Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
6, N
o
. 4
,
A
ugu
st
2016
, pp
. 15
95
~
1
601
I
S
SN
: 208
8-8
7
0
8
,
D
O
I
:
10.115
91
/ij
ece.v6
i
4.1
017
9
1
595
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Unusual Event Detection Using
Mean Feature Point Matching
Algorith
m
Chitr
a
Hegde,
Sh
akti Singh Chu
n
d
a
wat, Divya
S N
Department o
f
C
o
mputer Scien
c
e,
Amrita Vishwa Vidy
apeetham
U
n
i
v
e
r
s
i
ty
,
My
su
r
u
C
a
mp
u
s
Karnatak
a,
Indi
a
Article Info
A
B
STRAC
T
Article histo
r
y:
Received Feb 15, 2016
Rev
i
sed
May 18
, 20
16
Accepted
May 29, 2016
Analy
s
is
and detection of
unusu
a
l ev
ents in
pub
lic
and priv
ate surveillance
s
y
stem is a co
mplex task. Detecting
unusual
events in
surveillan
ce v
i
deo
requires the
appr
opriate definitio
n of si
m
ilarit
y
b
e
tween ev
ents
.
T
h
e ke
y
go
al
of the proposed s
y
stem is to
detect
b
e
havio
u
rs or actions
that
can b
e
cons
idered
as
a
nom
alies
.
S
i
nc
e
s
u
s
p
icious events differ from domain to
dom
ain, it r
e
m
a
i
n
s a chal
lenge
t
o
dete
ct those
e
v
ents in m
a
jor d
o
m
a
ins such
as airport, super malls, edu
c
ational
institutions etc. Th
e prop
osed Mean
Feature Point
Matching (MFPM) algorith
m is used for detecting unusual
events. Th
e Speeded-Up Robus
t Features (SURF) method is used
for featur
e
extra
c
tion
.
Th
e
M
F
PM
algorith
m
com
p
ares
the
featur
e poin
t
s
of the inpu
t
im
age with
the
m
ean fe
ature
points
of
tr
ained dataset. The
experimental
result shows th
at
the propos
ed
s
y
stem
is eff
i
cient and
accurate for wid
e
variety
of
surveillan
ce v
i
deos.
Keyword:
Ev
en
t d
e
tection
Spee
ded
-
up
r
o
bust
feat
u
r
es
Un
us
ual
e
v
ent
s
Vi
de
o m
i
ni
ng
Copyright ©
201
6 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
Chitra He
gde,
Depa
rt
m
e
nt
of
C
o
m
put
er Sci
e
nce,
Am
rita Vishwa
Vidy
a
p
eetham
U
n
ive
r
sity
, M
y
sur
u
Cam
pus,
11
4,
7
th
C
r
oss
,
B
oga
di
2
nd
Stag
e, Mysu
ru
-570
026
.
Em
a
il: ch
it.h
e
gd
e@g
m
ail.co
m
1.
INTRODUCTION
M
oni
t
o
ri
n
g
pu
bl
i
c
or
p
r
i
v
at
e
si
t
e
s has
bec
o
m
e
a bi
g i
ssue i
n
cl
udi
ng t
h
e m
oni
t
o
ri
n
g
of
rai
l
w
ay
stations, airports, hyper m
a
rkets,
education institutes and
inaccessible
or dangerous envi
ronm
ents. Recent
tech
no
log
i
cal
d
e
v
e
l
o
p
m
en
ts
h
a
v
e
b
e
en
adop
ted
b
y
cu
rren
t
surv
eillan
ce
syste
m
s to
p
r
odu
ce
fu
lly d
i
g
ital v
i
d
e
o
record
i
n
g
t
o
keep
track
o
f
su
sp
ici
o
u
s
ev
ents. Su
rv
eillan
c
e ca
m
e
ras are a g
r
eat way to p
r
o
v
i
d
e
security fo
r
hom
e or workplace. They
provi
de us
w
ith
vide
o foota
g
e
of a
n
y events
wh
ic
h m
a
y happe
n, a
nd als
o
act as
a
vi
si
bl
e det
e
r
r
e
n
t
t
o
cri
m
i
n
al
s. B
u
t
pr
ocessi
ng
of t
h
ese
vi
deo
s
rem
a
i
n
s chal
l
e
ngi
ng
fo
r
e
ver
.
Si
nce su
spi
c
i
o
us
event
s
di
ffe
r
f
r
om
dom
ai
n t
o
dom
ai
n, i
t
r
e
m
a
i
n
s chal
l
e
ngi
ng
t
o
det
e
c
t
t
hose
e
v
ent
s
i
n
eac
h
d
o
m
a
i
n
l
i
k
e
airpo
r
t,
sup
e
r
malls, ed
u
cation
institu
tes etc.
Un
us
ual
eve
n
t
can be
de
fi
ne
d
as an e
v
ent
w
h
i
c
h
devi
at
es f
r
om
norm
a
l
behavi
ou
r. T
h
ey
occu
r
very
rarely in e
n
tire vide
o se
que
n
ce.
These
events are
unpredictable as
well. Unusual
eve
n
t m
a
y indicate
im
port
a
nt
o
b
je
ct
s an
d e
v
ent
s
i
n
wi
de
va
ri
et
y
of
d
o
m
a
i
n
s.
On
e of t
h
e m
a
j
o
r d
i
fficu
lties i
n
su
rv
eillan
ce
v
i
d
e
o
an
alysis is th
e
h
u
g
e
am
o
u
n
t
o
f
d
a
ta, wh
ere on
ly a
sm
al
l
port
i
o
n
of
vi
deo
c
ont
a
i
ns i
m
port
a
nt
i
n
f
o
rm
at
i
on. T
h
e e
v
e
n
t
det
e
c
t
i
on i
n
vi
de
o
[
1
]
-
[
7
]
i
s
a
n
i
m
po
rt
an
t
task
w
h
en
w
e
r
eally fo
cu
s
o
n
secu
r
ity issu
es
of
an
o
r
g
a
n
i
zation
wher
e ev
er
y si
n
g
le actio
n
should
be
considere
d
for the
de
tection proces
s.
In
curren
t
surveillan
ce syste
m
,
it
n
eed
s human
o
b
s
er
v
e
r to
assess th
e v
i
d
e
o
th
at is b
e
in
g
g
e
n
e
rated.
M
oni
t
o
ri
n
g
al
l
t
h
e gene
rat
e
d
vi
de
os an
d fi
n
d
i
n
g t
h
e sus
p
i
c
i
ous e
v
ent
i
s
t
e
di
ou
s j
o
b
.
Fo
r exam
pl
e, assum
i
ng
th
at a rare acti
o
n
is
related
hu
m
a
n
activ
ity,
wh
ere a pers
on is using cell phone in t
h
e pla
ce whe
r
e its us
age is
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
15
95
–
1
601
1
596
pr
o
h
i
b
i
t
e
d. I
n
orde
r t
o
o
v
er
com
e
t
h
e dra
w
bac
k
s o
f
t
r
a
d
i
t
i
onal
sy
st
em
, a new t
echni
q
u
e i
s
pr
op
o
s
ed. T
h
i
s
m
a
kes t
h
e
way
t
o
red
u
ce t
h
e
use
of
m
a
n po
wer
by
el
i
m
i
n
at
i
ng t
h
e
ne
ed
o
f
hum
an o
b
se
r
v
er
.
Unu
s
u
a
l ev
en
t
reco
gn
itio
n [8]-[11
]
is a ch
al
len
g
i
ng task du
e t
o
m
a
n
y
reason
s su
ch
as
co
nfu
s
ion
i
n
act
i
ons,
am
bi
gui
t
y
i
n
defi
ni
n
g
t
h
e
no
rm
al
ity
et
c. Th
er
e
are m
a
ny existing algorithm
s
to
det
ect
t
h
e
u
nus
ua
l
events
.
Yue
Z
h
o
u
,
S
h
ui
che
n
g
Ya
n
,
Th
om
as S.Hua
n
g
p
r
ese
n
t
e
d
a
su
per
v
i
s
e
d
al
go
ri
t
h
m
[1
2]
f
o
r
com
put
i
n
g
t
h
e sim
i
l
a
rit
y
o
f
m
o
t
i
on t
r
ajec
t
o
ri
es usi
n
g pa
ram
e
t
e
rs of edi
t
di
st
ance. The
anom
al
y det
ect
i
on was co
nsi
d
er
e
d
as g
e
n
e
ral
ou
tlier d
e
tection
p
r
o
b
l
em
. Th
ere
was scop
e to ex
tend
th
e algo
rith
m
fo
r b
e
tter
u
tilizatio
n
of featu
r
e
space. Deepa
k
et al. develope
d m
odel
for de
tection of
dom
inant be
havi
our in
vide
os [13] using uns
upe
rvis
e
d
p
r
ob
ab
ilistic to
p
i
c
m
o
d
e
ls. They calcu
lated
n
o
r
m
a
lized
lik
el
ih
oo
d
m
easu
r
es. Bu
t it is
n
o
t
efficien
t wh
en
th
ere
is no
o
r
less ano
m
al
y in
th
e
data set.
Au
t
h
or Pen
g
et
al. p
r
o
p
o
s
ed
a p
a
ttern
m
i
n
i
n
g
app
r
o
a
ch
[1
4] wh
ich
u
tilize th
e p
a
tterns to
ad
dress th
e
k
e
y pr
ob
lem
s
in
v
i
d
e
o
m
i
n
i
ng
an
d under
s
tand
ing
f
i
eld
.
Th
e
H
i
dden
Marko
v
Mo
d
e
l
(
H
MM
)
[15
]
represen
tatio
n
o
f
obj
ect traj
ecto
ries en
ab
les t
h
e sim
ilari
ty
m
easure
s
bet
w
ee
n vi
de
o
e
v
e
n
t
s
by
cr
oss
l
i
k
el
i
h
oo
d,
but
s
u
f
f
ere
d
fr
om
t
h
e over
fi
t
t
i
ng p
r
o
b
l
e
m
due t
o
dat
a
sh
ort
a
ge. Fa
n Ji
a
ng a
n
d t
eam
[16]
pr
op
ose
d
a
DHC
(Dynam
i
c
Hierarchical Clustering) approac
h
, whe
r
e th
e H
M
M
s
are t
r
ai
ned o
n
m
a
ny
diffe
rent
sam
p
l
e
s an
d
th
e in
itial clu
s
tering
erro
rs cau
s
ed
b
y
ov
er
fi
ttin
g
are corrected
in
t
h
e iterat
i
v
e
pro
cess.
C
o
n
f
i
d
e
n
t
-
F
r
a
m
e-based R
e
c
o
g
n
i
z
i
n
g al
g
o
r
i
t
h
m
(C
FR
) [1
7]
was p
r
op
os
ed t
o
rec
o
gni
z
e
t
h
e hum
an
activ
ity, wh
ere h
i
gh
confid
ence v
i
d
e
o
fr
am
e
s
are used a
s
a
specialized mode
l in th
e classificatio
n
o
f
t
h
e rest
o
f
t
h
e v
i
d
e
o
fram
es. Fo
r act
iv
ities su
ch
as Fig
h
ting
an
d Ru
nn
ing
wh
ere th
e
GMM classifiers h
a
ve low
d
e
tectio
n rates.
Th
e rev
i
ewed
stu
d
i
es are an
alysed
b
a
sed
on fiv
e
asp
e
cts— su
rv
eillan
ce targ
et, an
o
m
aly
d
e
fin
ition
s
and as
sum
p
tions, the
feature
extraction processes,
learning m
e
thodol
ogi
es, an
d m
odelling algorithm
s
. The
sus
p
i
c
i
o
u
s
eve
n
t
s
di
f
f
er
f
r
o
m
dom
ai
n t
o
dom
ai
n an
d
remain
s ch
alleng
ing
to
d
e
tect th
o
s
e ev
en
ts
in
each
dom
ai
n l
i
k
e ai
rp
ort
,
s
u
per m
a
l
l
s
, educat
i
o
n
a
l
depa
rt
m
e
nt
s so o
n
.
The
r
ef
ore t
h
e p
r
o
p
o
s
e
d sy
st
em
overcom
es
t
h
e dra
w
bac
k
of exi
s
t
i
n
g sy
st
em
by im
pl
em
ent
i
ng M
ean
Feat
ure Poi
n
t
M
a
t
c
hi
ng (M
FPM
) al
g
o
ri
t
h
m
.
The
SUR
F
desc
ri
pt
or
i
s
us
ed
f
o
r
t
h
e
feat
ure
e
x
t
r
act
i
on
[1
8]
.
2.
R
E
SEARC
H M
ETHOD
2.
1.
Sys
t
em Over
view
Th
e system
tak
e
s v
i
d
e
o
s
fro
m
su
rv
eillan
ce ca
m
e
ra as in
pu
t
,
d
e
tects th
e
unu
sual ev
en
ts usin
g
trai
n
e
d
dat
a
set
.
T
h
e S
U
R
F
t
e
c
hni
que
i
s
ap
pl
i
e
d
o
n
eno
r
m
ous set
of
sam
p
l
e
im
ages
whi
c
h a
r
e
con
s
i
d
ere
d
a
s
un
us
ual
ev
en
ts. Co
nsider
cell p
h
o
n
e
usag
e as an
unusu
a
l ev
en
t in
pr
oh
ib
ited
ar
ea, sam
p
le
i
m
ag
es
o
f
p
e
o
p
l
e using
cell
phones i
n
different situat
i
o
ns are ta
ken to train t
h
e syst
e
m
. Figu
re
1
shows
fe
w sa
m
p
le images that are
considere
d
for
training the sy
ste
m
. Th
e syste
m
detect feature
points in e
ach sam
p
le image and e
x
tract
feature
descri
pt
o
r
s at
t
h
e i
n
t
e
rest
p
o
i
n
t
s
as s
h
ow
n i
n
t
h
e Fi
gu
re
2.
Fi
gu
re
1.
Sam
p
l
e
Im
ages
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Unu
s
u
a
l
Even
t
Detectio
n Usin
g Mea
n
Fea
t
ure Po
i
n
t Ma
tchin
g
Algo
rith
m
(Ch
itra
Hegd
e)
1
597
Fig
u
re
2
.
Visualizatio
n
of stro
ng
est
SURF po
in
ts
The system is
designe
d in s
u
ch a way tha
t
it c
onside
r
s 150 strong fea
t
ur
e points in
each sam
p
le
im
age, the m
e
an feature
poi
n
t is
com
pute
d
a
n
d
sto
r
ed
fo
r f
u
rthe
r p
r
ocessing. T
h
e
Mean Feat
ure Poi
n
t
Match
i
n
g
(MFPM) algorith
m is app
lie
d on
in
pu
t fram
es fo
r d
e
tecting
t
h
e unu
su
al
ev
en
ts
with
t
h
e
help
o
f
train
e
d sam
p
les.
2.
2.
Feat
ure E
x
tr
a
c
tion
To det
ect
bl
o
b
feat
ures
, Spee
ded
-
Up R
o
b
u
s
t
Feat
ures (S
U
R
F) m
e
t
hod i
s
im
pl
em
ent
e
d. The SUR
F
m
e
t
hod i
s
a sc
al
e and
rot
a
t
i
o
n i
n
vari
a
n
t
i
n
t
e
rest
p
o
i
n
t
det
ect
or a
nd
desc
ri
pt
o
r
. S
U
R
F
uses a
Hessi
a
n
m
a
t
r
i
x
wh
ich
is a secon
d
d
e
riv
a
tiv
e
matrix
for
feat
u
r
e ex
tracti
o
n.
For
feat
u
r
e
de
scri
pt
i
o
n, T
h
e
SUR
F
al
go
ri
t
h
m
uses
wav
e
let resp
on
ses i
n
horizon
tal an
d
v
e
rtical
di
rect
i
o
ns. A
n
e
i
g
h
b
o
u
r
h
o
od
of si
ze M
x
N i
s
t
a
ken aro
u
nd t
h
e key
poi
nt
. F
u
rt
her i
t
i
s
di
vi
ded i
n
t
o
su
b re
gi
o
n
s.
Fo
r each
su
b r
e
g
i
on
,
ho
r
i
zo
n
t
al an
d v
e
r
tical w
a
v
e
let
r
e
sponses ar
e tak
e
n
an
d a
v
ector
V
i
s
fo
r
m
ed
as
V= (
∑
d
x
,
∑
d
y
,
∑
|d
x
|,
∑
|d
y
|)
(
1
)
The sum
s
of
d
x
and |d
x
| are c
o
m
puted separat
e
ly for
d
y
< 0 a
n
d
d
y
≥
0 an
d t
h
e sum
s
of
d
y
and
|d
y
| are
sp
lit up
acco
r
din
g
t
o
th
e sign
o
f
d
x
, th
en
it do
ub
les t
h
e
n
u
m
b
e
r of
featu
r
e po
in
ts.
Th
e featu
r
e poin
t
s ex
tracted
fro
m
sa
m
p
le i
m
ag
es will b
e
in
th
e form
o
f
Mx
N m
a
trix
.
Th
is m
a
trix
will b
e
con
v
e
rted
in
to
sing
le d
i
m
e
n
s
io
n
array fo
r
furth
e
r
p
r
o
cessi
n
g
b
y
co
m
p
u
tin
g
th
e
m
ean
o
f
th
e
featu
r
e
poi
nts. T
h
e m
ean feat
ure
point is th
e ave
r
age
value i
n
eac
h
colum
n
of th
e MxN feature matrix.
The
com
puted
resu
lts will
b
e
co
nsid
ered
fo
r th
e
d
e
tection
of u
n
u
s
u
a
l
ev
en
t
s
.
2.
3.
Mea
n
Fe
at
ure
Poin
t
Ma
tchi
ng
(M
FPM
)
A
l
gori
t
hm
Th
e MFPM al
g
o
rith
m
tak
e
s i
n
pu
t in
th
e
form o
f
v
i
d
e
o
.
It
co
nv
erts th
e v
i
d
e
o
i
n
to
set of
i
m
ag
e fram
e
s
and
rem
oves t
h
e d
u
p
l
i
cat
e fram
e
s. The i
nput
i
m
age fram
e
s conve
rt
ed
i
n
t
o
g
r
ay
scal
e im
ages for
fu
rt
h
e
r
pr
ocessi
ng
.
Th
e
d
e
sired
o
u
tp
u
t
is to
ob
tain
a m
a
tch
o
f
fra
m
e fro
m
th
e train
e
d
d
a
taset.
Th
e
k
e
y fu
n
c
ti
o
n
a
lity lies
in observing
mean poi
nts and
patterns in
each fram
e
an
d studying t
h
e
resem
b
lance with the traine
d data.
Training of the
dataset a
r
e stored fo
r distinct
fram
e
, each one
represe
n
ting
vari
ous
test
cases. T
h
e al
gorithm
run
s
in
a
no
n-determin
istic p
o
lyn
o
m
ia
l ti
me
as th
ere is
n
o
gu
aran
tee th
at fra
m
e
m
a
tch
e
s
with
th
e trai
n
e
d
sets,
m
o
re possi
bl
e cases o
f
t
r
ai
ne
d dat
a
i
s
t
a
ken
i
n
o
r
de
r t
o
st
r
e
ngt
hen
t
h
e al
go
ri
t
h
m
.
The f
o
l
l
o
wi
ng
pa
ra
m
e
t
e
rs
are
used in t
h
e
algorithm
.
The i
n
p
u
t
vi
de
o
∑
is co
nv
erted
in
to
N
f
im
age fram
e
s, where F
0
b
e
th
e in
itial fra
m
e
. Each
F
i
where i
< N
f
i
s
c
o
nve
r
t
ed i
n
t
o
qua
d
r
a
n
t
s
Q
i
(Q
1
, Q
2
, Q
3
, Q
4
).
Eac
h
q
u
ad
ra
nt
i
s
c
o
nsi
d
e
r
ed
f
o
r t
h
e m
a
t
c
hi
ng p
r
o
cess.
Train
e
d
d
a
taset will h
a
v
e
N
t
num
ber
of
o
b
j
ect
s, w
h
ere
T
0
is th
e in
itial obj
ect. Feat
u
r
e po
in
ts
O
i
of eac
h Q
i
fo
r
all th
e F
i
is com
p
ared
with th
e
m
ean feature poi
n
t Obj
i
f
r
om
t
h
e t
r
ai
ne
d
dat
a
set
.
I
f
m
a
t
c
h f
o
un
d,
va
l
u
e
o
f
corres
ponding M[Obj
i
,Q
i
] will b
eco
m
e
1
an
d
coun
t C will
in
creased
b
y
1. Ratio
is calcu
l
a
ted
b
y
d
i
v
i
d
i
ng
t
h
e
cou
n
t
C
by
t
o
t
a
l
num
ber
of
fr
am
es N
f
. Th
e lo
op
co
n
t
i
n
u
e
s
till th
e en
d
o
f
t
h
e seq
u
e
n
ce. Th
e algo
rith
m
retu
rns
th
e m
a
trix
as a resu
lt. Th
e
Feat
u
r
e Matchin
g
(FM) m
a
tr
ix
M[Obj
i
,Q
i
] represen
t th
e
map
p
i
ng
b
e
tween
th
e
objects i
n
the
traine
d
dataset
and
each qua
d
rant of t
h
e input fram
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
15
95
–
1
601
1
598
Tab
l
e
1
.
Param
e
ters used in
M
FPM Al
g
o
rithm
Para
m
e
ters
Mean
in
g
∑
I
nput
video
F
0
Initial fra
m
e
in the
sequence
N
f
Total nu
m
b
er of fra
m
e
s
T
0
Initial fra
m
e
in trai
ned data set
N
t
T
o
tal nu
m
b
er
of trained data set
Q
i
Quadr
a
nts of input
fr
am
e
Obj
i
M
ean featur
e points of object in tr
ained data
O
i
Featur
e points of Q
i
R Ratio
of
m
a
tching
C
Total nu
m
b
er of
matches
M[
Ob
j
i
,Q
i
] Featur
e
M
a
tching
(FM
)
m
a
tr
ix
0f
0
t
i
i
i
i
0
0i
f
i
M
F
P
M
,
F
,
N
,
T,
N
,
Q,
O,
M
O
b
j
,
Q
,
R
,
C
ste
p
1
:
I
nput th
e
se
qu
en
c
e
s
t
ep
2
:
P
r
e
p
ro
ces
s
s
t
e
p
3
:
Cons
id
e
r
the
i
n
it
ia
l
f
r
a
m
e
F
st
ep
4
:
M
o
v
e
t
h
ro
ug
h
fram
e
s F
t
o
F
w
h
e
re i
N
st
e
p
5
:
Ea
ch F
i
s
se
gm
e
i
ii
i
0i
t
ii
ii
nte
d
into qu
ad
r
a
nts Q
s
t
ep
6
:
C
o
m
p
are O
o
f
F
w
i
t
h
O
b
j
i
n
the
tr
a
i
ne
d da
tase
t
T
to T
w
h
e
r
e
i
N
ste
p
7
:
Ea
c
h
e
n
tr
y
c
o
r
r
e
spond
ing
to
Q
of
the
F
a
n
d
Obj
of
the
T
is pl
otte
d
ii
ii
ii
f
ii
ii
1
,
if
O
is
m
a
tc
he
d w
ith O
b
j
0
,
other
w
ise
in M
O
bj
,
Q
st
e
p
8
:
M
O
bj
,
Q
=
ste
p
9
:
I
f
M
O
bj
,
Q
v
a
l
u
e
1
,
the
n
C
C
1
ste
p10
:
Ratio R
C
/
N
ste
p11
:
Re
tu
r
n
m
a
tr
ix M
O
bj
,
Q
Fi
gu
re
3.
M
F
P
M
Al
g
o
ri
t
h
m
3.
R
E
SU
LTS AN
D ANA
LY
SIS
The
pr
o
pose
d
sy
st
em
consi
d
er en
o
r
m
ous i
m
age sam
p
les
to extract t
h
e
featur
es
of
in
te
r
e
s
t
e
d
ar
e
a
.
The t
e
st
i
n
g i
s
per
f
o
r
m
e
d by
gi
vi
n
g
t
h
e ca
p
t
ure
d
vi
de
os a
s
i
n
put
.
The
s
y
st
em
i
s
abl
e
t
o
det
ect
cel
l
ph
o
n
e
usa
g
e. Figure 4 (a) s
h
ows the
sa
m
p
le im
age
whe
r
e the cell
phone is detected in the fi
rst quadra
nt. Figure
5 (a
)
shows t
h
e FM
matrix whic
h i
ndicates that t
h
e sixth
object
i
n
th
e trai
n
e
d
dataset is
m
a
tch
i
n
g
with
th
e
obj
ect in
th
e first
qu
adran
t
of th
e curren
t inp
u
t
im
ag
e. Figu
re
4
(d
) sho
w
s th
e sam
p
le i
m
age where t
h
e cell phone is
det
ect
ed i
n
t
h
e
seco
nd
qua
d
r
a
n
t
,
he
nce
ot
her
qua
dra
n
t
s
are
not c
h
ecke
d
further. On
ce the
objects are m
a
tched,
it will
m
o
v
e
to
th
e n
e
x
t
fram
e
in
th
e
sequ
en
ce. Th
u
s
it will red
u
c
e th
e tim
e co
m
p
lex
ity.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Unu
s
u
a
l
Even
t
Detectio
n Usin
g Mea
n
Fea
t
ure Po
i
n
t Ma
tchin
g
Algo
rith
m
(Ch
itra
Hegd
e)
1
599
(a
)
(b
)
(c)
(d
)
Fi
gu
re
4.
U
n
us
ual
eve
n
t
det
e
c
t
ed i
n
di
f
f
ere
n
t
im
ages
Th
e co
un
t v
a
ri
ab
le C is
m
a
in
tain
ed
thro
ugh
ou
t th
e pro
g
ram. At th
e en
d, C
v
a
lu
e ind
i
cate th
e nu
m
b
er
of
f
r
am
es wi
t
h
un
us
ual
e
v
ent
i
n
t
h
e
i
n
put
vi
d
e
o.
(a)
(b)
(c)
(d)
Fi
gu
re
5.
Feat
u
r
e M
a
t
c
hi
n
g
(F
M
)
m
a
t
r
i
x
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
15
95
–
1
601
1
600
To eval
uate the perform
a
nce of th
e MFPM
algorit
hm
, it is com
p
ared
with
th
e trajecto
r
y si
milarity
analysis algorithm
.
Figure
6 show
s t
h
e pe
rform
a
nce comprasi
on
of MFPM
algorithm
using
precision recall
curve.
(
2
)
(
3
)
Figure
6. Preci
sion-Recall curve
Tab
l
e 2 shows th
e test resu
lt o
f
t
h
e system
with
t
w
o
di
ff
e
r
ent
vi
de
o i
n
p
u
t
s
.
Vi
de
os a
r
e
capt
u
red
i
n
diffe
re
nt scena
r
ios. T
h
ese
videos are c
o
nve
rted into
im
age fram
e
s with fr
am
e rate 23 f
r
a
m
e
s per sec
o
nd
. To
reduce t
h
e
processing tim
e
, duplicate
fram
e
s
are
rem
oved. The
system
detected 17 im
a
g
e
fram
e
s as fram
e
s
with
un
usu
a
l ev
en
ts in
th
e fi
rst v
i
d
e
o
inpu
t. Th
e syste
m
ef
ficien
tly d
e
tectin
g
th
e cell p
h
o
n
e
u
s
ag
e in
all th
e
scenari
o
s. It ca
n
detect objects w
i
t
h
u
n
u
sual
beha
vi
o
u
r
desp
i
t
e
a scal
e cha
n
ge
or
i
n
-
p
l
a
ne
r
o
t
a
t
i
on.
Tabl
e 2.
T
e
st
R
e
sul
t
Sa
m
p
le
Videos
Total nu
m
b
er of
fr
am
e
s
N
u
m
b
er
of fr
am
e
s
Considered
Unusual events detected
(C)
Video1
3895
779
17
Video2
4352
870
6
4.
CO
NCL
USI
O
N
Th
e
p
r
o
p
o
s
ed
syste
m
i
m
p
l
e
m
en
ts th
e robust tech
n
i
qu
e for d
e
tecting
unu
sual ev
en
ts in su
rv
eillan
c
e
vide
os. T
h
e m
ean feature poi
n
ts m
a
tching
m
e
thod
will k
eep track
of t
h
e
unus
ual eve
n
ts and
helps i
n
taking
effectiv
e action
s
. Th
e ex
p
e
rimen
t
ed
resu
lts
sh
ows th
e
effi
ciency
in detection of
ce
ll phone usa
g
e
in differe
n
t
q
u
a
dran
ts
of the in
pu
t im
ag
e
fram
e. The
obje
cts in the
trained
datase
t are
match
e
d
with
i
n
pu
t im
ag
e d
e
sp
ite a
scale ch
an
g
e
or in-p
lan
e
ro
tatio
n
s
. In
fu
t
u
re, add
ition
a
l
test cases can
b
e
includ
ed in th
e trai
n
e
d
d
a
taset to
achieve
accura
te results. T
h
e
degree
of ra
rity in e
v
ents
can be
classified
using dy
nam
i
c c
l
assifiers to enhanc
e
the pe
rform
a
nce of the
algorit
h
m
.
REFERE
NC
ES
[1]
Bazm
i A. and
F
aez K.
, “
I
ncreasing th
e Ac
curac
y
of D
e
te
ction
and Reco
gnition in Visu
al S
u
rveil
l
an
ce
,
”
International Jo
urnal of
Electr
ical and Computer Engin
eering
, vo
l/issue: 2
(
3), pp.
395, 2012
.
[2]
Lee K
.
M
.
and
Kwon J
., “
A
unified fr
am
ework for event
s
u
m
m
ariz
ation
and r
a
r
e
even
t de
te
ctio
n,” In
2012
IEEE
Conference on
C
o
mputer Vision
and Pat
t
ern R
e
cognition
, pp
. 12
66-1273, 2012
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Unu
s
u
a
l
Even
t
Detectio
n Usin
g Mea
n
Fea
t
ure Po
i
n
t Ma
tchin
g
Algo
rith
m
(Ch
itra
Hegd
e)
1
601
[3]
M
o
rris
B. T.
an
d Trived
i M
.
M
., “
A
s
u
rve
y
of
vis
i
on-bas
e
d tr
aje
c
tor
y
l
earn
i
n
g
and ana
l
y
s
is
for s
u
rveill
ance
,
”
Circuits and
Sys
t
ems for Video
Technolog
y, IEEE Transactions on
., vol/issue: 18(8
)
, pp
. 1114-27
, 2
008.
[4]
Huang Z.
F
.
and
Mori G.,
“
S
F
U
at TR
ECVi
d 201
0: Surveillan
c
e
Event Detection,” In
T
R
ECVID
, 2010.
[5]
Xia Z.,
et a
l
., “PKU-NEC@ TRECVID 2012 SE
D: Uneven-sequ
e
nce b
a
sed even
t detection
in surveillance video
,
”
In
Proc.
T
REC
V
I
D
, 2012.
[6]
Chen M
.,
et a
l
.
,
“
S
em
antic even
t
extra
c
tion
us
ing
neural
network
ens
e
m
b
les
,
” In
S
e
mantic Computing, 2007
, ICSC
2007, Internatio
nal Conference
on,
pp
. 575-580
, 2007.
[7]
Noam
an R. A.,
et al.
, “
H
um
an Detec
tion Fram
e
w
ork for Automated Surve
ill
anc
e
S
y
stem
s,”
In
ternational Journ
a
l
of Electrical and
Comput
er
Eng
i
neer
ing (
I
JEC
E
)
,
vol/issue: 6(2),
2015.
[8]
Zhong H.,
et al.
, “Detecting unusual activity
in video,” In
Computer Vision and Pattern
Recognition, 2004. CVPR
2004. Proceed
in
gs of th
e 2004
I
EEE
Computer
Society Con
f
erence on
, vo
l. 2, p
p
. II-819
, 2004
.
[9]
Lu C.
,
et a
l
.
, “
A
bnormal event detection
at 1
50 fps in matlab,” In
Proceedings of th
e IEEE Internationa
l
Conference on
C
o
mputer Vision
, pp.
2720-2727
, 2013.
[10]
Sodemann A. A.,
et a
l
.
, “
A
review of ano
m
al
y
de
te
ction
in autom
a
ted
surveill
ance
,”
Sy
st
e
m
s,
M
a
n
,
an
d
Cybernetics, Par
t
C:
Application
s
and Reviews, I
EEE
Transactio
ns on.
, vol/issue: 42(6), pp. 1257-
72, 2012
.
[11]
Nievas
E. B.
,
et al.
, “Violence detection in vid
e
o usi
ng compu
t
er vision techniques,” In
Computer Analysis o
f
Images and Pa
tterns
, pp. 332-33
9, 2011
.
[12]
Zhou Y.,
et al
.,
“
D
etecting ano
m
al
y
in vid
e
os
from
trajector
y s
i
m
ilarit
y
an
al
ys
is
,” In
Multimed
ia
and Expo, 2007
IEEE Internatio
nal
Conf
er
ence
on
, pp
. 1087-10
90, 2007
.
[13]
A. Mukerjee,
et al.
, “Unsupervised Modelling o
f
‘Usual’ Even
ts a
nd Dete
cting
An
om
alies in
Vide
os”.
[14]
Cui P
.,
et al.
,
“
H
ierarchi
cal
vi
s
u
al ev
ent p
a
tt
ern mining and
its applications,”
Data Min
i
ng
and Knowledg
e
Dis
c
over
y,
vol/is
sue: 22(3), pp
. 4
67-92, 2011
.
[15]
Zhang D.,
et a
l
.
, “Semi-supervised adap
ted hmms
for unusual ev
ent detection
,
” I
n
Computer Vision and Pattern
Recogn
ition
,
20
05. CV
PR
2005.
IEEE
Computer
Society Con
f
erence on
, vo
l. 1, p
p
. 611-618
, 200
5.
[16]
Jiang F.,
et al.
,
“Abnor
mal even
t detection from su
rveillance v
i
d
e
o b
y
d
y
n
a
mic
hierar
chical
clus
tering
,” In
Imag
e
Processing, 200
7. ICIP 2007. IE
EE In
ternationa
l Conferen
ce on
,
vol. 5
,
pp
. V-14
5, 2007
.
[17]
Lin W
.,
et al.
, “
H
uman activity
recognition
for
video surveillan
ce,” In
Circuits and Systems, 20
08. ISCAS 2008.
IEEE Internatio
nal Symposium
on
, pp
. 2737-27
40, 2008
.
[18]
Chandrika L
., “
I
m
p
lem
e
ntation I
m
age Retriev
a
l
and Classificatio
n with SURF Techniqu
e,”
IJIS
ET-International
Journal of Innovativ
e Scien
c
e, Engineering
&
Te
chnology,
vo
l/issue: 1(4)
, 2014
.
Evaluation Warning : The document was created with Spire.PDF for Python.