TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 14, No. 1, April 2015, pp. 123 ~ 1
2
9
DOI: 10.115
9
1
/telkomni
ka.
v
14i1.725
7
123
Re
cei
v
ed
De
cem
ber 2
9
, 2014; Re
vi
sed
F
ebruary 20,
2015; Accept
ed March 1
5
, 2015
Evaluation of Moving Object Detection Methods based
on General Purpose Single Board Computer
Agung
Nugr
oho Jati*, Le
d
y
a No
v
a
mizanti, Mirsa Ba
y
u
Praset
y
o
, And
y
Ruhend
y
Putra
Dep
a
rtment of Comp
uter Engi
neer
ing, Sch
o
o
l
of
Electrical E
ngi
neer
in
g, T
e
l
k
om Univ
ersit
y
,
Jl.
T
e
lekomu
ni
kasi T
e
rusan Buah Batu, Ba
n
dun
g, Indon
esi
a
402
57, Ph. +
622
2-75
64
10
8
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: agun
gn
j@tel
k
omun
iversit
y
.
a
c.id
A
b
st
r
a
ct
RGA and SKD
A are tw
o different methods w
h
ich ca
n be us
ed to detect th
e obj
ect in i
m
a
ge bas
e
d
process
i
ng. In
order to su
p
port the
movi
ng surv
eil
l
anc
e ca
mera sys
tem w
h
ich
pr
opos
ed i
n
T
e
l
k
o
m
Univers
i
ty, RGA and SKDA h
a
ve tested to b
e
review
ed
w
h
i
c
h mor
e
reli
abl
e to be impl
emented i
n
a sing
l
e
boar
d co
mpute
r
. In this
pap
er, w
ill b
e
discuss
ed
abo
ut i
m
ple
m
e
n
tatio
n
a
n
d
testing
of tw
o d
i
fferent
meth
od
s
of obj
ect detect
i
on us
in
g back
g
rou
nds
su
btra
ction. For i
m
pl
ementati
on,
e
a
c
h of the
m
w
ill
be co
mbin
ed w
i
th
Extende
d K
a
l
m
an F
ilt
er i
n
a
R
a
spb
e
rry Pi. T
h
e p
a
ra
meter
w
h
ich
hav
e teste
d
ar
e
me
mory
and
CPU
usa
g
e
,
and syste
m
uti
l
i
z
a
t
io
n. The r
e
sult sh
ow
s that RGA is
mor
e
rel
i
ab
le th
an
SKDA to i
m
pl
emente
d
i
n
S
B
C
because of les
s
CPU usag
e and system
utiliz
a
tion.
Ke
y
w
ords
: ru
nni
ng
gauss
i
a
n
avera
ge (
R
GA), sequ
entia
l
ke
rne
l
d
ensity appr
oxi
m
ati
on (SKDA),
exten
d
e
d
kal
m
a
n
filter, singl
e bo
ard co
mp
uter (SBC), Rasp
berry Pi
Copy
right
©
2015 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
The stu
d
y of obje
c
t detecti
on ba
sed
on i
m
age p
r
o
c
e
s
sing i
s
the int
e
re
sting
way until this
time, especi
a
lly in the im
p
l
ementation.
Many
meth
o
d
s have mod
e
led and
dev
elope
d
b
e
fore,
tried to be im
plemente
d
in
many syste
m
. But
there are ma
ny co
nstrai
nts in e
m
bedd
ed sy
stem
impleme
n
tation. In an em
bedd
ed sy
ste
m
su
ch a
s
si
ngle bo
ard compute
r
(SB
C
), the
r
e’s l
a
ck
processor
with a limited m
e
mory capaci
ty. Besides
, limited power also will be
a constraint in
impleme
n
tation.
Ra
spb
e
rry Pi as a SB
C has
a sma
ll size an
d
comp
act fo
r embed
ded
system
impleme
n
tation [1]. As a comp
en
satio
n
, it can’t be
upgrade
d b
y
bigger m
e
mory capa
city or
highe
r processor. So,
whe
n
som
eon
e try to im
pleme
n
t a com
putin
g mechani
sm
onto SBC, they
must
con
s
id
e
r
r
e
s
our
ce
s li
mit
a
t
i
on.
A
s
we
kno
w
, the
impleme
n
tation of
sign
al p
r
ocessin
g
n
e
eds
much
en
oug
h re
so
urce
s i
n
a
comp
uter. So, in
this
pape
r
will be
evaluated
a
n
implem
enta
t
ion
usin
g SBC. What can be
happe
ned
whe
n
we mo
ve the comp
licated te
chni
que
s to limited
resou
r
ced co
mputer.
Last ye
ar, p
r
opo
sed
a m
o
ving
surveill
ance
came
ra
in Tel
k
o
m
University for high
er
reliability in
order to monitor
indoor building [15]. Th
us system
consi
s
ts
of camera
and SBC
whi
c
h
ca
n m
o
ve follo
w th
e hu
man
mo
vement. Con
s
ide
r
th
e limi
t
ation of SB
C, mu
st b
e
f
ound
best
method
with
high
a
c
curay
but
still manag
ed
by a
sm
all si
ze
memo
ry. Beside
s, t
he
processi
ng time still can be tolerated.
There are two kin
d
of mai
n
pro
c
e
s
s in
the sy
stem.
The first is o
b
ject d
e
tectio
n. Many
method
s a
r
e
develop
ed i
n
this
way.
Most of
th
e
m
use b
a
ckg
r
oun
d
sub
s
traction
techni
que
s,
su
ch a
s
Ru
n
n
ing Ga
ussia
n
Average,
Tempo
r
al
M
edian Filte
r
, Mixture of Gaussia
n
, Kernel
Den
s
ity Estimation, Sequ
ential Kern
el
Den
s
ity Ap
p
r
oximation, a
nd many mo
re [2-4]. in this
pape
r, will be
only discussed abo
ut RG
A and SKDA.
We o
n
ly tried
to impleme
n
t RGA a
nd SK
DA and
evalu
a
te them b
e
cause RGA is
said
as
a simple alg
o
r
ithm with an
acce
ptable a
c
cu
ra
cy and
only need
s le
ss me
mory consumption [
5
].
SKDA, is more complex al
gorithm
but has hi
gh accu
racy but
still needs lower time complexit
y
to
be p
r
o
c
e
s
se
d [6]. In the
end, you
ca
n se
e the
co
nclu
sio
n
, wh
ere i
s
fitter t
han oth
e
r to
be
impleme
n
ted in
SBC.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 14, No. 1, April 2015 : 123 – 12
9
124
The
se
con
d
pro
c
e
s
s is m
o
vement p
r
e
d
iction. Exten
ded Kal
m
an
Filter
(EKF) i
s
u
s
e
d
a
s
an algo
rithm for pre
d
ictin
g
the movement
[7].
The output from SKDA or RGA, proce
s
sed in EKF
to produ
ce th
e dire
ction. And then, SBC will comm
an
d the came
ra
to move dire
ctly to
the obje
c
t.
The p
r
esenta
t
ion is organi
zed a
s
follo
w: in t
he se
co
nd se
ction,
will be de
scrib
ed the metho
d
whi
c
h is im
plemented i
n
SBC. In the third
section, we’ll discuss
about the resul
t
of
impleme
n
tation, given the analysi
s
an
d evaluat
ion. And the last, c
oncl
u
si
on
s are given.
2. Rese
arch
Metho
d
2.1. Running
Gaussian
Av
erage (RGA)
Run
n
ing
Gau
ssi
an Averag
e (RGA) i
s
o
ne of
ma
ny method
s
in b
a
cg
rou
nd su
bstra
c
tion
[8]. This met
hod
can b
e
u
s
ed to d
e
tect
the obje
c
t, static or
dynam
ic. RGA i
s
al
so calle
d a
s
1-G
(One
Gau
s
sian) [2]. In RG
A, backgro
un
d pixel model
ed as G
a
u
ssi
an Di
stributio
n (µ,
∑
), wh
ere:
μ
s
,t
: mean point of each pixel
and time;
∑
s
,t
: covarian
ce
matrix of each pixel and time [5].
Covari
an
ce
matrix value
s
depen
d on
noise whi
c
h i
s
contain
ed
by the image
. Highe
r
noise
will produces higher covari
an
ce m
a
trix too. In the other
wo
rds, higher noi
se in the im
age
will also produce hi
ghe
r temporal gradi
ent
I
s,
t
−μ
s,
t
, so the p
i
xel can b
e
said as m
o
vin
g
pixel.
The mea
n
an
d covari
an
ce
are dyna
mic
values, an
d can be up
date
d
by the formulas:
μ
s
,t
+
1
=
(
1
−α
)
.
μ
s
,t
+
α
.I
s
,t
(
1
)
∑
s,
t
+
1
=
(
1
−α
)
.
∑
s,
t
+
α
.
(
I
s,
t
−μ
s,
t
)(
I
s,
t
−μ
s,
t
)
T
(
2
)
In SBC impl
ementation,
covari
an
ce
matrix
can b
e
optimi
z
ed
by cal
c
ulate
only the
diago
nal of t
he mat
r
ix. Beca
use in
ge
neral,
cova
ria
n
ce
matrix
consi
s
t of 3x3
matrix. Besi
des,
can be u
s
e
d
spatial agre
gation as li
ke
as morp
hol
o
g
ical filter to make b
e
tter perfo
rman
ce
of
backg
rou
nd substractio
n
techni
que [2].
2.2. Sequential Kernel De
nsit
y
Approximation
In Seque
ntial
Kern
el Den
s
i
t
y Approxima
t
ion (SKDA
)
method, m
e
a
n
-shift vecto
r
is u
s
e
d
to track the gradient
whi
c
h can
detect
main m
odes
from pdf (probability
densit
y
function) ti
me
from data sa
mple by minimal set assu
mption direct
l
y
[8]. SKDA u
s
e
s
Gau
s
sian
Kernel, note
d
by
x
i
(i = 1,…,n), where:
x
i
: Gaussian m
ean value;
π
: variance of
covari
an
ce d
x
d related wit
h
Gau
ssi
an [4].
The finction from den
sity point define a
s
:
/
∑
|
|
/
ex
p
,
,
(3)
Whe
r
e,
,
,
≡
(
4
)
SKDA is opti
m
ized
from
K
e
rnel
Den
s
ity Estimation
(KDE).
Whe
r
e
sa
mple f
r
am
e taken
pro
c
e
ss i
s
co
mpre
ssed
so
time pro
c
e
s
s
in more effici
ent. This i
s
h
appe
ned
because le
ss fra
m
e
is se
nt to be pro
c
e
s
sed, so the comp
utation is lee
s
complexity [8].
2.3. Extende
d Kalman Filter
Extended Kal
m
an Filter
(E
KF) is a
set
of f
unction
s
whi
c
h a
r
e u
s
ed to e
s
timate cu
rrent
system state and co
rre
ct
t
hem
by
a
s
e
d
on p
r
eviou
s
state. EKF uses t
w
o
cal
c
ul
ation pa
ram
e
ter,
they are pri
o
r state and po
sterio
r st
ate [
7
, 9]. You can see a
s
follo
ws:
Prior st
ate cal
c
ulatio
n given
as:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Evaluatio
n of Movin
g
Obje
ct Detection M
e
thod
s ba
sed
on Gene
ral
…
(Agun
g Nu
groh
o Jati
)
125
,
,
(
5
)
Posteri
o
r State Cal
c
ulation
given as:
(
6
)
Whe
r
e K is K
a
lman gai
n gi
ven belo
w
:
Σ
Σ
(
7
)
And h(X) i
s
a cal
c
ulatio
n function a
s
follo
ws:
̅
(
8
)
∑
: covaria
n
ce
erro
r, whe
r
e
can b
e
cal
c
ul
ate by:
Σ
Σ
R
(
9
)
Σ
K
H
Σ
(
1
0
)
H
&
F
: calcul
ation matrix, cal
c
ulate by Jaco
bian b
e
lo
w:
∂
h
(
x
t
,x
t
−
1
)
∂
x
t
−
1
for
H
(
1
1
)
∂
f
(
x
t
,x
t
−
1
)
∂
́
x
t
−
1
for
F
(12
)
Q: calculation
erro
r.
From tho
s
e f
o
rmul
as,
I
is
defined a
s
id
entity matrix and R
i
s
defin
ed as
cal
c
ul
ation erro
r.
Dynami
c
syst
em can be ta
ken by repl
acing value of
z
t
be
x
t
−
1
. Those techniqu
e optimize
s
the
dire
ction of camera dete
c
tion wh
en there are mo
re th
an one o
b
je
ct exist.
2.4. Object
Mov
e
ment Model
Obje
ct move
ment i
s
mo
deled
by a
linear move
ment. Gotte
n by
cal
c
ula
t
ing the
differen
c
e of curre
n
t frame
coordinate a
nd prev
io
us o
ne. So, the model given a
s
follows:
∆
,
;
(
1
3
)
∆
,
;
(
1
4
)
Whe
r
e
x
, is a hori
z
ontal axi
s
, y is ve
rtical
axis, and t is the time.
2.5. Raspb
e
r
r
y
Pi Model B
As
a SB
C, we us
e
Ras
p
berry Pi
model B.
Ra
spb
e
rry Pi use
s
So
C fro
m
Broa
dcom
m
BCM28
35
an
d con
s
ist
of
ARM11
70
0
MHz p
r
o
c
e
s
sor, 5
12 MB
RAM, an
d
G
P
U Vide
ocore IV.
This SBC u
s
e
s
SDCard for
booting
seq
u
ence and
storage sy
stem [10, 11].
2.6. Ev
aluation Crite
r
ia
There a
r
e
two ki
nd
of eval
uati
on
criteria
s. Fi
rst
criteri
a
is related
with ea
ch
ba
ckgrou
n
d
sub
s
tra
c
tion
method
s. Fo
r RGA, the p
a
ram
e
ters
a
r
e learning
ra
te (a)
and
d
e
viation sta
n
dart
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 14, No. 1, April 2015 : 123 – 12
9
126
con
s
tant (k) [5, 8]. And in
SKDA, they
are data vari
ances (
) and
t
(
threshold
) [2, 12]. Sec
o
nd
crite
r
ia, whi
c
h are
wante
d
to kno
w
to see p
e
rf
o
r
ma
nce
s
of impl
ementation i
n
SBC. They are
inclu
d
ing me
mory and
CPU usage,
an
d
system utilization.
Figure 1. Ra
spberry Pi model B [8]
3. Results a
nd Analy
s
is
3.1. Sy
stem
Design
System hard
w
are co
nsi
s
t
of came
ra, SBC Ra
spb
e
rry Pi, and servo motor for
dire
cting
came
ra to th
e obje
c
t. First, system has to detec
t the
object by RGA or SKDA. If there are
any
movement of
the obj
ect,
system
will f
o
llow that
m
o
vement.
In order
to sm
ooth
the system
movement, E
x
tended Kalm
an Filter is i
m
plemente
d
. In
the sim
p
le
way, system
d
e
scrib
ed
by this
picture belo
w
.
Ob
j
e
c
t
De
t
e
c
t
io
n
(u
s
i
n
g
RG
A
or
SK
D
A
)
Mo
v
e
m
e
n
t
Pr
e
d
i
c
t
i
o
n
(us
i
ng
Ex
t
e
n
d
e
d
K
a
l
man
F
ilt
e
r
)
Ca
m
e
r
a
Mo
v
e
m
e
n
t
Up
d
a
t
e
Im
ag
e
Ca
p
t
u
r
e
d
vi
d
e
o
fr
o
m
ca
m
e
r
a
Mo
t
o
r
co
n
t
r
o
l
Figure 2. System Impleme
n
tation Sche
me
3.2. RGA a
n
d SKD
A
Par
a
m
e
ter
s
An
a
l
y
s
is
As explained
before, SKDA has two import
ant pa
ra
meters. They are varian
ce
(
)
a
nd
threshold
(t).
We’ve alre
a
d
y tested im
plementat
io
n
of SKDA as a backg
roun
d sub
s
tractio
n
in
some of
scen
ario
s. The re
sults are d
e
scribed bel
ow.
Figure 2. SKDA Pre
c
isio
n Values
Dep
e
n
ds o
n
Threshold an
d Vari
ance
0
0.
1
0.
2
0.
3
0.
4
0.
5
0.
6
recall
/
p
recissio
n
Thre
shold
sigm
a
2
sigm
a
4
sigm
a
6
sigm
a
8
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Evaluatio
n of Movin
g
Obje
ct Detection M
e
thod
s ba
sed
on Gene
ral
…
(Agun
g Nu
groh
o Jati
)
127
Based
on
th
e test,
optim
al value
of t
he p
r
e
c
isi
o
n
of the
syste
m
are h
app
e
ned
wh
en
varian
ce
2
and th
re
shol
d 1e
-5. So,
that value
will be
used
for the im
p
l
ementation
and
perfo
rman
ce testing.
RGA also uses 2 paramet
ers, they are l
earni
ng rate
and thre
sh
old
.
We’ve try to
chang
e
the values of
para
m
eters in
orde
r to find optimal
syste
m
. The re
sult is de
scribe
d as follo
ws.
Figure 3. RG
A Preci
s
ion V
a
lue
s
De
pen
ds on L
e
arnin
g
Rate an
d Deviation Standart
Optimal
pre
c
i
s
ion
is a
c
hiev
ed
whe
n
a
se
t as
0.6 a
n
d
t 1. Based
on
the g
r
a
phi
c,
highe
r
threshold
will make preci
s
i
on less and becom
e
un
-optimal. This is
also happened for the higher
a value.
3.3. EKF Parameter T
esti
ng
EKF is used t
o
pre
d
ict the
next moveme
nt
of object.
EKF perform
ance dep
e
nd
s on the
Q (covaria
nce) pa
ram
e
ter. In this rese
arch, we’ve t
r
ied
some va
rian
ce
s of Q
value and t
he
result is de
scribed in the fi
gure b
e
lo
w.
(a)
(b)
Figure 4. (a)
amount of tru
e
predi
ction; (b) “spe
ed” of
predi
ction
From the graphi
cs ab
ove
,
shown that
higher
cova
rian
ce will i
m
pact the “speed
” of
predi
ction to
become le
ss.
For impleme
n
tation, Q
is set as 0.1 ca
use it has hi
g
hest “spe
ed
” and
still tolerated
“true” predi
ction. Q=0
can’t
be us
ed because will give many “jump” processes.
3.4. SBC Per
f
omanc
e
An
aly
s
is
SBC based system perfo
rmance can b
e
see
n
by
an
alyzing the m
e
mory an
d CPU usage
[13, 14], al
so the
system
utilizat
ion. T
h
is test is proposed to
know
whi
c
h technique is more
comp
atible to
be impleme
n
t
ed in SBC.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 14, No. 1, April 2015 : 123 – 12
9
128
Figure 5. RG
A and SKDA Comp
utation
Based
on th
e figure
abov
e, sho
w
n th
a
t
SKDA use
arroun
d 10%
CPU
re
sou
r
ce mo
re
than RGA. RGA uses CPU
arro
und
73.45%
av
erag
e
wh
en
is exe
c
ute
d
. And SK
DA, use
averag
e 8
8
.2
4% du
ring
execute
d
by th
e sy
stem.
De
spite
RGA
ha
s le
ss u
s
a
ge
of CPU, it
uses
large a
m
ou
nt of memory ca
pacity duri
ng
the pr
o
c
e
ss. I
t
’s described i
n
the figure b
e
low.
Figure 6. RG
A and SKDA Memory Usa
g
e
RGA uses m
o
re mem
o
ry reso
urce be
ca
use
it need
s
more me
mory space for d
e
viation
stand
art of e
a
ch
pixel (u
sed 2 fram
es
of ev
ery pro
c
ess). While
SKDA only u
s
e
s
mem
o
ry
for
putting 2 fram
es in every p
r
oce
s
s.
(a)
(b)
Figure 7. System Utilizat
ion (a)
RGA; (b) SKDA
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Evaluatio
n of Movin
g
Obje
ct Detection M
e
thod
s ba
sed
on Gene
ral
…
(Agun
g Nu
groh
o Jati
)
129
From
Figu
re
7 ab
ove,
we
can
say that
RGA i
s
more
“healthy
”
tha
n
SKDA. Th
e
system i
s
alway
s
in hig
h
utilization
when sy
stem i
m
plem
e
n
t SKDA as b
a
ckg
r
ound
sub
s
tra
c
tion techniq
ue.
SKDA has le
ss u
s
e
r
time but highe
r idl
e
time of the
pro
c
e
s
sor tha
n
RGA. It’s happe
ned be
cause
by using SKDA as a ba
ckg
ound
sub
s
tra
c
tion techniq
ue, system m
o
re often
sen
d
requ
est dat
a to
the hard
w
a
r
e
level.
4. Conclusio
n
Based
o
n
the
implem
entati
on, testin
g, a
nd a
nalysi
s
can
con
s
tructe
d some
con
c
l
u
sio
n
s
of the
study.
As a
limited
reso
ur
ce
s co
mputer, sin
g
l
e
bo
ard com
put
er can’t p
r
oce
s
s a
hig
h
level
sign
al processing
as
realti
me. The
r
e h
a
ve bee
n tested two
kind
of backg
rou
nd sub
s
tra
c
tio
n
techni
que
co
mbined
by EKF in a SBC. Both RGA and SKDA can be u
s
ed
as a b
a
ckg
r
o
und
sub
s
tratio
n te
chni
que to d
e
t
ect the movi
ng obje
c
t
in a
SBC. But it’s requi
red m
o
re than 25
6 MB
memory
cap
a
c
ity. RGA ind
eed, nee
ds
m
o
re m
e
mo
ry usa
ge (arrou
nd 0.1%) tha
n
SKDA to st
ore
deviation of e
a
ch pixel for
every frame. In the im
plem
entation, use
d
2 frame
s
of captu
r
ed vide
o.
Ho
wever, RG
A uses le
ss
reso
urce of CPU in the pro
c
e
ss than SK
DA. RGA just
needs
in average
7
3
.45% of CP
U resource
while SKD
A n
eed
s in
average 8
8
.24%.
Becau
s
e
of that,
SKDA needs more time processing
. A
nd it
also infl
uence the
sy
stem
utilizati
on. Based on the
study, known
that SKDA will make
sy
stem always in high utilization.
All of the results of this st
udy sho
w
n t
hat
RGA is
more reliabl
e
to be implemented in
SBC, esp
e
cia
lly Rasp
be
rry
Pi. Still and all, the pro
c
e
ssi
ng time n
e
eds to b
e
in
crea
sed
more. It
can b
e
hap
pe
ned by repl
ace the SBC with better CP
U spe
ed an
d hi
gher m
e
mo
ry capa
city.
In the further stage of stu
d
y, the result
from this st
udy can
be
expand
ed to try other
backg
rou
nd substractio
n
m
e
thod
s. And also in
thext rese
arche
s
, can be implem
ented directly
as
a surveil
a
n
c
e
came
ra to follow the hum
a
n
movement.
Referen
ces
[1]
T
he Rapsberr
y
Pi F
oundati
on:
http://
w
w
w
.
r
a
s
pberr
y
p
i
.org/. October 30th, 201
3.
[2]
Benez
eth Y, J
odo
in PM, Em
ile B,
Laur
ent
H,
Rose
nb
erg
e
r C. Com
par
ative Stu
d
y of
Backgro
un
d
Subtraction Algorithm.
Journ
a
l
of Electronic Ima
g
i
n
g
. 20
10: 1-30.
[3]
T
s
ai DM, Lai
SC. Indepen
dent Com
pon
ent Anal
ys
is-B
ased Back
gro
und Su
btractio
n for Indoo
r
Surveillance.
IEEE Transaction on
I
m
age P
r
ocessing.
20
0
9
; 18(1): 15
8-1
67.
[4]
Madd
ale
na
L, Petrosin
o A.
A Self-Orga
n
izin
g Ap
pro
a
c
h to Back
groun
d Su
btraction for V
i
su
al
Surveil
l
a
n
ce A
pplic
atio
ns.
IEEE Transactio
n
on Image Pr
ocessi
ng.
20
08
; 17(7): 116
8-1
177.
[5]
C W
r
en, A Az
arba
ye
ja
ni, T
Darel
l
, AP Pe
ntlan
d
. Pfind
e
r
:
Real T
i
me T
r
ackin
g
of the
Huma
n Bo
d
y
.
IEEE Trans. on Pattern Analys
is and Mac
h
in
e
Intellig
enc
e
. 1997; 19(
7): 780
-785.
[6] B
Han
DC.
S
equ
enti
a
l Ke
mel D
ensity A
p
proxi
m
ati
on t
h
roug
h Mo
de P
r
opa
gatio
n: Ap
plicati
ons t
o
Backgro
un
d Mode
lin
g
. Proce
edi
ngs of Asia
n Confer
enc
e
.
200
4.
[7]
Hua
ng S. Und
e
rstand
ing E
x
t
end
ed Ka
lman
F
ilt
er - Part III:
Exte
nde
d Kalm
an F
ilter. 201
0.
[8] Piccardi
M.
B
a
ckgro
und
Su
btractionT
ec
hn
iqu
e
s: a
Revi
ew
. IEEE Internationa
l Conferenc
e o
n
S
y
stems, Man
and C
y
b
e
rn
atic
s. 2004: 30
99-
310
4.
[9]
Ribe
i
ro
n M I. Kalma
n
and E
x
t
end
ed Ka
lman
F
ilter: Conce
p
t, Derivati
on, an
d Properti
es. 2004.
[10]
T
he Rapsberr
y
Pi F
oun
dati
on.
F
A
Qs |
Raps
berry Pi
. October 30th, 2013. Ra
p
s
berr
y
pi.
o
rg
:
http://
w
w
w
.
ras
pberr
y
p
i
.org/fa
qs.
[11]
T
he Raspberr
y
Pi F
o
u
n
d
a
ti
on.
Ab
ou
t u
s
| R
a
sp
be
rry Pi
. October 30t
h, 201
3. Rasp
berr
y
pi
org:
http://
w
w
w
.
ras
pberr
y
p
i
.org/a
b
out.
[12] Urbac
h ER, W
i
lkinso
n MH. Ef
ficient 2-
D Gra
y
sca
le
M
o
rp
hol
ogic
a
l T
r
ansfor
m
ations W
i
th
Arbitrar
y F
l
at
Structuring El
e
m
ents.
IEEE Tr
ansaction on Im
age Proc
essing.
200
8; 17(1)
: 1-7.
[13]
IBM. Linu
x Per
f
ormance a
nd
T
uning Guid
eli
nes
.
Ne
w
York:
IBM Group. 2004.
[14]
StrongMa
il S
y
s
t
em. Extreme L
i
nu
x Perform
a
n
c
e T
uning an
d Monitori
ng
.
StrongM
ail S
y
ste
m
.
[15]
And
y
R
u
h
e
n
d
y
, Agun
g NJ,
Le
d
y
a N.
S
i
ste
m
Movin
g
Surv
eil
l
ance
Ca
mera
Berbas
is Se
qu
entia
l Ker
n
e
l
Density Ap
prox
imatio
n da
n Extende
d Kal
m
an
F
ilter
. Seminar
Nasio
nal T
e
kn
olo
g
i T
e
rapan.
201
4: 304-
307.
Evaluation Warning : The document was created with Spire.PDF for Python.