Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
8
,
No.
6
,
D
ece
m
ber
201
8,
pp. 5
089
~5
097
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v8
i
6
.
pp50
89
-
5
097
5089
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Pos
t
Eve
nt Invest
igat
i
on o
f Multi
-
s
tream Vi
deo
Data Utilizin
g
Hadoop
Cluster
Jyoti
Pa
r
so
l
a,
Du
r
gaprasad
Ga
n
godk
ar, A
nkush Mi
ttal
Dep
ar
t
m
ent
o
f
C
om
pute
r
Scie
n
ce &
Engi
ne
eri
ng
,
Graphi
c
Era Uni
ver
sit
y
,
Indi
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ma
r 5
, 2
01
8
Re
vised
Ju
l
4
,
201
8
Accepte
d
J
ul
29
, 2
01
8
Rapi
d
adv
ance
m
ent
in
t
ec
hno
l
og
y
and
in
-
expens
ive
c
amera
ha
s
rai
sed
th
e
nec
essit
y
o
f
m
onit
oring
s
y
stems
for
surv
ei
l
la
n
c
e
applic
at
ions.
As
a
result
dat
a
ac
qu
ire
d
from
num
ero
us
ca
m
era
s
d
eplo
y
ed
for
surv
ei
llanc
e
is
tre
m
endous.
W
hen
an
ev
ent
is
t
r
igge
red
the
n
,
m
anua
l
l
y
inve
st
ig
at
ing
such
a
m
assive
dat
a
is
a
complex
ta
sk
.
Thus
it
is
esse
nti
al
to
expl
or
e
an
appr
o
ac
h
tha
t
,
c
an
store
m
assive
m
ult
i
-
strea
m
vide
o
data
as
well
as,
pro
ce
ss
the
m
to
find
use
ful
infor
m
at
ion.
To
addr
ess
the
challe
ng
e
of
stor
ing
and
proc
essing
m
ult
i
-
strea
m
vi
deo
data,
we
h
ave
used
Hado
op,
which
has
grown
int
o
a
le
ad
ing
computi
ng
m
odel
for
d
at
a
intensi
ve
ap
pli
c
at
ions.
In
th
is
pape
r
w
e
propose
a
novel
te
chn
ique
for
p
e
rform
ing
post
eve
nt
inv
esti
ga
ti
on
on
stored
surveil
l
anc
e
vid
eo
data.
Our
al
g
orit
hm
stores
vi
deo
data
in
HD
FS
in
such
a
wa
y
that
it
eff
i
ciently
ide
n
ti
fi
es
t
he
lo
ca
t
ion
of
da
ta
from
HD
FS
base
d
on
th
e
ti
m
e
of
occ
u
rre
nce
of
ev
ent
a
nd
per
form
furt
her
proc
essing
.
To
prove
eff
iciency
of
ou
r
proposed
work
,
we
have
per
for
m
ed
eve
nt
detec
ti
on
in
th
e
vide
o
b
ase
d
on
t
he
ti
m
e
per
iod
p
rovide
d
b
y
the
u
ser.
In
o
rde
r
to
e
stim
at
e
the
per
form
anc
e
o
f
our
appr
oac
h
,
we
eva
luated
th
e
storage
and
proc
essing
of
vide
o
d
at
a
b
y
v
ar
y
ing
(
i)
pix
el
resolut
ion
of
vi
deo
fra
m
e
(ii
)
si
ze
of
vid
eo
dat
a
(i
ii
)
num
ber
of
red
uce
rs
(workers)
exe
cu
ti
n
g
the
ta
sk
(iv)
th
e
num
ber
of
nodes
in
the
cl
uster.
Th
e
proposed
fra
m
ework
eff
ic
ie
n
tl
y
ac
h
ie
ve
s
pee
d
up
of
5.
9
for
la
rg
e
fi
les
of
1024X
1024
pixe
l
r
esolut
ion
vide
o
fra
m
es
th
us
m
ake
s
it
appr
opriate
for
t
he
f
ea
sib
le pra
c
t
ic
a
l
dep
lo
y
m
ent
in
an
y
app
li
c
at
io
ns
.
Ke
yw
or
d:
Hado
op D
ist
ri
bu
te
d
Fil
e Syst
e
m
Ma
p
Re
duce
Re
du
ce
rs
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
:
Jyoti
Par
s
ola
,
Dep
ar
t
m
ent
of
Com
pu
te
r
Scie
nce & E
ng
i
nee
rin
g,
Gr
a
phic
Er
a
Unive
rsity
,
De
hradun,
In
dia
.
Em
a
il
:
j
yote
e.ne
gi@
gm
ail.co
m
1.
INTROD
U
CTION
In
te
ll
igent
vi
de
o
surveil
la
nce
syst
e
m
(V
SS)
has
evo
l
ved
a
s
an
act
ive
stud
y
area
in
com
pu
te
r
visio
n
because
of
it
s
nu
m
erous
real
tim
e
app
li
cat
ion
s
f
or
s
ocial
secur
it
y.
It
intends
to
detect
,
identify
an
d
tra
ck
the
obj
ect
in
va
rio
us
vi
deo
fr
am
es
or
im
age
sequ
e
nce.
T
he
m
o
t
ive
beh
i
nd
is
to
est
ablish
an
intel
li
gen
t
visu
al
su
r
veill
ance
sy
stem
and
rein
s
ta
te
the
tra
diti
on
al
sur
veill
ance
syst
em
du
e
to
dep
l
oym
ent
of
m
ulti
ple
cam
eras
for
c
on
ti
nu
ous
m
on
it
or
ing.
W
hen
an
e
ven
t
oc
cur
s
the
n
t
he
capab
il
it
y
to
pe
rfor
m
scal
able
an
d
ti
m
ely
analy
ti
cs
to
this
e
xtensi
ve
acc
um
ulated
data
is
a
hi
gh
prefe
ren
c
e
for
e
ver
y
int
el
li
gen
t
VSS.
Ther
e
f
or
e
t
he
m
ajo
r
chall
enges
fac
ed
by
vid
e
o
s
urveil
la
nce
sy
stem
are
;
a)
S
tora
ge
of
c
on
t
inuousl
y
incre
asi
ng
gig
a
ntic
data,
gen
e
rated
by
the
m
ulti
ple
su
rv
ei
ll
ance
cam
eras.
b)
Prom
pt
pr
oces
sin
g
of
progressively
rising
data
w
he
n
a
n
even
t i
s
trig
gered
Pr
oc
essin
g
a
nd
sto
rin
g
c
onsist
ently
grow
i
ng
da
ta
with
c
onve
ntio
nal
net
work
sto
rag
e
a
nd
data
base
syst
e
m
is
no
t
an
easy
ta
sk.
Hado
op,
w
hic
h
wa
s
ori
gi
nally
design
e
d
by
Goo
gle,
has
e
vo
l
ved
i
nto
do
m
inant
processi
ng
m
od
el
f
or
s
uch
a
pp
li
cat
io
ns
whic
h
are
data
e
xh
a
us
ti
ve
[
1].
It’s
e
xtensi
ble,
tolera
nt
to
er
r
or
an
d
sp
li
ts
and
co
py
data,
sen
ds
t
he
com
pu
ta
ti
on
w
he
re
data
r
esi
des.
Hado
op
ha
s
recei
ved
so
m
uch
rec
ogniti
on
because
of
i
ts e
asy
accessi
bili
ty
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
87
08
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
5
089
-
5
097
5090
The
str
uctur
e
of
Ha
doop
is
ver
y
rigid
s
o
it
is
no
t
trivia
l
to
dev
el
op
and
de
plo
y
the
com
plex
al
gorithm
s
to
t
he
Ma
pRe
du
c
e
m
od
el
.
Alt
hough
a
lot
of
re
search
ha
ve
be
en
pe
rfo
rm
ed
fo
r
vid
e
o
proce
ssing
[2]
-
[6]
with H
a
doop,
ye
t
it
has
not
b
ee
n
util
iz
ed
f
or p
os
t
e
ve
nt
in
vestigat
io
n.
T
hus
t
he
m
otivati
on
o
f
our
w
ork
is give
n
as
foll
ow
s:
Handle m
ulti
ple stream
s f
ro
m
v
a
rio
us
s
urvei
ll
ance cam
eras.
Stor
a
ge
a
nd ti
m
el
y analy
sis of exte
ns
ively
accum
ulate
d
m
assive
data to
iden
ti
fy a
n
e
ve
nt of i
nterest.
Sp
ee
d up in
pe
rfor
m
ance
We
ha
ve
use
d
Hado
op
for
st
or
a
ge
an
d
proc
essing
si
ng
le
s
tream
su
rv
ei
ll
a
nce
data
on
a
sing
le
no
de
cl
us
te
r
as
disc
us
se
d
in
[
7].
I
n
this
pa
pe
r,
we
pro
po
se
a
f
ram
ewo
rk
f
or
po
st
eve
nt
in
ve
sti
gation
a
s
s
how
n
in
Fig
ure
1,
w
hi
ch
st
or
es
the
m
ult
i
-
stream
data
acc
um
ulate
d
from
m
ulti
ple
cam
eras
de
plo
ye
d
for
vid
e
o
su
r
veill
ance a
ppli
cat
ion
, i
nto
t
he HD
FS.
If an
ev
e
nt o
cc
urs t
hen,
us
er
sends
query to
an
al
y
ze the r
e
quire
d data
al
ong
with
th
e
tim
e
du
rati
on
w
hen
the
e
ve
nt
is
s
uspect
ed
to
occur.
Ba
s
ed
on
the
ti
m
e
durati
on,
the
syst
e
m
identifie
s
the
locat
ion
of
th
e
data
residi
ng
i
n
the
Data
Nod
e
in
the
cl
us
te
r
an
d
com
pu
ta
ti
on
is
execu
te
d
by
Hado
op
Ma
pR
edu
ce
.
O
ur
pr
opose
d
a
ppro
ac
h
for
post
eve
nt
inv
est
igati
on
of
s
uch
a
m
ass
ive
data,
overc
om
es
the
nee
d
f
or
a
naly
sis
of
e
ntire
data
ge
ner
at
ed
by
the
set
of
vid
e
o
cam
eras
de
plo
ye
d
f
or
m
on
it
or
in
g
purpose
thu
s
re
du
ci
ng the c
om
pu
ta
ti
on
ti
m
e.
The
r
em
ai
nin
g
pap
e
r
is
a
rr
a
ng
e
d
i
n
the
f
ollow
i
ng
way,
sect
ion
II
dis
cusses
t
he
rel
evan
t
st
ud
y
perform
ed
by
var
i
ou
s
re
sear
cher
s
.
S
ect
ion
III
disc
us
ses
the
arc
hitec
tur
e
of
post
eve
nt
in
vestigat
io
n
wit
h
Hado
op
an
d
a
naly
ses
the
al
gorithm
pr
op
ose
d
for
m
ulti
-
stream
vi
deo
data
proce
ssin
g
a
nd
stora
ge
with
Hado
op secti
on
IV
w
hich
sho
ws results a
nd
perform
ance an
al
ysi
s and sec
ti
on
V discus
se
s the c
oncl
us
io
n.
Fig
ure
1
.
Fr
am
ewor
k
f
or
po
st
even
t i
nv
e
sti
ga
ti
on
with
hado
op
2.
RELATE
D
W
ORK
It
has
bee
n
show
n
in
[
1]
that
Had
oo
p
Ma
pr
e
duce
is
app
r
opriat
e
for
processi
ng
te
xt
data
wh
ic
h
require
sam
e
com
pu
ta
ti
on
to
be
perform
e
d
in
the
e
ntire
m
assive
data
residin
g
in
t
he
H
DF
S
.
Th
eref
or
e
init
ia
ll
y
Ma
pR
edu
ce
wa
s
use
d
f
or
the
pro
bl
e
m
s
li
ke
searc
hing,
s
ort
ing
la
rg
e
data,
la
rg
e
scal
e
ind
e
xing,
gr
a
ph
com
pu
ta
ti
on
[
8
]
m
a
trix
com
p
utati
on
[
9
]
.
Som
e
research
er
s
hav
e
trie
d
to
use
it
fo
r
i
m
age
proces
si
ng
a
nd
la
rg
e
scal
e
query
pr
ocessin
g
an
d
query
s
earc
h
as
well
[
10
]
,
[
11]
.
I
n
[12]
par
al
le
l
exec
ution
of
scat
te
red
datab
ase
is
perform
ed.
A
colo
red
im
age
is
con
ver
te
d
into
gray
scal
e
im
age
and
pa
ra
ll
el
y
featur
es
are
dr
a
w
n
ou
t.
High
reso
l
ution
im
ages
a
re
pro
c
essed
a
nd
fea
tures
a
re
rem
ov
e
d
with
H
adoo
p
Ma
pRe
du
ce
[
13
]
.
H
adoo
p
Ma
pRed
uce
fra
m
ewo
r
k
is
al
so
util
iz
ed
for
app
li
c
at
ion
li
ke
i
m
age
retrieval
based
on
the
con
te
nt
[14].
An
i
m
age
ref
inem
ent
m
et
ho
d
with
Ma
pRe
du
ce
is
discu
ssed
in
[
15]
.
It
nee
ds
i
m
ages
to
b
e
st
ream
ed
on
ly
once
com
par
ed
to
ot
her
file
syst
e
m
wh
ic
h
ne
eds
e
ach
ti
m
e
entire
i
m
age
or
par
t
of
im
age
stream
ed
after
a
pp
l
yi
ng
filt
er.
[
16
]
Pro
po
s
ed
a
dist
ribu
te
d
im
age
proces
sin
g
syst
e
m
nam
ed
SEIP
,
w
hich
is
buil
t
on
Ha
doop,
a
nd
e
m
plo
ys
exten
sible
in
node
arch
it
ect
ure
to
sup
port
va
r
iou
s
kinds
of
i
m
age
proce
ssing
al
gorith
m
s
on
distrib
uted
plat
form
s w
it
h
G
P
U
acce
le
rat
or
s
.
[17
]
Hav
e
use
d hado
op for cl
us
te
rin
g
cat
e
go
rical
d
at
a.
Few
resea
rc
h
inv
e
sti
gators
ha
ve
us
e
d
vi
deo
data
processi
ng
[18]
(v
i
de
o
transc
odin
g)
wit
h
Hado
op
Ma
pRed
uce
f
r
a
m
ewo
r
k
as
di
scusse
d
by
[2]
.
[3
]
Pe
rfor
m
s
pa
rall
el
vid
e
o
a
naly
sis
and
processi
ng
w
her
ea
s
vid
e
o
play
in
g,
sh
ari
ng
a
nd
stora
ge
[
4]
w
it
h
Hado
op
cl
us
te
r.
Ha
doop
has
bee
n
use
d
f
or
la
r
ge
vid
e
o
m
anag
em
ent
[
5]
an
d
f
or
O
bj
ect
detect
io
n
an
d
cl
ass
i
ficat
ion
[6
]
.
T
he
w
ork
im
ple
m
ented
in
[
19
]
is
the
distrib
uted
vis
ual
enh
a
ncem
e
nt
us
in
g
histo
gram
equ
al
iz
at
ion
al
gorit
hm
on
i
m
age
datab
ase
fr
om
su
rv
e
il
la
nce
ca
m
eras.
T
he
e
xp
e
rim
ent is cond
ucted
i
n ps
eudo dist
rib
ute
d
m
od
e
unde
r Ha
do
op MapR
edu
ce
arc
hitec
ture
.
Vn
V2
V1
…
O
u
t
p
u
t
V
i
d
e
o
S
tr
e
a
m
s
Us
e
r
Q
u
e
r
y
S
la
v
e
n
S
la
v
e
1
M
a
ste
r
S
la
v
e
2
S
la
v
e
i
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
& C
om
p
Eng
IS
S
N: 20
88
-
8708
Pos
t
Eve
nt Inv
est
iga
ti
on
of M
ulti
-
s
tre
am Vid
eo Da
t
a Uti
li
zi
ng Ha
doop Cl
us
te
r
(
Jyoti
P
arsola
)
5091
Hen
ce
base
d
on
the
a
bove
li
te
ratur
e
s
urv
ey
it
is
evident
that
Hado
op
Ma
pRed
uce
has
no
t
bee
n
util
iz
ed
for
po
st
even
t
inv
e
sti
gation
a
pp
li
ca
ti
on
s
an
d
m
or
eov
e
r
the
pro
bl
e
m
of
m
ulti
ple
strea
m
s
storag
e
an
d
processi
ng of s
urveil
la
nce
data is sti
ll
a ch
al
le
ng
e
.
3.
ARCHITE
CT
UR
E
F
OR
H
ADOOP
A
N
D
A
N
ALY
S
I
S
OF
MU
LT
I
STREA
M
VID
E
O
D
AT
A
US
I
NG HA
D
OOP
We
hav
e
de
sign
ed
our
e
ve
nt
inv
est
igati
on
syst
e
m
based
on
a
Ma
pRe
du
ce
f
ram
ewo
rk
f
or
data
stora
ge
a
nd
d
at
a
proces
sin
g.
Ha
do
op
is
a
n
op
e
n
sourc
e
s
oft
war
e
f
ram
ewo
r
k
li
cen
sed
by
an
a
pac
he,
it
is
use
d
for
distrib
uted processi
ng
[2
0]
,
[21]
,
[
26]
,
[27]
and
d
ist
rib
uted
de
posit
or
y of
e
xtensiv
e
da
ta
set
acro
ss gr
oup
of
nodes
bu
il
d
on
low
pr
ic
e
d
com
pu
te
rs.
Trait
s
intrinsi
c
to
Hado
op
are
data
part
it
ion
ing
a
nd
par
al
le
l
com
pu
ta
ti
on
of
la
rg
e
dat
aset
s.
Its
stora
ge
a
nd
com
pu
ta
ti
on
al
capab
il
it
ie
s
scal
e
with
the
add
it
io
n
of
hos
ts
to
a
Hado
op
cl
us
te
r,
a
nd
can
rea
ch
vo
l
um
e
of
siz
es
in
the
pe
ta
byte
s
on
cl
us
te
rs
with
t
housa
nds
of
ho
sts.
It
com
pr
ise
s
of
t
wo
pr
i
ncipal
el
e
m
ents
as
discuss
e
d
in
[
7].
First
is
Hadoop
Distrib
ute
d
Fil
e
Syst
e
m
(H
D
FS
)
us
e
d
f
or
distri
bu
te
d
file
syst
e
m
,
seco
nd
is
Ma
pRed
uce
wh
ic
h
is
the
e
xecu
ti
on
e
ngin
e
or
data
proc
essin
g
fr
am
ewo
r
k
as
s
how
n
in
Fig
ure
2
.
The
a
naly
sis o
f
the m
ulti
-
stream
v
ideo
data
usi
ng H
a
doop is
don
e
in
t
hr
ee
diff
e
re
nt phase
s
Stor
i
ng the m
ulti
-
strea
m
v
ide
o data t
o
t
he H
DF
S
Pr
oc
essin
g
m
ulti
-
stream
v
ideo
d
at
a
with m
apr
ed
uce
Accu
m
ulati
ng
al
l t
he
res
ults a
nd d
is
play
ing t
he result.
In
VS
S
,
data
a
ccum
ulate
d
from
var
iou
s
ca
m
eras
dep
l
oyed
f
or
m
on
it
or
i
ng
pur
pose
is
m
assive
an
d
con
ti
nu
ously
keep
on
inc
reas
ing
.
Q
uestion
is
to
store
s
uc
h
an
e
xtrem
ely
la
rg
e
data.
More
ov
e
r
the
issue
beco
m
es
m
or
e
com
plex
wh
e
n
an
e
ve
nt
is
trigg
e
red
a
nd
the
data
is
to
be
pro
cesse
d
to
ext
ract
the
us
ef
ul
inf
or
m
at
ion
re
gardin
g
a
ny
e
ven
t.
Tra
diti
on
al
m
et
ho
d
us
e
d
for
e
xtracti
ng
us
ef
ul
in
f
orm
at
ion
is
to
c
heck
th
e
entire
databas
e,
w
hic
h
is
c
om
pu
ta
ti
on
al
ly
exp
e
ns
i
ve.
Th
ere
s
houl
d
be
so
m
e
m
easur
e
w
her
e
the
use
r
c
a
n
search
the
pa
rtic
ular dat
a
base
d
on the tim
e o
f occu
rr
e
nce
of eve
nt,
in
ste
ad of
searc
hing t
he
en
ti
re
databa
se.
Ther
e
f
or
e
f
or
t
his
pur
pose
Ha
doop
H
DF
S
is
us
e
d
.
Data
in
HDFS
is
proce
ssed
in
batc
hes
.
Ther
e
f
or
e
stream
s
are
buff
e
red
i
nto
loc
al
m
e
m
or
y
and
the
n
data
i
s
trans
ferred
i
nto
the
HDFS.
More
ov
e
r
Hadoop
wa
s
or
i
gin
al
ly
design
e
d
f
or
te
xt
processi
ng
th
us
,
there
is
no
su
pp
or
t
in
Ha
doop
f
or
proc
essing
vid
e
o
da
ta
.
We
extract
fr
am
es
from
vid
eo
strea
m
and
store
them
as
Seq
ue
nce
fi
le
in
the
HDFS
.
Se
que
nc
e
fi
le
s
are
Had
o
op
par
ti
cula
r
archi
ve
file
la
youts
ver
y
m
uch
li
ke
to
ta
r
and
zi
p.
It
bri
ngs
to
gethe
r
the
file
set
with
a
key
and
a
value
com
bin
a
ti
on
wh
e
re
key
is
the
file
nam
e
an
d
value
is
the
co
nte
nt
of
file
.
T
he
gen
e
r
at
ed
se
qu
e
nce
file
is
m
os
tly
half
the
siz
e
of
the
or
i
gin
al
data
an
d
hen
ce
ta
ke
s
lim
it
ed
m
e
m
or
y
area
in
HD
F
S
conver
ti
ng
it
s
tora
ge
eff
i
ci
ent.
T
hes
e
file
s
can
be
separ
at
e
d
an
d
processe
d
in
pa
rall
el
.
Fo
r
vide
o
analy
ti
c
app
li
cat
ion
s
li
ke
m
ot
ion
detect
ion,
rat
he
r
tha
n
c
om
par
in
g
eve
ry
al
te
rn
at
e
fr
am
e
it
is
su
f
fici
ent
to
process
e
ve
ry
al
te
rn
at
e
fift
h
fr
am
e
[22
].
3.1.
Data S
to
r
age
R
at
her
tha
n
stori
ng
al
l
the
vi
deo
fr
am
es
we
store
eve
ry
al
te
rn
at
e
fift
h
fra
m
e
wh
ic
h
f
urt
her
re
duce
s
the
sto
rag
e
spa
ce
in
HDFS.
W
e
us
e
a
novel
te
ch
nique
f
or
sto
rin
g
the
se
qu
e
nce
file
s
(
fr
am
es)
us
i
ng
Algorithm
1
.
Ever
y
vi
deo
ca
m
era
is
identifie
d
with
a
un
i
qu
e
ide
ntifie
r
li
ke
V1,
V
2,
…
.
,Vn
w
hile
storin
g
the
seq
uen
ce
file
we
ge
ner
at
e
th
e
nam
e
of
seq
ue
nce
file
by
co
ncatenati
ng
ca
m
era
identifie
r
,
date,
tim
e
and
fr
am
e
nu
m
ber
,
eg
.
V
1_1_07
_2016_
10_12_
22_
1
w
her
e
V
1
is
t
he
nam
e
of
t
he
c
a
m
era
or
strea
m
,
1_
07_2
016
is
the
date
(
Day_M
onth
_Y
ea
r
form
at
)
,
10
_12_22
is
the
tim
e
(Hour
_Minu
te
s
_Sec
onds
f
or
m
at
)
an
d
1
is
the
fr
am
e
nu
m
ber
.
T
his
appr
oach
f
aci
li
ta
te
s
app
r
opria
te
identific
at
ion
of
DataN
ode
wh
e
re
the
fr
a
m
e
has
been
s
tore
d.
Th
us
overc
omi
ng
the
tim
e
r
equ
i
red
to
sea
rch
entire
data
accum
ulate
d
so
far.
This
da
ta
in
the
HDFS
is
separ
at
e
d
into
blo
c
ks
(
def
a
ult
siz
e
of
bl
ock
i
s
64
M
b
)
a
nd
f
ur
t
her
sto
re
d
i
n
va
rio
us
no
de
s
of
the
cl
us
te
r
.
Each
blo
c
k
is
rep
li
ca
te
d
with
3 co
pi
es
in t
he
m
achines
of the cl
us
t
er.
Alg
o
rith
m
1:
Data
sto
rage in HDF
S
Inp
u
t:
m
u
ltip
le
vid
eo
strea
m
s
Ou
tp
u
t: Sequ
en
ce fil
e
Step
1:
Fo
r
ev
er
y
v
id
eo
f
ra
m
e
(
)
o
f
Sto
re
with
the n
a
m
e as
+ date + ti
m
e+
Fra
m
e
n
u
m
b
er;
Vs +5
;
Step
2:
Co
n
v
ert
st
o
red
to
seq
u
en
ce fil
e
Step
3:
Co
p
y
S
eq
u
en
ce fil
e to th
e H
DFS
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
87
08
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
5
089
-
5
097
5092
3.2.
Data Pr
ocessi
ng
The
Ha
doop
Ma
pRed
uce
w
ork
flo
w
as
show
n
in
Fig
ure
3,
us
e
r
ente
rs
the
query
with
the
tim
e
and
sen
ds
it
to
the
m
ast
er.
In
the
m
ast
er
Jo
btrac
ker
div
ide
s
the
j
ob
to
va
rio
us
ta
sk
s
an
d
sen
ds
it
to
the
Tasktrack
er
residin
g
in
the
sla
ve
nodes
.
The
ta
sk
s
are
ex
ecut
ed
by
the
m
ap
and
re
duc
e
fu
ncti
on
res
pe
ct
ively
.
The
proces
s
of
Data
Node
identific
at
io
n
and
data
proce
ssing
al
gorith
m
is
fu
rthe
r
di
scusse
d
in
Algorithm
2
Re
s
ults
is
accum
ulate
d
by
the
Tas
kTr
ac
ker
an
d
final
outp
ut
is
ge
ne
ra
te
d.
T
o
pro
ve
t
he
e
ff
ic
ie
ncy
of
ou
r
propose
d
w
ork
we
are
fin
ding
m
otion
in
the
vid
e
o
data
on
the
basis
of
t
he
tim
e
of
occ
urren
ce
of
eve
nt
as
pro
vid
e
d
by
the
us
er
. Movi
ng obj
ect
detect
io
n base
d on m
otion
se
gm
entat
ion
is it
sel
f
a c
halle
ng
e
in VS
S.
A
lot
of
resear
ch
ha
ve
bee
n
pe
rfor
m
ed
for
m
o
t
ion
segm
entat
ion
a
nd
has
been
broad
ly
cl
assifi
ed
int
o
Ba
ckgrou
nd
s
ub
t
racti
on
a
nd
tem
po
ral
dif
f
eren
ci
ng.
I
n
ba
ckgr
ound
sub
tract
ion
m
otion
[
23
]
is
dete
ct
ed
by
fin
ding
the
di
f
fer
e
nce
betwee
n
the
prese
nt
f
ram
e
and
the
r
efere
nce
bac
kgrou
nd
w
her
ea
s
in
case
of
t
em
poral
diff
e
re
nce
[
24
]
m
otion
is
dete
rm
ined
by
cal
culat
ing
the
pixe
l
wise
diff
e
re
nce
bet
ween
t
he
pr
ese
nt
fr
am
e
and
the
earli
er
f
ra
m
e.
The
m
otion
detect
io
n
a
lgorit
hm
as
propose
d
i
n
[
25]
,
is
use
d
in
our
syst
em
fo
r
fin
ding
m
ov
ing
o
bject
s.
O
ne
of
the
e
ff
ic
ie
n
t
m
et
hods
o
f
f
ram
e
diff
ere
ncin
g
is bl
ock
m
at
ching
, f
or
ide
ntifyi
ng m
ov
ing
obj
ect
s.
In
bloc
k
m
at
ching
a
s
s
how
n
i
n
Fi
gure
7
the
c
urren
t
f
ram
e
is
div
ide
d
int
o
bl
ocks
a
nd
sim
ilarly
pr
e
vious
f
ram
e
is
al
so
di
vid
e
d
i
nto
bl
ock
s
a
nd
the
b
loc
ks
of
are
sea
rch
e
d i
nto
so
if
a b
lo
ck
of
is
fou
nd in dif
fere
nt p
i
xel locati
on in
it
i
m
plies
that m
otion
is
pro
du
ce
d.
Fig
ure
3
.
Ma
p red
uce
job exe
cution fl
ow for
m
otion
d
et
ect
ion
Alg
o
rith
m
2:
Iden
tif
icatio
n
of
DataNod
e &
data P
roces
sin
g
Inp
u
t: Data,
ti
m
e
a
n
d
du
ration
o
f
an even
t
Ou
tp
u
t: Resu
ltan
t m
o
tio
n
vecto
r
.
Step
1:
Use
r
sen
d
s th
e qu
ery
in th
e
f
o
r
m
of
date and
ti
m
e
to th
e
m
aste
r.
Step
2:
M
aster
se
n
d
s th
e co
m
p
u
tatio
n
to th
e Jo
b
Tr
acke
rand
Na
m
e
No
d
e
id
en
tif
ies
d
ata locatio
n
Step
3:
Jo
b
Tr
acke
r
sp
lits th
e job
into
the Task
T
racker
Step
4:
The
co
m
p
u
tatio
n
is execu
ted
f
u
rther by
m
ap
an
d
r
ed
u
ces in
the
d
ata
r
esid
i
n
g
in th
e data no
d
e
and
r
esu
lt is sen
t
b
ack to
the
Task
Tr
acker
.
Step
5:
Jo
b
Tr
acke
r
accu
m
u
l
ates th
e
r
esu
lt f
ro
m
T
ask
T
ra
ck
er
an
d
f
o
rwar
d
s
it to th
e
m
aster
Ma
ny
te
chn
iq
ues
ha
ve
been
propose
d
by
var
i
ou
s
sc
hola
rs
f
or
perform
ing
m
at
ching
com
pu
ta
ti
on.
Su
m
of
Ab
s
olut
e
Diff
e
re
nce
(
SAD)
is
us
e
d
i
n
[
25
]
f
or
m
easur
i
ng
the
dif
f
eren
ce
bet
wee
n
the
tw
o
vi
de
o
f
ram
e
blo
c
k,
a
s it
is hi
gh
ly
ef
fici
ent.
The
l
ow
e
r valu
e o
f
S
AD m
ea
ns
the
h
i
gh
e
r
si
m
il
arity b
et
we
en
the
tw
o bloc
ks
. It
is cal
culat
ed us
ing
e
q.1.
(
1
,
2
)
=
∑
∑
|
(
,
,
1
)
−
=
−
1
=
0
(
+
1
,
+
2
,
2
)
|
=
−
1
=
0
(1)
I
n
p
u
t
Jo
b
C
h
u
n
k
S
l
a
v
e
Jo
b
M
a
st
e
r
U
se
r
Jo
b
C
h
u
n
k
M
a
p
Da
ta
S
p
li
t
Da
ta
S
p
li
t
R
e
d
u
c
e
R
e
du
c
e
M
a
p
M
a
p
O
u
t
p
u
t
H
D
F
S
F
r
a
m
e
1
F
r
a
m
e
2
F
r
a
m
e
n
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
& C
om
p
Eng
IS
S
N: 20
88
-
8708
Pos
t
Eve
nt Inv
est
iga
ti
on
of M
ulti
-
s
tre
am Vid
eo Da
t
a Uti
li
zi
ng Ha
doop Cl
us
te
r
(
Jyoti
P
arsola
)
5093
wh
e
re
(
1
,
2
)
={
-
z
≤
1
,
2
≤ z
}
Moti
on v
ect
or
(m
v)
= (
1
,
2
))
|m
i
n
sa
d (
1
,
2
)
)
(2)
In
(1)
&
(
2)
(
,)
a
nd
(
,
)
dis
play
s
the
intensit
y
of
pi
xel
s
in
the
earli
er
an
d
prese
nt
fr
am
e
su
bse
que
ntly
.
SAD(
1
,
2
)
is
the
total
value
of
abs
olu
te
diff
e
r
ence
at
the
pi
xe
l
locat
ion
1
,
2
[
-
z,
z]
is
the
search
a
rea in
t
he
searc
h win
dow
. m
v
ind
ic
at
es the m
otion
v
ect
or at sm
all
est
r
at
e o
f SA
D
com
pu
te
d
i
n fr
am
es
distant
with
ti
m
e
t
1
and
t
2
.
Dim
ension
of
m
ot
ion
blo
c
k
i
s
ch
os
e
n
for
16
x
16
a
nd
sea
rc
h
window
siz
e
is
of
32
x32 pixels
.
The
jo
b
of
co
m
pu
ti
ng
m
otio
n
vect
or
in
t
he
scop
e
betwe
en
[
-
31,
32
]
i
nto
f
ram
es
of
a
vi
deo
is
com
pu
ta
ti
on
al
ly
hig
h
pr
ic
e
d,
as
a
resu
lt
sear
ch
is
sta
rted
w
hen
,
m
otion
bl
ock
is
on
t
he
s
a
m
e
po
sit
ion
of
the
ref
e
ren
ce
f
ram
e
i.e.
blo
ck
a
nd
searc
h
wind
ow
are
c
oin
ci
di
ng
in
m
idd
le
.
If
there
is
no
change
the
n
val
ue
of
SAD
is
zero
a
nd
if
a
bl
ock
i
nc
lud
es
m
ov
em
e
nt
then
blo
c
k
presents
the
m
axim
u
m
value
of
abs
olu
te
d
iffe
ren
ce
SAD
o
is c
om
pu
te
d wit
h
t
he
e
x.(3):
0
=
∑
∑
|
(
,
,
1
)
−
=
−
1
=
0
=
−
1
=
0
(
,
,
2
)
(3)
wh
e
re a
, b i
ndic
at
es p
os
it
io
n of pi
xel in ea
rlie
r
(
ref
e
re
nce) a
nd prese
nt
f
ra
m
e.
Adding
to
this
a
thres
hold
(
th
)
is
e
nforce
d
to
SAD
o
to
decr
ease
proce
ssing
ti
m
e.
It
helps
t
o
dete
r
m
ine
wh
et
her to
i
niti
al
iz
e the searc
h or n
ot
on
the
basis
of ex
.
(
4).
Searc
h Deci
sion
=
(4)
Fo
r
eac
h
bl
oc
k
m
i
ld
th
can
be
fixed
as
par
t
of
25
6
X
15=
3840
w
her
e
,
256
is
16
X
16
blo
c
k
pi
xel
value. Pa
rt v
al
ue
li
es
within t
he ran
ge of
( 0.
4
,
0 .1).
The
resu
lt
a
nt
is
a
set
of
m
otion
vect
or.
Acc
um
ulati
ng
al
l
the
re
su
lt
s
for
e
ver
y
fr
am
e
of
a
vid
e
o
final
m
ot
ion
is
pl
otted.
I
n
orde
r
to
fin
d
out
the
m
ot
ion
detect
ion
w
e
us
e
SAD
to
detect
th
e
m
otion
in
th
e
vi
de
o
fr
am
es
as
afor
e
m
entione
d.
Ma
p
f
un
ct
io
n
read
s
t
wo
fra
m
es
as
an
in
put
an
d
s
plit
s
each
f
ram
e
into
32
by
32
pix
el
siz
e
blo
c
ks
a
nd
each
bl
ock
is
assig
ne
d
a
key
a
nd
val
ue
c
on
ta
ini
ng
the
32
by
32
bl
ock
an
d
t
his
outp
ut
is
cal
le
d
as
interm
ediat
e
data.
Each
key
val
ue
pair
is
passe
d
to
the
re
du
c
e
functi
on
in
s
uch
a
m
ann
er
that
the
values
co
ntaini
ng
the
sam
e
ke
y
is
passe
d
to
the
sam
e
reduc
er.
T
he
ta
s
k
of
m
otion
detect
ion
is
pe
rfo
rm
e
d
by
the r
e
duce f
un
ct
ion
a
n
d i
t fin
ds
t
he
m
ov
in
g object
base
d o
n
m
otion
se
gme
ntati
on
.
3.3.
Result
A
cc
um
ulat
i
on
Finall
y
al
l
the
resu
lt
s
c
om
pu
te
d
by
va
rio
us
re
ducers
f
or
al
l
the
bl
oc
ks
of
t
he
vi
de
o
f
ram
e
are
accum
ulate
d
a
final
outp
ut
i
s
obta
ine
d
dis
play
ing
the
m
ov
i
ng
obj
ect
on
t
he
c
orres
pond
i
ng
vid
e
o
fra
m
es.
Algorithm
3
s
hows
t
he
data
accum
ulati
on
process
.
T
his
a
ppr
oach
ca
n
be
us
e
d
to
detect
the
eve
nt
in
m
ulti
ple
stream
s
wh
ere
the
po
s
sible
ti
m
e
of
occ
urre
nce
of
the
ev
e
nt
is
pro
vid
e
d
by
the
use
r.
It
can
be
obser
ve
d
from
the abo
ve
e
xp
l
anati
on that
ou
r propose
d
a
pp
ro
ac
h
ca
n
ac
hi
eve th
e
foll
owing
Eff
ic
ie
nt
Sto
r
age
of
m
ulti
-
stream
vid
eo
data
acc
um
ula
te
d
f
ro
m
num
ero
us
cam
eras
de
plo
ye
d
f
or
su
r
veill
ance i
nto
the
HDF
S.
Extract
data
ba
sed o
n
the
tim
e
of
occurre
nce
of eve
nt
prov
i
de
d by the
us
e
r.
Ana
ly
ti
cs of th
e m
assive d
at
a
with Ma
pRed
uc
e in s
hort ti
m
e.
Alg
o
rith
m
3: Dat
a
Accu
m
u
latio
n
Inpu
t
:
Motio
n
Ve
cto
r
f
o
r
ev
ery f
ra
m
e
O
utp
ut
:
Mov
in
g
Object
Step1
:
Motio
n
vecto
rs ar
e
ob
tain
ed
f
o
r
ev
ery
set
of
v
id
eo
f
ra
m
es.
Step
2
: Resu
lts ar
e
accu
m
u
lat
ed
.
Step
3
: M
o
tio
n
vecto
rs ar
e
plo
tted
ac
co
rdin
g
to th
e us
er
q
u
ery
4.
RESU
LT
S
&
PERFO
R
MANC
E E
V
ALU
ATIO
N
We
hav
e
ana
ly
zed
perform
ance of
our pro
posed fram
ewo
r
k i
n
the
foll
owin
g
m
ann
er;
By
m
easur
in
g
t
he
c
om
pu
ta
ti
on
ti
m
e fo
r vary
ing
1
If
SAD
o
<th
0
oth
erwise
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
87
08
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
5
089
-
5
097
5094
1)
nu
m
ber
of r
e
ducers
(w
orker
s
)
2)
nu
m
ber
of no
de
s buil
ding cl
ust
er
Com
pu
ta
ti
on
al
eff
ic
ie
nc
y f
or
higher
p
i
xel r
e
so
luti
on
vid
e
o fr
am
es b
y va
ryi
ng
t
he
siz
e
of
vid
e
o fr
am
e.
(a)
(b)
(c)
(d)
Fig
ure
4
.
Sh
adow
[
28]
(
a
) or
i
gin
al
fr
am
e (b)
seg
m
ented
m
ov
in
g object a
nd
B
as
el
ine
[
28
]
(
c
) Ori
gin
al
fr
a
m
e
(
d
)
Se
gm
ented m
ov
ing
obj
ect
(a)
(b)
(c)
(d)
Fig
ure
5
.
I
nter
mitt
ent Objec
t
Motio
n
[
28]
(
a
) or
igi
nal fram
e (
b)
segm
ented
m
ov
in
g obje
ct
an
d
G
rou
nd
Tru
th
[29] (
c
) Ori
gina
l fr
am
e (
d
)
Se
gm
ented
m
ov
ing o
bject
(a)
(b)
(c)
(d)
Fig
ure
6
.
CAVI
AR_
Meet
_Wa
l
kToget
her
1
[30
]
(a)
Or
i
gin
al
f
ram
e (b
) Segm
ented
m
ov
i
ng
obj
ect
a
nd
CA
VIAR
_
walk3
[30]
(
c
) Ori
gin
al
fram
e
(
d
) Segm
ented
m
ov
ing
obj
e
ct
The
pro
posed
work
is
i
m
plem
ent
ed
on
I
nt
el
cor
e
i5
3.10
GH
z
with
4
GB
of
m
e
m
or
y
on
Ubu
ntu
12.04
as
a
n
op
erati
ng
syst
e
m
Hado
op
ve
rsion
is
1.2
.1.
We
hav
e
us
e
d
5
a
nd
10
nodes
cl
ust
er
f
or
perf
orm
ance
evaluati
on. W
e
hav
e
e
valuate
d
pe
rfor
m
ance
by
var
yi
ng
file
siz
e
an
d
data
s
iz
e
of
a
cl
us
te
r
and
it
s
aff
ect
o
n
the
com
pu
ta
ti
on
ti
m
e.
Detect
ion
of
m
otion
is
dri
ven
on
vid
e
o
fr
am
es
(graysc
al
e)
with
pix
el
siz
e
256
X
25
6,
512
X
512
a
nd
1024
X
10
24.
C
olored
vi
deo
f
ram
e
are
conv
erted
t
o
gr
ay
s
cal
e
befor
e
pr
ocessin
g.
Ca
lc
ulate
d
mo
t
ion
detect
ion o
n
var
i
ou
s
vid
e
o
sequ
e
nces
is
dis
play
ed
on
fig
ure
4
to
6.
Ja
va
CV
wh
ic
h
i
s
wr
ap
per
f
or
O
pe
nCV
li
br
a
ry
[31
]
is
us
ed
to
plo
t
m
otion
vector
.
Fi
g
ure
4
to
6
show
s
th
e
or
igi
nal
fr
am
e
and
the
c
orre
sp
on
ding
segm
ented
fr
am
es
wh
e
re
m
ot
ion
is
i
dent
ifie
d
.
E
xperi
m
ent
is
com
pute
d
on
sta
nda
rd
dataset
acce
ssible
ope
nl
y
Chan
ge
Detect
ion
Be
nc
hm
ark
[
28
]
,
Lab
or
at
ory
f
or
Im
age
&
Me
dia
U
ndersta
ndin
g
(LI
M
U)
[29
]
,
C
on
te
xt
Aw
a
re
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
& C
om
p
Eng
IS
S
N: 20
88
-
8708
Pos
t
Eve
nt Inv
est
iga
ti
on
of M
ulti
-
s
tre
am Vid
eo Da
t
a Uti
li
zi
ng Ha
doop Cl
us
te
r
(
Jyoti
P
arsola
)
5095
Visio
n
Using
I
m
age
-
Ba
sed
A
ct
ive
Re
cogniti
on
(CA
VIAR)
[30
]
.
For
analy
zi
ng
Ma
pRe
du
ce
perfor
m
ance
for
m
ot
ion
detect
i
on
al
gorithm
,
Hado
op
Ma
pR
edu
ce
Cl
us
te
r
i
s
est
ablishe
d.
The
cl
us
te
r
co
ns
ist
s
of
9
sal
ve
nodes
and 1 m
ast
er n
od
e
.
Our
propose
d
fr
am
ewo
r
k
e
ff
i
ci
ently
reduces
the
sto
rag
e
spa
ce
in
H
DFS
a
nd
the
resu
lt
s
of
data
siz
e
reducti
on
are
s
how
n
in
Ta
ble
1
.
Fi
rst
col
umn
of
the
ta
ble
r
epr
ese
nts
the
ori
gin
al
data
siz
e,
seco
nd
col
um
n
is
the
data
siz
e
w
hich
is
re
duce
d
w
he
n
only
al
te
rn
at
e
fift
h
f
r
a
m
e
is
stored
t
he
data
reducti
on
pro
duced
is
about
80
-
85
%
an
d
thir
d
colum
n
disp
la
ys
the
data
siz
e
red
uctio
n
achieved
by
the
com
pr
essio
n
due
to
the
sequ
e
nc
e
file
ge
ner
at
io
n
and
the
c
om
pr
ession
is
ab
out
80%.
The
res
ult
cl
early
sho
ws
the
ef
fici
en
cy
of
our
a
ppr
oach
i
n
te
rm
s o
f
sto
rage.
Table
1
.
Stora
ge
sp
ace
r
e
du
ct
i
on in H
DF
S
.
Origin
al Data
Size
Red
u
ced Data size
Seq
u
en
ce Fil
e co
m
p
ressed
data
5
0
0
M
B
1
0
0
M
B
2
0
M
B
1
GB
2
0
4
.8 MB
4
0
.96
M
B
1
.5 GB
3
0
7
.5 MB
6
1
.6 MB
2
GB
4
0
9
.6 MB
8
1
.92
M
B
2
.5 GB
5
1
1
M
B
1
0
3
.2 MB
4.1.
Perfo
r
ma
nce
Eva
lu
at
i
on
on Mul
ti Node
Cl
ust
er
1)
A
naly
sing
ta
sk
e
xec
ution t
i
m
e w
it
h
va
ryi
ng num
ber
of
node
s in
the cl
ust
er.
The
e
xtensi
bili
ty
and
rob
us
tn
ess
of
the
fram
ewor
k
is
eval
ua
te
d
by
a
naly
sing
the
m
ulti
st
ream
vid
eo
data
on
va
rio
us
no
des
of
the
cl
us
te
r
.
E
xper
i
m
ent
is
exec
ut
ed
with
diff
e
r
ent
nu
m
ber
of
node
s
t
o
be
a
ble
t
o
unde
rstan
d
s
pe
ed up. Pa
rall
el
sp
ee
d up
S
p
is
m
e
asur
e
d
a
s
giv
e
n by e.
q.
(
5
)
=
1
(5)
wh
e
re
T
1
is
th
e
total
e
xecu
ti
on
ti
m
e
cal
cul
at
ed
in
on
e
node
cl
us
te
r
a
nd
T
n
is
the
total
exec
utio
n
ti
m
e
cal
culat
ed
in
n
node
cl
ust
er
wer
e
n
>
1.
va
lue
of
S
p
s
hows
the
num
ber
of
tim
es
par
al
le
l
execu
ti
on
is
fa
ste
r
than
r
unning
t
he
sam
e
Ma
pRedu
ce
al
gorith
m
on
the
si
ng
l
e
node
cl
us
te
r
.
I
f
it
is
gr
eat
er
than
1,
it
entai
ls
that
there
i
s
at
le
a
st
so
m
e
gain
from
do
in
g
t
he
w
ork
i
n
paral
le
l.
Exec
utio
n
ti
m
e
fo
r
vi
deo
f
ram
es
of
pixe
l
reso
l
ution
25
6
x
25
6,
512
X
512
a
nd
10
24
X
10
24
in
se
qu
e
ntial
(a
si
m
ple
j
ava
pro
gr
am
)
an
d
Ma
pRed
uce
cl
us
te
r of v
a
rio
us
nodes
and
c
om
p
uted
s
pee
d u
p
is s
how
n
i
n F
ig
ur
e
7.
T
he pr
ocessi
ng tim
e is the t
otal t
im
e to
cal
culat
e
m
oti
on
detect
ion
in
the
required
da
ta
siz
e
and
w
e
hav
e
searc
he
d
100
MB
data
in
the
HD
F
S
as
well
as in se
quentia
l and f
ur
t
her pe
rfor
m
ed
m
otion
detect
ion i
n
t
he respecti
ve
da
ta
.
Fig
ure
7
.
S
pee
d up f
or m
otion
detect
ion al
gorithm
o
f
a
) 2
56
X 256
pix
el
r
esolutio
n vide
o fr
am
e w
it
h
diff
e
re
nt num
ber
of
node
s in
a Ma
pRed
uce
cl
us
te
r b) 5
12
X 51
2 pixel
res
olu
ti
on
vid
e
o f
ram
e
with d
if
f
eren
t
nu
m
ber
of no
de
s in
a
MapRe
du
ce
cluster
c)
1024
X 102
4 p
ixel
res
olu
ti
on
vid
e
o fr
am
e wi
th d
if
fer
e
nt
num
ber
of no
des
in
a
Ma
pRed
uce cl
us
te
r.
4.2.
Analysin
g
T
ask Exec
ut
io
n T
im
e
by
V
ar
yin
g Numb
er
of
Redu
cer
s
(
Wor
kers
)
Perf
or
mi
ng
th
e
Job
We
ha
ve
al
so
analy
zed
the
pe
rfor
m
ance
of
m
ot
ion
detect
ion
al
gorit
hm
b
y
var
yi
ng
the
nu
m
ber
of
reducer
s
(
wor
ke
rs).
Fig
ure
8
sh
ows
t
he
ou
tc
om
e
of
dif
fer
e
nt
num
ber
of
r
edu
ce
r
for
va
ri
ou
s
volum
es
of
data
and
va
rio
us
pi
xel
siz
e
vid
e
o
fr
am
es.
W
e
al
s
o
te
ste
d
exec
ut
ion
ti
m
e
by
va
ryi
ng
m
ap
ta
sks
but
r
esults
w
ere
no
t
0
1
2
3
4
5
0
0
MB
1
GB
1
.
5
GB
2
GB
2
.
5
GB
5
nodes
(se
cs)
0
1
2
3
4
5
0
0
MB
1
GB
1
.
5
GB
2
GB
2
.
5
GB
5
node
s
1
0
no
des
0
2
4
6
8
5
0
0
MB
1
GB
1
.
5
GB
2
GB
2
.
5
GB
5
node
s
1
0
no
des
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
87
08
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
5
089
-
5
097
5096
rem
ark
able.
E
xe
cution
ti
m
e
fo
r
sm
al
le
r
data
vo
l
um
e
is
almost
sim
i
la
r
but
for
la
r
ger
data
vo
l
um
e
red
ucti
on
in
processi
ng tim
e is achie
ved c
on
si
der
a
bly.
T
he
ta
ble
cl
early
shows that i
t i
s not
necessary
that
Fo
r
lo
w res
olu
t
ion
vid
e
o fr
am
e 250
-
300 re
ducers
on a
n
a
ve
rag
e
provide
s
good
resu
lt
s.
Fo
r
h
i
gh r
e
so
l
ut
ion
vid
e
o fr
a
m
e 5
00
-
70
0
re
du
ce
rs o
n
a
n
a
ver
a
ge pr
ovide
s good res
ults.
Th
us
this
gi
ve
s
pri
or
in
form
at
ion
t
o
set
th
e
num
ber
of
r
edu
ce
rs
f
or
co
m
pu
ta
ti
on
as
f
ind
in
g
t
he
nu
m
ber
of r
e
ducers
pr
ov
i
ding e
ff
ic
ie
nt
res
ul
t i
s a tedio
us
t
ask.
(a)
(b)
(c)
Fig
ure
8
.
Moti
on d
et
ect
io
n
c
om
pu
ta
ti
on
tim
e
for
(a) 2
56
X 256
pix
el
s
(in
s
econds
)
siz
e
vi
deo f
ram
e
of
var
i
ou
s
d
at
a
siz
e w
it
h va
ryi
ng
nu
m
ber
of r
e
du
ce
rs
,
(
b) 51
2 X
512 (i
n
sec
onds)
pix
el
siz
e
vid
e
o of va
rio
us
data siz
e
with
var
yi
ng
num
ber
of
re
du
ce
rs
,
(
c)
1024
X 102
4 (in sec
onds)
pix
el
size
vid
e
o fr
am
e o
f vari
ou
s
data siz
e
with
var
yi
ng
num
ber
of
re
du
ce
rs
5.
CONCL
US
I
O
N
We
ha
ve
pro
pose
d
and
im
ple
m
ented
an
effi
ci
ent
app
r
oac
h
for
pe
rfor
m
ing
po
st
eve
nt
inv
est
igati
on
on
m
assive
volum
e
of
su
r
veill
ance
data
w
hi
ch
is
on
e
of
t
he
chall
en
ges
of
Vide
o
S
urv
ei
ll
ance
syst
e
m
.
W
e
hav
e
use
d
Ha
doop
HDFS
f
or
distri
bu
te
d
stora
ge
a
nd
Hado
op
M
ap
Re
du
ce
f
or
pa
rall
el
and
di
stribu
te
d
processi
ng
of
m
assive
accu
m
ula
te
d
m
ulti
-
stream
vid
eo
data.
We
ha
ve
pro
posed
a
n
al
gorithm
fo
r
eff
ic
ie
nt
storing
vid
e
o
da
ta
in
the
H
DFS.
He
nce
wh
e
n
an
eve
nt
is
trigg
e
re
d
we
a
uto
m
atical
ly
extract
data
base
d
on
t
he
tim
e
of
occurr
ence
of
eve
nt
a
nd
process
it
f
ur
t
her
to
fin
d
use
fu
l
in
form
at
i
on.
To
pro
ve
the
com
petence
of
our
pro
po
se
d
a
ppr
oach
i
n
ha
ndli
ng
a
nd
pro
cess
ing
e
xtrem
el
y
huge
data,
we
hav
e
im
ple
m
e
nted
m
otion
de
te
ct
i
on
al
gorithm
in
H
adoo
p
cl
ust
er.
Hado
op
cl
us
te
r
c
on
si
sts
of
m
a
xim
u
m
of
10
node
s.
O
ur
e
xpe
rim
ent
resu
lt
pr
eci
sel
y
ind
ic
at
es
that
the
com
pu
ti
ng
peri
od
is
s
horte
ne
d,
w
hen
pix
e
l
reso
l
ution
of
vid
e
o
fr
am
e
i
s
increase
d.
We
al
so
analy
zed
the
perform
ance
by
m
easur
ing
t
he
com
pu
ta
ti
on
tim
e
fo
r
va
ryi
ng
num
ber
of
re
du
ce
rs
(
wor
kers).
N
et
w
ork
la
te
ncy
al
so
aff
ect
s
th
e
e
xecu
ti
on
ti
m
e
in
a
c
luster.
To
so
lve
thi
s
issue
exec
ution
ti
m
e
can
be
fu
rt
her
im
pr
ov
e
d
.
More
ov
e
r
th
rough
the
i
ncr
e
m
ent
in
nu
m
ber
of
no
des
of
a
cl
us
te
r
,
c
ompu
ta
ti
on
tim
e
can
be
c
ut
do
wn
m
or
e
.
Our
f
ram
ewo
r
k
is
r
obus
t
a
nd
can
c
op
e
with
var
yi
ng
nu
m
be
r
of
nodes
i
n
the
cl
us
te
r
as
w
el
l
as
increasin
g
data
vo
l
um
e.
Had
oop
perform
s
e
xcell
ent
for
ap
plica
ti
on
w
hic
h
nee
d
sim
il
ar
ta
sk
to
be
perf
or
m
ed
in
disti
nct
data
siz
es
;
hen
ce
ap
plica
ti
on
re
qu
i
rin
g
dif
fer
e
nt
job
s
to
be
pe
rfo
rm
ed
in
var
ious
data
set
s
in
a
li
gn
ed
m
ann
er
is
no
t
po
s
sible
with
Hado
o
p M
apR
edu
ce
.
REFERE
NCE
S
[1]
J.
Dea
n
and
S.
Ghem
awa
t,
"
Map
Reduc
e:
Sim
pl
ifi
ed
D
ata
Proce
ss
ing
on
La
rge
Cluste
r
"
,
ACM
Comm
un
.
,
vol.
5
1,
no.
1
,
pp
.
107
–
1
13,
2008
.
[2]
R.
Pere
ira,
K.
B
rei
tman,
"
A
Clo
ud
Based
Arch
itect
ur
e
for
Im
proving
Video
Co
m
pre
ss
ion
Ti
m
e
Eff
i
ci
en
c
y
:
Th
e
Split
&Me
rge
A
pproa
ch
"
,
In
3rd
IEEE
In
t. Conf
.
on
Cloud
Comp
uti
ng
(
CL
OU
D
)
,
pp.
482
–
489,
2
010.
[3]
H.
Ta
n
and
L.
Chen,
"
An
appr
oac
h
for
fast
an
d
par
a
ll
e
l
vid
eo
proc
essing
on
A
p
ac
he
Hadoop
c
luste
rs
"
,
IE
EE
I
nt
.
conf
.
on
Mu
lt
ime
dia
and
Ex
po
(
ICME)
,
vol.
1,
no
.
6
,
2014
.
[4]
C.
L
iu,
et al
.
,
"
A
distri
but
ed
v
ideo share
s
y
stem b
ase
d
on
Hadoop
"
,
in
IEEE
3rd I
nt.
Con
f.
on
Clo
ud
Computing
a
nd
Inte
lligen
ce Sy
st
ems (
CCIS
)
,
pp.
587
-
590,
Novem
ber
,
2014
.
[5]
X.
Li
u,
et
al
.
"
A
Distribut
ed
Video
Mana
geme
nt
Cloud
Plat
for
m
U
sing
Hadoo
p
"
,
IE
EE
Acce
ss
,
vol.
3
,
no.
1,
p
p.
2637
-
2643,
201
5.
0
200
400
600
800
1000
1200
1
0
0
1
5
0
2
0
0
2
5
0
3
0
0
3
5
0
0
200
400
600
800
1000
1200
5
0
0
MB
1
GB
1
.
5
GB
2
GB
2
.
5
GB
1
0
0
1
5
0
2
0
0
2
5
0
3
0
0
3
5
0
0
500
1000
1500
3
0
0
5
0
0
7
0
0
8
0
0
1
0
0
0
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
& C
om
p
Eng
IS
S
N: 20
88
-
8708
Pos
t
Eve
nt Inv
est
iga
ti
on
of M
ulti
-
s
tre
am Vid
eo Da
t
a Uti
li
zi
ng Ha
doop Cl
us
te
r
(
Jyoti
P
arsola
)
5097
[6]
Yasee
n,
et
al
.
,
"
Cloud
-
base
d
sc
al
ab
le
obj
ec
t
d
e
te
c
ti
on
and
cl
as
sific
a
ti
on
in
v
id
eo
strea
m
s
"
,
Fu
t
ure
Gene
ration
Computer
Sys
tem
s
,
vol. 80, pp.
286
-
298,
2018
.
[7]
J.
Parsola,
e
t
al
.,
"
Eff
icien
t
Stor
age
and
Proc
essing
of
Video
Data
for
Moving
Objec
t
De
te
c
ti
o
n
using
Hadoop
MapReduc
e
"
,
in
Int.
Conf.
on
S
i
gnal,
Net
works
,
Computing
and
Syste
ms
(
ICNCS
-
2016)
,
JN
U,
N
ew
Delhi
,
Ind
ia
,
2016
.
[8]
J.
Cohen,
"
Graph
twiddl
ing
in
a
MapReduc
e
wor
ld
"
,
Computing
in
Sci
ence
&
En
gine
ering
,
vol.
1
1,
no.
4,
pp.
29
-
41,
2009
.
[9]
J.
M
y
ung,
S.G
Le
e
,
"
Expl
o
it
ing
int
er
-
op
erati
on
par
allelism
for
m
at
rix
ch
ai
n
m
ult
iplication
usin
g
MapReduce
"
,
J.
on
Super
Compu
ti
ng.
vol.
66
,
no
.
1,
pp
.
594
-
609
,
2013.
[10]
H.
W
ang,
et
a
l
.,
"
Eff
icient
quer
y
pro
ce
ss
ing
fr
a
m
ework
for
big
data
ware
hous
e
:
an
al
m
ost
joi
n
-
fre
e
appr
oa
ch
"
,
Fronti
ers of
Co
mputer
Scienc
e
,
vol.
9
,
no
.
12
,
pp
.
224
-
236
,
2015
.
[11]
J.
Ahn
et
al
.
"
SigM
R:
MapReduc
e
base
d
SP
ARQ
L
quer
y
pro
ce
s
sing
b
y
signa
tur
e
enc
oding
and
m
ult
iwa
y
joi
n
"
,
J.
on
Super
Compu
ti
ng
,
vol
.
7
1,
no.
10,
pp
.
3695
-
37
25,
2015
.
[12]
M.
Yam
amoto
a
nd
K.
Kan
eko,
"
Para
llel
image
d
at
ab
ase
pro
ce
ss
i
ng
with
m
apr
ed
uce
and
p
erf
orm
anc
e
ev
al
ua
ti
on
in
pseudo
distri
bu
t
ed
m
ode
"
,
Int. J.
on
E
lectronic C
omm
erc
e
Studies
,
vol
.
3
,
no
.
2
,
pp
.
211
-
228.
[13]
H.
D
Zhu,
et
a
l
.,
"
Para
ll
e
l
Im
age
Te
xtur
e
Feat
ur
e
Ext
ra
ct
ion
und
er
Hadoop
Cloud
Plat
form
"
,
Int
el
l
i
gent
Computing
Theory.
Springe
r Int
.
Publishing
.
459
-
465
.
[14]
W
.
Prem
cha
isw
adi
,
e
t
al
.,
"
Im
proving
per
f
or
m
anc
e
of
cont
e
nt
-
base
d
image
ret
rie
v
al
sche
m
es
using
Hadoop
MapReduc
e
"
,
I
E
EE
In
t. c
onf
.
on
High
Pe
rform
an
ce
Comput
ing
a
nd
Simulation (
HPCS)
,
pp.
615
-
620,
2013
.
[15]
T.
D.
Gam
ag
e,
e
t
al
.,
"
Im
age
filt
eri
ng
with
Map
Red
uce
in
pseu
do
-
distri
buti
on
m
ode
"
,
IEEE
co
nf.
on
Moratuwa
Engi
ne
ering
R
ese
arch
Conf
ere
nc
e
(
MER
Con)
.
16
0
-
164,
2015
.
[16]
T.
Li
u
,
et
a
l
.
,
"
SEIP:
S
y
stem
for
Eff
icient
Im
age
Pro
ce
ss
ing
on
D
istri
bute
d
Pl
at
for
m
"
,
Journal
of
Computer
Sci
ence
and
Technol
og
y,
vol
.
30,
no.
6,
p
p.
1215
-
1232
,
2
015.
[17]
G.
Bat
hl
a,
e
t
al
.,
"
A
Novel
Approac
h
for
c
lust
eri
ng
Big
Dat
a
base
d
on
MapReduc
e
"
,
Int
.
J
.
o
f
El
e
ct
rica
l
and
Computer
Engi
n
ee
ring (
IJE
C
E)
,
vol.
8
,
no
.
3
,
pp
.
1711
-
1719,
201
8.
[18]
A.
Anjum
,
e
t
al
.,
"
Vid
eo
stre
a
m
ana
l
y
sis
in
clouds
:
An
object
detec
t
ion
and
c
la
ss
ifi
c
at
ion
fra
m
ework
for
hig
h
pe
rform
anc
e
vid
eo
an
aly
t
ic
s
"
,
IE
EE
Tr
ansacti
ons
on
Cloud
Computing
,
2016.
[19]
A.
Azli
,
"
Distri
bute
d
visua
l
en
hanc
ement
on
surveil
l
anc
e
v
id
eo
with
Hadoo
p
Mapre
duce
a
nd
per
form
anc
e
eva
lu
at
ion
in
pse
udo
distri
bu
te
d
m
ode,
"
Australi
an
J. of B
asi
c
an
d
Applied
S
cienc
es,
vo
l.
8,
no.
9
,
p
p.
38,
2014
[20]
P.S.G
Aruna
Sri
and
M.
Anus
ha,
"
Big
dat
a
surve
y
"
,
Indone
sian
J
ournal
of
El
ectri
cal
Engi
n
ee
ring
and
Informatic
s
,
vol.
4
,
no
.
1
,
201
8.
[21]
S.
A.
Tha
n
eka
r
et
al
.
,
"
A
Stud
y
on
MapR
edu
ce
:
Ch
al
l
enge
s
and
Tre
nds
"
,
In
donesian
Journ
al
of
E
le
c
tric
al
Engi
ne
ering
and
Computer
Sc
ie
n
ce
,
vol
.
4
,
no
.
1,
2016.
[22]
D.
Gangodka
r
et
al
.
"
Robust
segm
ent
at
ion
of
m
oving
vehi
cles
under
complex
outdoor
con
dit
ions
"
,
IEEE
Tr
ansacti
ons on Intel
l
.
Tr
ansp.
S
ys.
,
vol.
13,
no.
4,
pp
.
1738
-
175
2,
2012
.
[23]
W
.
E.
L
Grim
son
and
C.
St
auf
fer
,
"
Adapti
v
e
ba
c
kground
m
ixt
ure
m
odel
s
for
re
al
-
ti
m
e
tracki
ng
"
,
In
IE
EE
Con
f
.
Comput
Vi
sion
a
nd
Pattern
R
ec
o
gnit
ion
,
vol
.
1,
p
p.
22
–
29
,
1999
.
[24]
Z.
Yu
and
Y.
Chen,
"
A
rea
l
-
t
ime
m
oti
on
det
ecti
on
al
gorit
hm
for
tra
ffic
m
onit
ori
ng
sy
st
ems
base
d
on
conse
cut
iv
e
te
m
pora
l
d
iffe
r
e
nce
"
,
In7th
Conf
of Asian
Contro
l
Conf
ere
nce
(
ACC)
,
pp
.
1594
-
1
599,
2009
.
[25]
D.Gangodka
r
e
t
al
.
"
Segm
entat
ion
of
m
oving
obje
c
t
s
in
visib
l
e
and
the
rm
al
vide
os
"
,
Int
.
Co
nf
of
Comput
er
Comm
unic
ati
on
and
Informatic
s
(
ICCCI
)
,
pp.
1
-
5
,
2012
.
[26]
W
hit
e. T
.
:
Hado
op:
Th
e
Def
ini
t
i
ve
Guide
.
Yahoo
Press
(2010)
[27]
Holm
es,
A.:
H
ad
oop
in
pr
actice.
Manning
Publi
c
at
ions Co
.
(2012)
[28]
Change
De
tecti
o
n
Benc
hm
ark
.
A
vai
l
abl
e
:
ht
tp:
/
/wordpre
ss
-
jodoi
n.
dm
i.
usherb.
c
a/
d
at
ase
t2014/
[29]
La
bora
tor
y
for
I
m
age
&
Med
ia
Understa
nding
(
LIMU).
Avai
la
b
le
:
htt
p
:/
/
li
m
u.
ait
.
k
y
ushu
-
u.
a
c.
jp
/
dat
ase
t/
en
/
[30]
Conte
xt
Aw
are
Vision
Us
ing
I
m
age
-
Based
Acti
ve
Rec
o
gnit
ion
(CAV
IAR).
Avai
la
bl
e
htt
p://hom
epa
ge
s.inf
.
ed
.
a
c.
uk
/rbf
/CAVIA
RDA
T
A1/
[31]
Open
Source
Co
m
pute
r
Vision
(
OpenCV) [Onli
n
e]
.
Ava
ilable:
ht
tp:
//
op
enc
v
.
will
o
wgara
ge
.
com/wi
ki/
Evaluation Warning : The document was created with Spire.PDF for Python.