outl
ie
r
det
ect
ion
m
od
el
is
trai
ne
d
us
in
g
on
ly
no
rm
al
vid
e
os
with
norm
al
beh
avi
ours
t
o
pr
e
dict
the
optim
al
po
ly
no
m
i
al
coef
fici
ents
.
W
e
assum
e
t
hat
an
insta
nt
segm
ent
has
a
norm
al
/a
bn
orm
al
beh
a
viou
r
bas
ed
on
the
pr
e
dicti
on
sc
or
e
of
the
ad
opte
d
regressi
on
m
od
el
instea
d
of
us
i
ng
pr
e
def
i
ned
thres
ho
l
d.
I
n
th
is
le
vel,
t
he
vid
e
o
f
ram
es
are
se
gm
ented
into
set
of
pr
i
m
itive
cl
ips
or
shots
isolat
ed
by
th
e
corres
pondin
g
ind
e
xes
of
temporal
fr
am
es
diff
e
ren
ces
.
The
global
ou
tl
ie
rs
are
ex
plo
it
ed
to
segm
ent
the
vid
e
o
fr
am
es into
cl
ips
.
F
i
n
a
l
l
y
,
t
h
e
l
o
c
a
l
o
u
t
l
i
e
r
s
’
d
e
t
e
c
t
i
o
n
l
e
v
e
l
i
s
i
m
p
l
e
m
e
n
t
e
d
w
h
e
r
e
t
h
e
c
o
n
t
e
x
t
u
a
l
o
u
t
l
i
e
r
s
w
i
t
h
i
n
e
a
c
h
v
i
d
e
o
c
l
i
p
w
a
s
t
a
k
e
n
i
n
o
u
r
c
o
n
s
i
d
e
r
a
t
i
o
n
.
P
r
a
c
t
i
c
a
l
l
y
,
t
h
e
f
r
a
m
e
w
o
r
k
o
f
a
n
e
f
f
i
c
i
e
n
t
a
n
o
m
a
l
y
e
v
e
n
t
d
e
t
e
c
t
o
r
s
h
o
u
l
d
b
e
e
l
i
g
i
b
l
e
t
o
u
p
d
a
t
e
i
t
s
d
e
t
e
r
m
i
n
a
t
i
o
n
o
f
n
o
r
m
a
l
e
v
e
n
t
s
i
n
a
c
o
n
t
i
n
u
o
u
s
m
a
n
n
e
r
.
T
h
u
s
,
t
h
e
r
e
c
e
n
t
a
p
p
r
o
a
c
h
i
s
a
b
l
e
t
o
u
p
d
a
t
e
t
h
e
r
e
g
r
e
s
s
i
o
n
m
o
d
e
l
t
h
r
o
u
g
h
f
e
e
d
i
n
g
t
h
e
d
e
t
e
c
t
e
d
n
o
r
m
a
l
s
e
g
m
e
n
t
s
t
o
t
h
e
t
r
a
i
n
i
n
g
s
a
m
p
l
e
s
f
r
o
m
a
n
y
a
c
q
u
i
r
e
d
v
i
d
e
o
.
F
i
g
u
r
e
1
i
l
l
u
s
t
r
a
t
e
s
t
h
e
w
o
r
k
f
l
o
w
o
f
t
h
e
p
r
o
p
o
s
e
d
a
n
o
m
a
l
y
e
v
e
n
t
d
e
t
e
c
t
i
o
n
m
e
t
h
o
d
.
Figure
1.
Wo
r
kfl
ow
of
t
he pr
opose
d an
om
al
y
eve
nt d
et
ect
io
n
m
et
ho
d
2.1
.
Te
mp
or
al
f
r
ames
d
if
f
erence
s
e
s
timat
i
on
As
a p
re
-
proce
ssing
of
v
i
deo
analy
sis, w
e segm
ent the v
ideo
into ind
i
vidu
al
f
ram
es
an
d
conve
rt each
fr
am
e to Y
Cb
Cr for
m
at
in
order
t
o
e
xtract
t
he
lum
inance l
ay
er.
I
n
t
his p
a
per, an ef
fici
en
t underst
an
ding
of
t
he
vid
e
o
cl
ip
be
ha
viours
has
be
en
pro
vide
d
by
form
ulati
ng
the
tem
po
ral
rele
van
ce
am
ong
s
equ
e
nce
d
f
ram
es.
I
n
order
to
detect
an
d
e
xtract
t
he
f
or
e
gro
und
obj
ect
s
(m
ov
in
g
obje
ct
s)
over
the
vid
e
o
se
qu
ence,
the
te
m
po
ral
diff
e
re
ncin
g
be
tween eac
h
tw
o
co
ns
ec
utive
f
ram
es is est
i
mate
d.
T
he
te
m
po
ral
fr
am
es d
if
fer
e
nces a
re ob
ta
ined
base
d
on
t
wo
m
ai
n
integrate
d
par
am
et
ers;
m
ot
ion
est
im
ation
a
nd
c
olour
histo
gr
am
dif
fer
e
nce.
The
m
ot
ion
par
am
et
er
bet
ween
f
r
am
es
F
i
,
F
i
-
1
is
ob
ta
ined
t
hro
ugh
bin
a
rizat
ion
t
he
subtract
e
d
i
m
age
(
F
i
-
F
i
-
1
)
with
a
Th
res
hold
val
ue
that
is
s
pecifi
ed
acc
ordin
g
to
th
e
inte
ns
it
y
range
of
the
s
ubtract
ed
im
age
(
F
i
-
F
i
-
1
)
a
nd
Otsu
'
s
m
et
ho
d [
2
9
]
as
il
lustrate
d
i
n
(
1)
:
=
|
−
−
1
|
>
ℎ
ℎ
(1)
Me
anwhil
e,
th
e
foregr
ound
obj
ect
s
are
e
xt
racted
an
d
la
belle
d
usi
ng
(
8
–
c
onnecte
d)
neig
hbouri
ng
pix
el
s
of
M
O
i
i
m
age
w
hich
re
pr
ese
nts
the
f
oreg
rou
nd
ob
j
ec
ts
in
the
c
urre
nt
fr
am
e
F
i
,
whe
re
i=
2,3,….N
,
an
d
N
is
the
total
n
um
ber
of
f
ram
es
in
the
acq
uire
d
vid
e
o
sam
ple.
A
fter
ward,
th
e
area p
r
operty
AR
i
is
e
xtracte
d
f
r
om
MO
i
data
po
i
nt
s
to
highli
ght
the
interest
e
d
reg
i
on
s
,
m
ini
m
iz
ing
the
co
m
pu
ta
ti
on
tim
e
as
well
as
to
m
ake
a
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
2
4
, N
o.
2
,
N
ove
m
ber
20
21
:
106
3
-
107
3
1066
decisi
on
a
bout
occ
urrin
g
m
otion
betwe
e
n
th
e
seq
ue
nce
fr
a
m
es.
W
e
a
ssum
ed
that
the
hi
gh
e
r
values
of
AR
i
m
easur
ed
in pi
xels (w
hite pix
el
s)
re
ferred
to m
ot
ion
ca
pturing bet
ween t
he
seque
nce fram
es.
The
sec
ond
pa
ram
et
er
us
ed
in
our
f
ram
ewor
k
f
or
te
m
po
ral
fr
am
es
dif
fer
e
nces
e
stim
ation
i
s
represe
nted
by
the
sim
i
la
rity
ben
c
hm
ark
com
m
on
ly
use
d
i
n
m
achine
le
ar
ning
a
nd
data
analy
sis.
T
he
Hau
s
dor
ff
dist
ance
(
HD
)
dis
ta
nce
m
easur
e
m
ent
[
30
]
is
a
li
gh
twei
gh
t
ye
t
eff
ect
ive
sim
il
arity
m
eas
ur
e
f
or
qu
a
ntifyi
ng
th
e
pro
xim
it
y
of
two
grap
hs
.
The
Hausd
orff
(
HD
)
dista
n
c
e
m
easur
e
qu
a
ntifie
s
the
dif
f
eren
c
e
betwee
n
two
s
equ
e
nce
d
fr
am
es
based
on
bl
ock
-
based
c
olour
hist
ogram
d
iffer
e
nces.
T
he
tem
po
ral
diff
e
ren
ce
betwee
n
eac
h
s
equ
e
nce
fr
am
es
F
i
, F
i
-
1
us
in
g
Ha
us
do
rff (
HD
)
distance
m
e
asur
e
is c
om
pu
te
d
acco
r
ding t
o (2)
:
=
∑
(
,
−
1
)
(2)
wh
e
re
Diff
i
para
m
et
er
denotes
the
weig
hte
d
s
um
of
blo
c
ks
di
ff
ere
nces
bet
ween
seq
ue
nce
fr
am
es
F
i
and
F
i
-
1
and
k
is
t
he
c
orrespo
nd
i
ng
bl
oc
k
in
de
x
in
the
two f
ram
es
F
i
,
F
i
-
1
.
A
li
near
i
nteg
rati
on
of
the
t
wo
pa
ram
et
ers
(
AR
i
a
nd
Diff
i
)
is
sp
eci
fie
d
to
f
or
m
ulate
t
he
sc
or
e
of
tem
po
ral
diff
e
r
ences
betwee
n ea
ch
tw
o seq
ue
nced f
ram
es
as g
ive
n
i
n
(
3):
=
+
(
1
−
)
(3)
wh
e
re
α
is
a
scal
ar
sta
nd
at
0.6
an
d
it
is
dete
rm
ined
throu
gh
the
e
xp
e
rim
e
nts.
O
bvio
us
ly
,
area
featu
re
AR
is
aug
m
ented
by
scal
ar
α
due
t
o
it
s
eff
ect
ive
ne
ss
of
ref
le
ct
in
g
the
dissim
il
ari
ty
between
tw
o
se
quence
f
ra
m
es.
The
te
m
po
ral
diff
e
re
nces
est
i
m
ation
strat
e
gy
and
it
s
visu
a
li
zat
ion
res
ult
of
sam
ple
vid
e
o
are
il
lustrate
d
in
Figure
2
a
nd Fi
gure
3
res
pecti
vely
.
Figure
2
.
Bl
oc
k diag
ram
o
f
te
m
po
ral fr
am
e d
iffe
re
nces esti
m
at
ion
strateg
y
(a)
(b)
Figure
3
.
V
is
ua
li
zat
ion
r
es
ult o
f
te
m
po
ral f
r
a
m
e d
iffe
ren
ce
s estim
at
ion
b
a
sed o
n;
(
a
)
area
(blue li
ne
)
a
nd
HD
distance m
easu
re (re
d
li
ne)
;
(
b
)
Tem
pD
iff
sc
or
e
of
vid
e
o
sa
m
ple
(0
4) from
AVEN
U data
s
et
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
An
omaly eve
nt d
et
ect
io
n and l
oca
li
za
ti
on
of
vi
deo
cli
ps
us
i
ng
… (
Sawsen
Ab
du
l
hadi M
ahm
ood
)
1067
3.
MO
DELL
IN
G NOR
M
AL
BE
HAV
I
OUR
AND A
NOMAL
Y
S
COR
E
3.1
.
Gl
obal o
ut
li
er detec
tio
n model
Ba
sed on
t
he
a
ssu
m
ption
; a d
at
a sa
m
ple is con
si
der
e
d
as
g
l
ob
al
outl
ie
r
if it
d
eviat
es s
uffici
ently
f
ro
m
the
w
hole
data
set
,
the
first
le
vel
of
detect
in
g
ou
tl
ie
rs
is
re
al
iz
ed
base
d
on
e
xtracti
ng
th
e
global
ou
tl
ie
r
usi
ng
the
re
gr
essi
on
m
od
el
.
Re
gr
ession
a
naly
sis
is
a
fun
dam
ental
an
d
com
m
on
con
ce
pt
adopted
in
m
achine
le
arn
in
g
m
od
el
s.
S
pecifica
ll
y,
it
fall
s
under
su
pe
r
vised
le
ar
ning
i
n
wh
ic
h
the
trai
ni
ng
m
od
el
is
trai
ne
d
us
in
g
input
pr
op
e
rtie
s
an
d
ou
t
pu
t
l
abels.
It
assist
s
in
in
vestigat
ing
a
n
d
est
a
blishin
g
the
rele
van
ce
of
t
he
m
od
el
var
ia
bles th
rou
gh guessi
ng ho
w on
e
v
a
riable
influ
e
nces t
he othe
r.
In
the
case
of
occurri
ng
a
bru
pt
il
lu
m
inati
on
change
or
s
udden
obj
ect
m
ot
ion
al
ong
vi
de
o
seq
uen
ce
,
we
need
to
la
be
l
the
c
orrespo
nd
i
ng
vi
de
o
f
r
a
m
es
with
a
ppropr
ia
te
outl
ie
r
sco
re.
I
n
this
con
te
xt,
a
no
n
-
li
near
regressio
n
m
od
el
is
em
plo
ye
d
to
ac
hieve
a
po
ly
no
m
ia
l
fitting
process
wi
th
the
ra
w
data
of
te
m
po
ral
fra
m
e
s
diff
e
re
nces
Te
m
pD
iff.
Co
ns
e
qu
e
ntly
,
the
no
rm
al
beh
avio
ur
an
d
the
gl
ob
al
outl
ie
rs
of
the
entire
vid
e
o
fra
m
es
are
est
i
m
a
te
d
and
detect
ed
r
especti
vely
.
A
s
a
con
se
qu
e
nc
e,
global
ou
tl
ie
rs
are
detect
ed
by
le
ver
a
gi
ng
f
r
om
the
resi
dual
er
rors
wh
ic
h
s
uf
fici
ently
descri
bed
data
de
viati
on
from
the
fitt
ed
po
ly
no
m
ia
l
cur
ve.
H
ence
,
pr
e
dicti
ng
th
e
set
of
pote
ntia
l
values
of
po
ly
no
m
ial
coef
f
ic
ie
nts
will
info
rm
us
of
how
ap
pro
pr
ia
te
ly
our
current
m
od
el
is
able
to
des
cribe
the
ra
w
da
ta
beh
avi
our.
The
m
ai
n
idea
of
ad
opti
ng
a
regressio
n
m
od
el
for
global
ou
tl
ie
r
detect
ion
in
t
he
w
ho
le
vi
deo
fr
am
es
is
ba
s
ed
on
the
ass
um
ption
that
t
he
te
m
po
ral
fra
m
es
diff
e
re
nces
ca
n
be
represe
nted
by
no
n
-
li
ne
ar
po
ly
nom
ia
l
m
od
el
.
The
ou
tl
ie
r
s
are
th
os
e
tem
po
ral
dif
f
eren
c
e
instants
that
ar
e
no
t
well
re
presente
d
by
th
e
regressio
n
m
od
el
.
In
t
his
le
vel,
the
s
udde
n
or
ab
rupt
ch
ang
e
s
al
ong
vi
de
o
f
ra
m
es
are
detect
ed
w
hich
co
rr
e
sp
on
de
d
to
the
diff
e
re
nces
sa
m
ples
that
far
su
f
fici
ently
fro
m
the
interva
l
[
p+2∆
,
p
-
2∆
]
,
wh
e
re
p
re
pr
ese
nts
t
he
po
ly
nom
ia
l
coe
ff
ic
ie
nts
a
nd
∆
ref
e
rs
t
o
the
resi
dual
er
ror.
I
n
this case, t
he fr
a
m
es ly
ing
out
of this
range a
r
e label
le
d
as
g
l
ob
al
outl
ie
r
f
ra
m
es as sho
wn in F
i
gure
4.
Hen
ce
f
or
th
,
th
e
pr
op
os
e
d
fr
a
m
ewo
r
k
of
gl
ob
al
ou
tl
ie
r
de
te
ct
ion
ta
sk
is
designed
as
a
reg
res
sio
n
m
od
el
fitting
,
wh
ic
h
co
ns
ide
rs
a
certai
n
segm
ent
of
the
vid
eo
has
a
bnor
m
al
beh
avio
ur
based
on
re
gressi
on
pr
e
dicti
on
sco
r
e
instea
d
of
a
doptin
g
a
pre
de
fine
d
th
res
ho
l
d
for
detect
ing
ano
m
al
y.
Wh
e
n
the
re
a
re
a
nom
al
y
changes
,
the
ano
m
al
y
scor
e
ascents
si
gn
i
ficantl
y.
It
w
or
t
h
noti
ng
t
ha
t
the
detect
e
d
global
outl
ie
rs
are
exp
l
oited
to
autom
at
ic
ally
s
egm
ent
vid
eo
fr
am
es
into
prim
i
ti
ve
sh
ots
or
cl
ips,
ye
t
to
find
an
d
local
iz
e
the
local
ou
tl
ie
rs
within
eac
h
cl
ip.
Nei
ghbo
ur
fr
am
es
of
the
detect
ed
glob
al
ou
tl
ie
r
fr
am
es
are
extracte
d
an
d
form
ulate
d
into
disti
ng
uish
i
ng
re
gions
a
nd yet
add
e
d
to
t
he
m
e
m
or
y
bu
ffer
to
c
ollec
t
the
ou
tl
ie
r
reg
i
ons
al
ong
vid
e
o
sam
ple.
Fig
ure
5
s
how
s
the
detect
ed
global
outl
ie
rs
con
stric
te
d
by
the
pr
im
itive
sh
ots
bounda
ri
es
of
vid
e
o
sam
ples 0
4,
06 take
n from
AV
EN
U
da
ta
set
.
Figure
4. Gl
obal
o
utli
er
detec
ti
on
base
d on poly
no
m
ia
l fit
tin
g st
rategy
of
vid
e
o
sam
ple (04)
from
AV
E
NU
(a)
(b)
Figure
5.
Dete
ct
ion
of cli
ps
boun
dar
ie
s
of:
(a
) vide
o
sam
ple 04, (
b)
vid
e
o sam
ple 0
6
f
r
om
AV
ENU
da
ta
set
,
X
-
a
xis
represe
nts fram
es ind
e
xes;
Y
-
a
xis
de
no
te
s
the tem
po
ral
dif
fer
e
nce
s scores
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
2
4
, N
o.
2
,
N
ove
m
ber
20
21
:
106
3
-
107
3
1068
3.2.
L
oca
l
outli
ers detec
tio
n
L
o
c
a
l
o
u
t
l
i
e
r
s
c
a
n
b
e
d
e
f
i
n
e
d
a
s
t
h
e
p
a
t
t
e
r
n
s
t
h
a
t
p
r
om
i
n
e
n
t
f
r
o
m
t
h
e
i
r
n
e
i
g
h
b
o
u
r
h
o
o
d
i
n
a
s
p
e
c
i
f
i
c
d
a
t
a
r
e
g
i
o
n
[
2
4
]
.
In
t
h
i
s
s
e
c
t
i
o
n
,
w
e
w
i
l
l
e
x
p
l
a
i
n
h
o
w
t
o
d
e
t
e
c
t
t
he
l
o
c
a
l
o
u
t
l
i
e
r
s
i
n
t
h
e
a
c
q
u
i
r
e
d
v
i
d
e
o
s
a
m
p
l
e
b
a
s
e
d
o
n
c
l
i
p
s
d
e
n
s
i
t
y
.
T
h
e
v
i
d
e
o
f
r
a
m
e
s
a
r
e
s
e
gm
e
n
t
e
d
a
ut
om
a
t
i
c
a
l
l
y
i
n
t
o
n
o
n
-
o
v
e
r
l
a
p
p
e
d
K
c
l
i
p
s
o
r
s
h
o
t
s
b
a
s
e
d
o
n
t
h
e
g
l
o
b
a
l
o
u
t
l
i
e
r
d
e
t
e
c
t
i
o
n
p
r
o
c
e
s
s
i
l
l
u
s
t
r
a
t
e
d
i
n
t
h
e
p
r
e
v
i
o
u
s
s
e
c
t
i
o
n
.
W
e
c
o
p
e
w
i
t
h
e
a
c
h
s
e
gm
e
nt
e
d
c
l
i
p
a
s
c
l
u
s
t
e
r
w
i
t
h
v
a
r
i
a
b
l
e
s
i
z
e
S
a
n
d
a
s
s
um
e
d
t
h
e
a
v
e
r
a
g
e
s
c
o
r
e
o
f
t
e
m
p
o
r
a
l
f
r
a
m
e
d
i
f
f
e
r
e
n
c
e
o
f
e
a
c
h
c
l
u
s
t
e
r
i
s
t
he
c
e
n
t
r
e
Co
m
o
f
c
l
u
s
t
e
r
C
m
.
A
f
t
e
r
w
a
r
d
,
t
h
e
d
e
n
s
i
t
y
o
f
e
a
c
h
c
l
u
s
t
e
r
i
s
d
e
t
e
r
m
i
n
e
d
b
a
s
e
d
o
n
t
h
e
m
e
a
n
d
i
s
t
a
nc
e
Md
b
e
n
c
hm
a
r
k
b
e
t
w
e
e
n
e
a
c
h
s
a
m
p
l
e
x
j
b
e
l
o
n
g
s
t
o
m
th
c
l
u
s
t
e
r
a
nd
i
t
s
c
e
n
t
r
o
i
d
Co
m
a
c
c
o
r
d
i
n
g
t
o
(
4
)
:
(
)
=
1
∑
‖
−
‖
(4)
wh
e
re
m
=
1,
.
K
,
an
d
K
is
t
he
nu
m
ber
of
the
detect
ed
cl
ips
or
cl
us
te
rs
.
Ba
se
d
on
t
he
m
ean
dista
nce
Md
m
entioned
i
n
(
4),
a
lower
value
of
Md
re
fers
to
pr
ese
nt
a
dense
cl
us
te
r
wh
il
e
a
higher
value
of
Md
i
ndic
at
es
that
the
cl
us
te
r
is
scat
te
red
or
sp
a
rse.
Since
the
Md
value
is
inv
ersely
relat
ed
with
the
cl
us
te
r
densi
ty
,
the
densi
ty
DS
of t
he
cl
ust
er
C
m
i
s ide
ntifie
d
t
o be the
in
ver
se
of
Md
(
C
m
)
as
il
lustrate
d
in
(5)
:
(
)
=
1
(
)
(5)
Accor
ding
to
c
luster
de
ns
it
y
def
i
niti
on
,
t
he
cl
us
te
r
with
hi
gh
e
r
de
ns
it
y
le
ads
to
be
it
s
da
ta
el
e
m
ent
s
are
cl
os
e
r
t
o
it
s
centr
oid.
In
con
t
rast,
lo
w
-
de
ns
it
y
cl
us
te
rs
te
nd
to
ha
ve
data
el
em
ents
on
ave
ra
ge
f
ar
from
cl
us
te
r
ce
ntre.
Re
gards
to
loc
al
ou
tl
ie
r’s
d
et
ect
ion
ta
s
k,
w
e
ha
ve
sel
ect
e
d
the
hi
gh
e
r
de
ns
it
y
cl
us
te
rs
w
hich
include
d
den
se
ou
tl
ie
rs
an
d
re
pr
ese
nt
ed
a
ca
nd
i
date
an
om
a
ly
even
t
with
r
espect
to
the
w
ho
le
vi
de
o
seq
uen
c
e
as
sh
ow
n
in
Figure
6.
The
c
orrespo
nd
i
ng
f
ra
m
es
of
this
ty
pe
of
cl
us
te
r
s
ar
e
la
belle
d
as
ano
m
al
y
even
t
fr
am
es.
The
sel
ect
ion
be
nch
m
ark
is
ba
sed
m
ai
nly
on
a
pr
e
def
i
ned
th
reshold
t
o
ta
ke
a
decisi
on
a
bout
lowe
r
an
d
hi
gh
e
r
densi
ty
cl
us
te
r
s
ta
king
i
n
our
co
ns
ide
rati
on
it
s
m
ean
va
lue
.
Furthe
r,
t
he
c
entr
oid
Co
of
lowe
r
de
ns
it
y
cl
us
te
r;
i
.e
(
DS
≈
0)
is
consi
der
e
d
t
o
be
sin
gula
r
outl
ie
r
within
cl
ust
er
or
cl
ip
a
nd
the
co
rr
es
po
nding
fr
am
e
is
la
belle
d
as an
om
al
y fr
am
e as il
lustrate
d
in
Fig
ure
6.
Figure
6
.
Dem
on
st
rati
on r
es
ul
ts of
si
ngular
and colle
ct
ive
ano
m
al
ie
s f
ra
m
es captur
e
d f
ro
m
v
ide
o
sam
ple (0
4)
from
AV
E
NU
4.
RES
ULT
S
AND
DI
SCUSSION
In
t
his
sect
io
n,
detai
le
d
descri
ption
of
c
omm
on
even
t
det
ect
ion
dataset
s
ad
op
te
d
in
ou
r
ex
pe
rim
ents
is
pr
ese
nte
d
as
well
pe
rfor
m
ance
e
valuati
on
m
e
tric
s
of
a
no
m
al
y
detect
ion
sco
re
are
com
pu
te
d
a
nd
sim
ulate
d
gr
a
phic
al
ly
.
Two
m
ai
n
publi
c
dataset
are
e
m
plo
ye
d
in
ou
r
ex
per
im
ents
to
eval
uate
the
perform
ance
of
t
he
pro
po
se
d
a
no
m
al
y
detect
ion
ta
sk
in
vid
e
o
se
qu
e
nce
.
I
n
the
trai
ni
ng
m
od
e,
vi
deo
sam
ples
with
norm
al
beh
a
viou
r
are
dep
l
oyed
in
ou
r
exp
e
rim
ents
i
n
orde
r
to
est
im
at
e
the
reg
re
ssion
m
od
el
coef
fici
ents.
T
he
t
est
ing
m
od
e
is
sta
nd
i
ng
to
validat
e
t
he
occurri
ng
of
outl
ie
r’
s
sco
r
es
in
the
acq
ui
red
vi
deo
sam
ple.
Fi
gure
7
presents
so
m
e
fr
am
es
e
xam
ples
of
no
rm
al
and
anom
al
y
beh
avio
ur
ta
ke
n
from
UCSD
[
31
]
an
d
ave
nu
[
26
]
dataset
s
resp
ect
ively
.
Av
e
nue
datase
t
[2
6]
co
ns
is
ts
of
16
trai
ning
vid
e
os
an
d
21
te
sti
ng
vid
e
o.
Eac
h
vi
deo
sa
m
ple
ta
kes
a s
hort i
nt
erv
al
, a
rou
nd (1
-
2) m
inu
te
s wi
th f
ram
e reso
l
ution 6
40×3
60
pix
el
s.
4.1.
UCSD
d
atase
ts
(
Ped
1 and Ped
2)
The
UC
SD
dat
aset
[31]
com
po
ses
of
vid
e
o
s
a
m
ples
of
pedest
rian
wal
kw
a
ys.
The
cr
owd
den
sit
y
in
the
wal
kw
ay
s
was
var
ia
ble
,
r
ang
i
ng
from
sp
arse
t
o
ver
y
c
rowded
.
I
n
the
norm
al
set
t
ing
,
the
vi
deo
c
on
ta
in
s
on
ly
pe
destria
ns
.
Abn
or
m
al
even
ts
a
re
due
to
ei
the
r:
the
ci
rcu
la
ti
on
of
non
-
pe
des
tria
n
entit
ie
s
in
th
e
walk
ways
an
d
ano
m
al
ou
s
pe
de
stria
n
m
otion
patte
rn
s
.
Peds1
dataset
co
nta
ins
34
trai
ning
vid
eo
sam
ples
an
d
36
te
sti
ng
vide
o
sam
ples
with
a
res
olu
ti
on
of
238×
158
pix
el
s.
Peds2
dataset
co
ntains
16
t
rainin
g
vide
o
sam
ples
and
12
te
sti
ng
vid
e
o
sam
p
le
s
with
a
reso
l
utio
n
of
360×
240
pix
el
s.
All
te
sti
ng
s
a
m
ples
are
ass
ociat
ed
with
a
m
anu
al
ly
-
colle
ct
ed
fr
a
m
e
-
le
vel
ground
tr
uth
rele
va
nt
to
abno
rm
al
even
ts
an
nota
ti
on
.
Fig
ure
7
sh
ow
s
fr
am
es exam
pl
es of
norm
al
a
nd anom
al
y events take
n fro
m
A
ven
u an
d UCSD
d
at
aset
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
An
omaly eve
nt d
et
ect
io
n and l
oca
li
za
ti
on
of
vi
deo
cli
ps
us
i
ng
… (
Sawsen
Ab
du
l
hadi M
ahm
ood
)
1069
(a)
(b)
Figure
7. Exa
m
ples o
f
no
rm
al
an
d an
om
al
y
fr
am
es in; (a
) UCSD
-
Ped2
da
ta
set
, (
b)
A
ve
nu
e
d
at
aset
,
the anom
al
y regions a
re
dem
on
strat
ed
b
y
re
d rect
an
gle
4.2.
Im
plem
e
nt
ati
on
d
e
t
ails
The
vi
deo
fr
a
m
es
are
extra
ct
ed
an
d
resiz
ed
into
reso
l
ut
ion
(
320×
200)
pix
el
s,
the
n
c
onve
rt
each
fr
am
e
into
YC
bCr
c
olor
m
odel
to
ex
plo
it
t
he
lum
inance
c
om
po
ne
nt
Y
i
n
the
im
ple
m
entat
ion
f
ram
ewo
r
k.
I
n
order
to
est
im
a
te
the
an
om
al
y
sco
re
f
or
c
urr
ent
f
ram
e
F
i
,
w
e
a
dopt
a
Ha
us
do
rff
distanc
e
m
easur
e
as
well
as
m
ot
ion
fe
at
ur
e
s
betwee
n
t
he
two
se
quence
fr
am
es
(
F
i
,
F
i
-
1
).
T
he
n,
th
e
an
om
aly
scor
e
is
est
i
m
at
ed
bas
ed
on
integrati
ng
the
se
two
feat
ur
e
s
by
us
in
g
li
ne
ar
(
3)
.
T
he
e
sti
m
at
ed
ano
m
aly
scor
es
value
s
are
norm
al
ized
int
o
[0,
10
]
inter
va
l
to
facil
ity
the
com
pu
ta
ti
on
pr
oce
ss.
T
he
norm
al
iz
a
ti
on
te
chn
iq
ue
ad
opte
d
in
this
w
ork
ha
s
been
pe
rfor
m
ed
acc
ordin
g
to
the
f
ollow
i
ng
e
qu
at
io
n:
x
n
=
(
x
-
x
m
)/
std
,
w
here
x
n
re
pr
ese
nts
the
norm
al
iz
ed
data
po
i
nt,
x
m
an
d
st
d
are
t
he
m
ean
and
sta
ndar
d
de
viati
on
of
al
l
data
points
res
pecti
vely
.
The
validat
io
n
w
orkf
l
ow
of
the
pro
pose
d
an
om
al
y
ev
ent
detect
io
n
in
vi
deo
is
c
om
po
sed
of
tw
o
m
ai
n
ph
ases
;
trai
nin
g
a
nd
te
sti
ng
m
od
el
s as ill
us
trat
ed
in
the
f
ollow
in
g
sub
sect
ion
s
.
4.3.
Tr
aining
m
odel
We
purs
ue
to
determ
ine
anom
al
ie
s
in
a
give
n
vid
e
o
se
que
nce
with
lo
we
r
com
pu
ta
ti
on
tim
e
as
well
as
sp
eci
fyi
ng
the
re
qu
i
red
false
al
arm
rate.
Sign
i
ficantl
y,
the
pro
pose
d
ano
m
al
y
detect
i
on
a
ppr
oach
a
s
su
m
es
that
any
ano
m
a
lou
s
patte
r
n
w
ou
l
d
hold
f
or
a
n
obscu
re
pe
riod
o
f
tim
e.
Mean
w
hile,
we
ha
ve
no
pr
e
-
kn
owle
dge
about
the
an
om
al
y
even
t
in
t
he
vid
e
o
se
qu
e
nce.
T
hus,
the trainin
g
m
od
e
i
s
desire
d
t
o
a
ddress
an
d
desc
r
ibe
the
norm
al
beh
avi
our
of
the
fr
a
m
e
sequ
e
nces.
The
t
rainin
g
m
od
e
is
config
ur
e
d
by
sel
ect
ing
the
vid
e
os
sam
ples
V
=
{
vi
:
i
=1,
2,
....,
M
}
with
norm
al
beh
avi
our
to
pe
rfo
rm
the
fitt
in
g
pr
ocess
with
a
prede
fine
d
regr
ession
m
od
el
with
4
th
po
ly
nom
ia
l
e
qu
at
io
n
ha
vi
ng
four
coe
ff
ic
ie
nts
(
b0,
b1,
b2
,
b3
)
an
d
one
tem
po
ral
dif
fe
ren
ce
scor
e
(in
dep
e
ndent
var
ia
ble
).
The
te
m
po
ral
fr
am
es
diff
er
ences
Te
mpDiff
f
or
eac
h
vi
de
o
sam
ple
hav
e
been
com
pu
te
d.
Co
ns
e
qu
e
ntly
,
we
app
li
ed
the
po
ly
no
m
ial
fitting
m
od
el
up
on
the
diff
e
re
nces
values
to
extr
act
the
su
it
able
c
oeffici
ents
an
d
resi
du
al
er
rors
of
the
re
gressi
on
m
od
e
l.
In
this
con
te
xt,
t
he
glo
bal
outl
ie
rs
de
no
t
e
tem
po
ral
diff
e
r
ences i
ns
ta
nts t
hat are
not
well
r
ep
rese
nted b
y t
he
est
im
a
te
d
re
gr
essi
on m
od
el
.
4.4.
Te
s
ting
m
od
el
An
y
ne
w
ra
w
data
of
dif
fer
e
nt
values
obta
ined
from
the
t
e
m
po
ral
di
ff
e
r
entia
l
ph
ase
a
r
e
fitt
e
d
with
the
trai
ne
d
re
gressi
on
m
od
el
to
detect
the
outl
ie
r
in
the
ra
w
data.
The
poly
no
m
ia
l
reg
r
ession
m
od
el
f
it
s
a
curve
li
ne
to
th
e
te
m
po
ral
dif
f
eren
ces
da
ta
as
il
lustrate
d
i
n
F
igure
8.
Global
outl
ie
rs
are
de
te
ct
ed
base
d
on
the
assum
ption
tha
t
ou
tl
ie
rs
a
re
f
ar
aw
ay
from
the
fitt
ed
m
odel
and
ha
ve
a
higher
resi
du
al
error
tha
n
t
he
oth
e
r
data
sam
ples.
The
co
rr
es
pondin
g
fr
am
es
of
this
ty
pe
of
an
om
alies
are
la
belle
d
as
ano
m
a
ly
fr
a
m
e.
Fu
rth
er,
th
e
local
ou
tl
ie
rs
of
the
vid
e
o
sa
m
ple
hav
e
bee
n
detect
ed
at
cl
ip
le
vel
based
on
de
ns
it
y
cl
us
te
ring
strat
e
gy
.
Th
e
corres
pondin
g
fr
am
es
of
t
his
ty
pe
of
cl
us
te
r
s
are
la
belle
d
as
an
om
al
y
fr
am
e
even
t.
The
sel
ect
ion
ben
c
hm
ark
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
2
4
, N
o.
2
,
N
ove
m
ber
20
21
:
106
3
-
107
3
1070
is
base
d
m
ai
nl
y
on
a
prede
fin
ed
th
res
ho
l
d
to
ta
ke
a
decisi
on
a
bout
lo
wer
and
hi
gh
e
r
densi
ty
cl
us
te
rs
ta
king
i
n
our
c
onside
rati
on it
s m
ean
as the cen
t
re
of ea
ch
cl
ust
er.
Figure
8. Vis
ua
li
zat
ion
r
es
ults o
f
the
pr
opose
d
a
no
m
al
y event d
et
ect
io
n m
et
ho
d.
(a
)
ra
w
r
ep
rese
nts tem
po
ral
fr
am
es d
iffe
rence
s,
(b)
ra
w re
pr
ese
nts t
he glob
al
outl
ie
r
det
ect
ion
,
(c)
ra
w dem
on
strat
es t
he fit
ti
ng
proce
ss
with
regressio
n m
od
el
,
(
d)
ra
w descri
bed th
e local
outl
ie
r dete
ct
ion,
(e
)
r
aw rep
rese
nts the
e
xtracted
anom
aly
reg
i
on
s
or
cl
ust
ers,
last
r
a
w st
at
es the
den
sit
y of eac
h cl
us
t
er
DS
4.5.
Per
fo
r
m
an
ce
m
etrics
In
our
e
xperim
ents,
the
qua
ntit
at
ive
resu
lt
s
-
base
d
pe
rfo
rm
ance
e
valuati
on
ha
s
be
en
c
on
du
ct
e
d
us
in
g
a
receiver
oper
at
ing
cha
racter
ist
ic
(ROC)
cu
rv
e
a
nd
co
rr
es
pondin
g
area
unde
r
the
c
urve
(AUC)
m
et
ric
s.
I
n
this
w
ork,
the
ano
m
al
y
even
t
is
detect
e
d
ba
sed
on
a
cl
ip
(
even
t
)
le
vel
a
nd
f
ram
e
le
vel.
Co
ns
e
quently
,
we
record
the
be
gi
n
an
d
e
nd
of
each
a
no
m
al
y
even
t
with
ass
ist
ance
of
gro
und
t
ru
t
h
re
su
l
ts
pro
vid
e
d
for
each
te
sti
ng
vid
e
o
i
n
orde
r
to
co
nst
ru
ct
a
su
it
abl
e
groun
d
trut
h
-
cl
ip
le
vel
res
ults
for
validat
ion
an
d
c
om
par
iso
n
pur
po
ses
. To
obta
in ROC cur
ve
dem
on
strat
i
on
for
the test
ed
vid
e
o
sam
ples, tru
e posit
ive
rate (TPR)
a
nd f
al
se
po
sit
ive
rate
(
FPR)
m
et
rics
are
cal
culat
ed
at
cl
i
p
le
vel.
AU
C
area
m
eas
ur
es
a
re
com
pu
te
d
f
or
eac
h
vid
e
o
in
bo
t
h
dataset
s,
t
hen
we
a
ver
a
ge
the
obta
ined
resu
lt
s
to
s
how
the
AU
C
for
each
dataset
c
om
par
ed
wit
h
recent
m
et
ho
ds
t
hat
ba
sed
on
neural
netw
orks
s
uch
as
[25]
an
d
[
2
7
]
.
Table
1
e
xh
i
bits
the
co
rr
ect
detect
ion
sc
ore
and
false
al
arm
ben
chm
ark
f
or
A
venu,
Ped1
a
nd
Pe
d2
dataset
s
res
pecti
vely
.
Fig
ure
9
il
lustr
at
es
the
ROC
curves
qu
a
ntit
at
ive r
es
ults f
or Ave
nu,
Ped2 a
nd Pe
d1
dataset
s r
es
pe
ct
ively
.
Table
1.
A
c
om
par
ison
resu
l
ts of eve
nt le
ve
l detec
ti
on in
t
erm
o
f
co
rr
ect
detect
ion an
d f
al
se ala
rm
r
at
e
Metho
d
Co
rr
ect detectio
n
/
f
alse alar
m
Av
en
u
Ped
1
Ped
2
An
o
m
a
ly
even
ts
47
40
12
M.
H
asan
et al
.
[
2
5
]
4
5
/4
3
8
/6
1
2
/1
S.
Yan
et
al
.
[
2
7
]
3
4
/6
3
8
/5
1
2
/0
Y.
S.
Ch
o
n
g
an
d
Tay
[
2
8
]
4
3
/8
3
6
/1
1
1
2
/3
The p
rop
o
sed
work
4
5
/3
3
8
/4
1
2
/1
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
An
omaly eve
nt d
et
ect
io
n and l
oca
li
za
ti
on
of
vi
deo
cli
ps
us
i
ng
… (
Sawsen
Ab
du
l
hadi M
ahm
ood
)
1071
A
c
omm
on
ly
,
the
highe
r
va
lues
of
the
AU
C
area
val
ue
ref
e
rr
e
d
t
o
highe
r
perfor
m
ance
of
t
he
detect
or
or
cl
a
ssifie
r
m
et
ho
d.
As
s
how
n
i
n
F
igure
8,
t
he
pr
opos
e
d
m
et
ho
d
pr
e
sents
a
hi
gher
A
UC
value
w
he
n
app
li
ed
on
ave
nu
e
dataset
t
ha
n
Pe
d1
a
nd
P
ed2.
T
he
c
r
ow
ded
sce
nes
i
nc
lud
e
d
i
n
vid
e
o
sam
ples
are
e
ff
ect
s
m
ai
nly of
inc
re
asi
ng
false al
ar
m
r
at
es as sh
own
in Fi
g
ure
10. Fr
om
o
ther han
d, we
g
et
c
om
petit
ive r
esults a
nd
higher
pe
rfor
m
ance
f
or
a
venu
an
d
Pe
d2
data
set
s
w
hen
com
par
e
d
with
the
pro
posed
ap
proach
to
sta
te
of
the
art m
et
ho
ds
as
il
lustrate
d
in
T
able 2.
Figure
9. Roc
curves
vis
ualiz
at
ion
resu
lt
s
of the
pro
po
se
d work at
cli
p
le
vel for
ave
nue
,
Ped2 a
nd Pe
d1
dataset
s
Figure
10. T
w
o
e
xam
ples o
f fai
lure
c
ases
of the
pro
po
se
d
a
no
m
al
y detec
tor
Table
2.
A
c
om
par
ison
resu
l
ts of f
ram
e level
d
et
ect
ion i
n t
erm
o
f
A
UC
m
et
ric
Metho
d
AUC (
%)
Av
en
u
Ped
1
Ped
2
G.
Pan
g
et al
.
[
1
2
]
-
8
3
.2
7
1
.7
M.
H
asan
et al
.
[
2
5
]
7
5
.2
90
8
1
.5
Lu
et al.
[
2
6
]
6
5
.5
-
6
3
.8
S.
Yan
et
al.
[
2
7
]
7
9
.6
7
1
.9
7
5
.0
The p
rop
o
sed
wo
rk
83
76
82
Fr
om
the
e
xp
e
rim
ents,
we
f
ound
t
hat
the
cl
us
te
r
with
de
nsi
ty
cl
os
e
to
ze
ro
val
ue
is
c
on
sidere
d
as
a
singular
an
oma
ly
fr
am
e
(sudd
e
n
cha
nge),
w
hile
the
cl
us
te
rs
with
hig
he
r
de
ns
it
y
r
epr
ese
nt
a
n
a
no
m
al
y
colle
ct
ive
f
ram
e
in
te
rm
of
a
nom
aly
even
t
de
te
ct
ion
in
the
a
cqu
i
red
vi
deo
s
a
m
ple.
F
ur
the
r
,
the
area
pa
ra
m
et
er
us
e
d
f
or
fore
gro
und
obj
ect
s
extracti
on
is
m
or
e
accurate
and
ref
le
ct
s
a
n
ef
fici
ent
crit
erio
n
of
the
se
qu
e
nce
fr
am
es
beh
a
vi
our
c
om
par
ed
to
the
sim
ilarity
distance
m
easur
e
bet
ween
each
tw
o
se
qu
e
nce
d
fr
am
es.
Fu
rt
her
m
or
e,
t
he
pr
opos
e
d
a
no
m
al
y
even
t
detect
ion
sc
he
m
e
is
validat
ed
ba
sed
on
th
e
avail
able
da
ta
set
without
need
i
ng
to
perform
an
a
ug
m
entat
ion
of
the
dataset
s
in
co
ntra
st
to
dee
p
le
a
rn
i
ng
-
an
om
al
y
eve
nt
detect
ion
-
b
a
se
d
m
et
ho
ds w
hi
ch req
uire
d
a l
arg
e
d
at
aset
t
o l
earn
t
he
a
no
m
al
y beh
a
viours.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
2
4
, N
o.
2
,
N
ove
m
ber
20
21
:
106
3
-
107
3
1072
4
CONCL
US
I
O
N
We
ha
ve
sug
ge
ste
d
an
d
pres
ented
a
n
an
oma
ly
even
t
detect
ion
an
d
loca
li
zat
ion
schem
e
based
on
determ
ining
th
e
global
an
d
local
outl
ie
rs
in
vi
deo
se
que
nce.
T
he
prob
l
e
m
of
global
ou
tl
ie
r
detect
ion
was
assum
ed
as
re
gr
essi
on
m
od
e
l
fitt
ing
in
ord
er
to
segm
ent
the
vi
deo
f
ram
es
into
no
n
overlap
ping
cl
ips
.
T
he
pro
po
se
d
m
odel
is
able
to
up
date
the
trai
ne
d
re
gr
es
sio
n
m
od
el
us
ing
ne
wly
detect
ed
ano
m
al
y
clips
.
The
global
outl
ie
rs
are
ex
plo
it
ed
to
segm
ent
the
vid
e
o
f
ram
es
into
cl
ips.
Furt
her
m
or
e,
t
he
local
outl
ie
rs
at
eac
h
cl
ip
are
detect
ed
us
in
g
de
ns
i
ty
of
each
cl
ip
ta
kin
g
in
our
consi
der
at
io
n
the
co
ntextual
ou
tl
ie
rs
withi
n
each
vid
e
o
cl
ip.
Ba
sed
on
the
e
xperim
ents
that
i
m
ple
m
ented
on
public
an
om
aly
even
t
de
te
ct
ion
datas
et
,
we
exh
i
bited
that
the
pro
posed
a
no
m
al
y
detecto
r
si
gn
i
ficantl
y
has
a
com
petit
ive
per
f
or
m
ance
com
par
e
d
to
the
sta
te
-
of
-
the
-
art
m
e
tho
ds i
n t
er
m
s o
f
detect
io
n
acc
ur
acy
a
nd
f
al
se ala
rm
r
at
es at cl
ip level.
ACKN
OWLE
DGE
MENTS
This
re
searc
h
was
par
ti
al
ly
su
pp
or
te
d
by
com
pu
te
r
sci
e
nce
dep
a
rtm
ent
-
colle
ge
of
e
du
c
at
io
n
i
n
Mustansiriy
ah
Un
i
ver
sit
y.
REFERE
NCE
S
[1]
G.
Pang,
C.
Shen,
L
.
Cao,
and
A.
Van
Den
Henge
l
,
"
De
ep
le
arn
ing
for
ano
m
aly
de
tecti
on
:
a
rev
ie
w
,"
ACM
Computing
Surv
ey
s
,
vo
l. 54,
no.
2,
pp
.
1
-
38
,
Mar
.
2021
,
doi:
10.
11
45/3439950
.
[2]
K.
Yan,
X.
You,
X.
Ji,
G.
Yin
,
and
F.
Yang,
"A
hy
br
id
outlier
det
ection
m
et
h
od
for
hea
lt
h
c
a
re
big
data,
"
20
16
IEE
E
Inte
rnat
io
nal
Confe
r
enc
es
on
Bi
g
Data
and
Cloud
Computi
ng
(
BDCloud)
,
Soci
al
Comput
ing
and
Ne
tworki
ng
(
Soci
alCom)
,
Sustainabl
e
Computing
and
Comm
unic
ati
ons
(
SustainCom
)
(
BDCl
oud
-
Soci
alCom
-
SustainCom)
,
Oct.
2016,
pp
.
157
-
1
62,
doi
:
10
.
1109
/BDCloud
-
SocialCom
-
Sus
ta
inCom
.
2016.
34
.
[3]
Y.
W
ang,
Z
.
W
u
,
Y.
Zhu
,
and
P.
Zha
ng,
"Resea
r
c
h
on
anomal
y
d
e
te
c
ti
on
al
gori
th
m
base
d
on
general
i
za
t
ion
l
at
en
c
y
of
te
lecom
m
unic
ation
net
work,
"
Fut
ure
Gene
ration
Computer
Syste
ms
,
vol.
85,
pp.
9
-
18,
Aug.
2018
,
doi:
10.10
16
/
j
.
f
uture.
2018.
02.
022
.
[4]
W
.
L
i
,
V
.
M
a
h
a
d
e
v
a
n
,
a
n
d
N
.
V
a
s
c
o
n
c
e
l
o
s
,
"
A
n
o
m
a
l
y
D
e
t
e
c
t
i
o
n
a
n
d
L
o
c
a
l
i
z
a
t
i
o
n
i
n
C
r
o
w
d
e
d
S
c
e
n
e
s
,
"
i
n
I
E
E
E
T
r
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
c
h
i
n
e
I
n
t
e
l
l
i
g
e
n
c
e
,
v
o
l
.
3
6
,
n
o
.
1
,
p
p
.
1
8
-
3
2
,
J
a
n
.
2
0
1
4
,
d
o
i
:
1
0
.
1
1
0
9
/
T
P
A
M
I
.
2
0
1
3
.
1
1
1
.
[5]
Y.
Cong,
J.
Yu
an
,
and
J.
L
iu,
"S
par
se
rec
onstr
uct
ion
cost
for
abnor
m
al
ev
ent
det
e
ct
ion
,
"
CV
PR
2011
,
2011,
pp.
3449
-
3456
,
doi:
10
.
1109/CV
PR
.
2011.
599543
4.
[6]
T.
Xiao
,
C.
Zh
a
ng
,
and
H.
Zh
a,
"Le
arn
ing
to
de
te
c
t
anomalies
i
n
surveil
la
n
ce
v
i
deo,
"
in
IE
EE
S
ignal
Proce
ss
in
g
Lett
ers
,
vo
l. 22,
no.
9
,
pp
.
1477
-
1481,
Sept
.
2015
,
doi
:
10
.
1109/LS
P.2015.
2410031
.
[7]
L.
Krat
z
and
K.
Nishino,
"A
nom
aly
de
tecti
on
i
n
ext
remel
y
cro
wded
sce
nes
using
sp
atio
-
te
m
po
ral
m
oti
on
patte
rn
m
odel
s,"
2009
IEE
E
Confe
r
enc
e
on
Computer
Vi
sion
and
Pat
te
rn
Re
c
ognit
ion
,
2009
,
pp.
1446
-
1453,
doi:
10
.
1109/CV
PR
.
2009.
520677
1
.
[8]
V.
Sali
gr
ama,
J.
Konrad
,
and
P.
Jodoin,
"V
id
eo
Anom
aly
Ide
n
tification,
"
in
IEEE
Signa
l
Proc
essing
M
agazine
,
vol.
27
,
no
.
5
,
pp
.
18
-
33
,
Sept
.
20
10,
doi
:
10
.
1109
/MS
P.2010.
937393
.
[9]
K.
Min
and
J.
Corso,
"TAS
ED
-
Net:
Te
m
pora
l
l
y
-
agg
reg
a
ti
ng
spatial
rnc
oder
-
d
ec
oder
ne
twork
for
vide
o
salie
n
c
y
det
e
ct
ion
,
"
201
9
IEEE/CVF
I
nte
rnational
Co
nfe
renc
e
on
C
omputer
Vi
sion
(
ICCV)
,
2019,
pp.
2394
-
240
3,
doi:
10
.
1109/IC
CV.2019.
00248
.
[10]
L.
W
ang
et
al.
,
"
Te
m
pora
l
segm
e
nt
net
works
:
tow
ard
s
good
pra
ct
i
ce
s
for
d
ee
p
acti
on
rec
ogn
it
ion
,
"
2016
European
Confe
renc
e
on
C
omputer
Vi
sion
(
ECCV
)
,
vol
.
991
2
,
2016
,
pp
.
20
-
36
,
doi
:
10
.
1007
/978
-
3
-
319
-
4648
4
-
8_2
.
[11]
Y.
Zhu
and
S.
News
am,
"M
o
ti
on
-
awa
r
e
fe
at
u
re
for
improve
d
vide
o
anomal
y
de
tecti
on
,
"
arXiv
:1907
.
10211
,
pp.
1
-
12
,
2019
.
[12]
G.
Pang,
C.
Yan,
C.
Shen,
A.
Van
,
Den
Henge
l
,
and
X.
Bai
,
"S
e
lf
-
t
ra
ine
d
d
e
ep
o
rdina
l
r
egr
ession
for
e
nd
-
to
-
e
nd
v
ide
o
a
nom
aly
d
et
e
ct
ion
,
"
2020
IEE
E/
C
VF
Co
nfe
renc
e
on
Computer
Vi
sion
and
Pat
te
rn
Rec
ognit
ion
(
CVP
R
)
,
2020,
pp
.
12170
-
12179,
do
i
:
10
.
1109/CVPR
42600.
2020.
01219
.
[13]
G.
Pang,
C
.
She
n
and
A.
V
an
D
en
Henge
l,
"D
e
e
p
anomal
y
d
et
e
c
ti
on
with
deviation
net
works
,
"
2
020
Proceedi
ngs
of
the
25th
ACM
SIGKD
D
Inte
rnational
Con
fe
re
nce
on
Knowle
d
ge
Discov
ery
&
Data
Mini
ng
,
20
20
,
pp.
353
-
362
,
doi:
10
.
1145/32
92500.
3330871
.
[14]
F.
La
nd
i, C. G.
Snoek
,
and
R.
C
ucc
hi
ara,
"A
nom
aly
loc
a
li
t
y
in
vi
deo
survei
ll
an
ce,
"
arXiv
:1901
.
10
364
,
2019
.
[15]
W
.
Sulta
ni,
C.
Chen
,
and
M.
Shah,
"Rea
l
-
wo
rld
anomal
y
d
etec
t
ion
in
survei
ll
an
ce
vide
os
,
"
2018
IEE
E/
C
VF
Confe
renc
e
on
C
omputer
Vi
sion
and
Pattern
R
ecogniti
on,
2018
,
pp.
6479
-
6488
,
doi:
10
.
1109/CV
PR
.
2018.
00678
.
[
1
6
]
F
.
J
i
a
n
g
,
J
.
Y
u
a
n
,
F
.
J
i
a
n
g
,
J
.
Y
u
a
n
,
S
.
A
.
T
s
a
f
t
a
r
i
s
,
a
n
d
A
.
K
.
K
a
t
s
a
g
g
e
l
o
s
,
"
A
n
o
m
a
l
o
u
s
v
i
d
e
o
e
v
e
n
t
d
e
t
e
c
t
i
o
n
u
s
i
n
g
s
p
a
t
i
o
t
e
m
p
o
r
a
l
c
o
n
t
e
x
t
,
"
C
o
m
p
u
t
e
r
V
i
s
i
o
n
a
n
d
I
m
a
g
e
U
n
d
e
r
s
t
a
n
d
i
n
g
,
v
o
l
.
1
1
5
,
p
p
.
3
2
3
-
3
3
3
,
2
0
1
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
c
v
i
u
.
2
0
1
0
.
1
0
.
0
0
8
.
[17]
Ce
Li,
Z.
H
an,
Qixi
ang
Ye,
a
nd
J.
Jiao
,
"V
isual
abnor
m
al
beha
vior
d
et
e
ct
i
on
base
d
on
tr
aj
e
ct
or
y
sparse
rec
onstruc
ti
on
a
naly
s
is,"
Neuro
c
omputing
,
vo
l. 1
19,
pp
.
94
-
100
,
2013
,
doi
:
10
.
10
16/j
.
n
euc
om
.
201
2.
03.
040
.
[
1
8
]
M
.
G
.
N
a
r
a
s
i
m
h
a
n
a
n
d
S
.
K
a
m
a
t
h
,
"
D
yn
a
m
i
c
v
i
d
e
o
a
n
o
m
a
l
y
d
e
t
e
c
t
i
o
n
a
n
d
l
o
c
a
l
i
z
a
t
i
o
n
u
s
i
n
g
s
p
a
r
s
e
d
e
n
o
i
s
i
n
g
a
u
t
o
e
n
c
o
d
e
r
s
,
"
M
u
l
t
i
m
e
d
i
a
T
o
o
l
s
A
p
p
l
.
,
v
o
l
.
7
7
,
n
o
.
1
1
,
p
p
.
1
3
1
7
3
-
1
3
1
9
5
,
J
u
n
.
2
0
1
8
,
d
o
i
:
1
0
.
1
0
0
7
/
s
1
1
0
4
2
-
0
1
7
-
4
9
4
0
-
2.
Evaluation Warning : The document was created with Spire.PDF for Python.