Indonesian Journal of
Electrical
Engineer
ing and
Computer Science
V
o
l. 10
, No
. 3, Jun
e
20
18
, pp
. 98
9
~
99
9
ISSN: 2502-4752,
DOI: 10.
11591/ij
eecs.v10
.i3.pp989-999
9
89
Jo
urn
a
l
h
o
me
pa
ge
: http://iaescore.c
om/jo
urnals/index.php/ijeecs
A New Method for Ball
Tracking Based on
α
-
β
, Linear Kalm
an
and Extended Kalman Filters
Via Bubble Sort Algorithm
Ha
thi
r
am
Ne
na
va
th,
R
a
vi
Kum
a
r
Ja
to
th
Department o
f
Electronics
and C
o
mm
unication Engineer
ing, Na
ti
onal Inst
itut
e
of
Tec
hno
log
y
, Warangal, Ind
i
a
Article Info
A
B
STRAC
T
Article histo
r
y:
Received
Ja
n 19, 2018
Rev
i
sed
Mar
27
, 20
18
Accepted Apr 11, 2018
Object
tracking
is one of th
e ch
allenging
issues in computer vision
and vid
e
o
processing, whic
h has several pot
enti
al a
ppl
ica
tio
ns. In this paper,
initia
ll
y,
a
moving object is selected b
y
fr
ame
differen
c
in
g method and extracted the
object b
y
segment threshold
i
ng.
The
bubble sort algorithm (BSA) arranges
the regions (l
arg
e
to sm
all) to m
a
ke sure
th
at th
ere is at least one big region
(object) in
object detection pro
c
ess.
To track
th
e object, a motion model is
construct
e
d to set the s
y
stem
models of Alph
a-Beta (
α
-
β
) filter, Linear
Kalm
an filte
r (L
KF) and Extend
ed Kalm
an filt
er
(EKF). Man
y
e
xperim
e
nts
have been cond
ucted on balls
with di
fferen
t
s
i
zes
in im
age s
e
quences
and
compared th
eir
tracking p
e
rf
ormance in
no
rmal light and
bad light
conditions
. The
parameters obtained are
th
e root
m
ean s
quare err
o
r (RM
S
E),
absolute er
ror (AE), obje
c
t tr
ac
king e
rror (OTE), Tr
acking d
e
tection rate
(TDR), and peak signal-to-noise ratio (PSNR) an
d they
ar
e compared to fin
d
the
algorithm th
at p
e
rforms the b
e
st for
two cond
itions.
K
eyw
ords
:
Back
gr
oun
d sub
t
r
action
Bu
bb
le so
rt al
go
rith
m
EKF
LKF
Object trac
king e
r
ror
α
-
β
filter
Copyright ©
201
8 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
:
Hath
iram
Nen
a
v
a
th
,
Depa
rt
m
e
nt
of
El
ect
roni
cs
an
d C
o
m
m
uni
cati
on
En
gi
nee
r
i
n
g,
Natio
n
a
l
In
stitu
te of Techn
o
l
o
g
y
Warang
al,
W
a
r
a
ng
al, Ind
i
a 50
600
4.
Em
a
il: h
a
th
iram
.
iisc@g
m
ai
l.co
m
1.
INTRODUCTION
C
o
m
put
er
vi
si
on
i
s
t
h
e
a
dva
ncem
ent
i
n
sci
e
nce a
n
d t
ech
n
o
l
o
gy
t
h
at
i
n
cl
ude
m
e
t
hods
f
o
r
acq
ui
ri
ng
,
pr
ocessi
ng
, an
al
y
z
i
ng, an
d u
nde
rst
a
n
d
i
n
g i
m
ages. As a sci
e
nt
i
f
i
c
di
sci
p
l
i
n
e, com
put
er
vi
si
on i
s
fret
f
ul
wi
t
h
th
e th
eory b
e
hin
d
artificial syste
m
s th
at ex
cerp
t
inform
at
ion
fr
om
im
ages. The i
m
age dat
a
can t
a
ke
vari
ous
fo
rm
s, such as vi
deo se
qu
e
n
ces, vi
e
w
s fr
om
m
a
ny
cameras, o
r
m
u
l
t
i
-
di
m
e
nsi
onal
d
a
t
a
from
a
m
e
di
cal
scann
e
r.
As a
scien
tific d
i
scip
lin
e, co
m
p
u
t
er v
i
sion
pursu
es to
relate its t
h
eories and
mo
d
e
ls to
t
h
e d
e
sig
n
of
co
m
p
u
t
er v
i
si
o
n
syste
m
s. Man
y
research
ex
ten
d
s to
in
clu
d
e
:
u
s
es
in
security and
su
rv
eillan
ce, v
i
d
e
o
com
m
uni
cat
i
on an
d c
o
m
p
ressi
on
, Na
vi
g
a
t
i
on,
di
spl
a
y
t
echn
o
l
o
gy
,
Hi
g
h
-Le
v
el
V
i
deo
Anal
y
s
i
s
,
t
r
affi
c
co
n
t
ro
l, Metro
l
og
y, v
i
d
e
o ed
itin
g
,
augmen
ted
reality an
d
Hu
m
a
n
-
C
o
m
p
u
t
er in
terfaces to
med
i
cal
im
agi
ng [
1
-
2
]
.
Gi
ven, t
h
e m
o
st
im
port
a
nt
st
at
e (e.g.,
po
si
t
i
on an
d
vel
o
ci
t
y
) of a t
a
rget
o
b
ject
i
n
t
h
e l
e
a
d
i
n
g
im
age, the objective of tracki
ng is to
estim
a
te the states
of the target
i
n
the su
bse
que
nt
fram
e
s. Even t
h
o
u
g
h
object tracki
n
g has bee
n
studied for m
o
re than a fe
w
dec
a
des an
d si
g
n
i
f
i
cant
ev
ol
ut
i
o
n has
been m
a
de i
n
recent yea
r
s [3-7], it
rem
a
in
s a challenging
problem
.
Ob
j
ect illu
st
ratio
n
is
on
e
of th
e
k
e
y wo
rks in
an
y
p
i
ctorial track
ing
algo
rith
m
,
an
d man
y
ou
tlin
es
have
been
pr
o
pos
ed [
8
]
.
Si
nc
e t
h
e pri
m
ary
wo
rk
of L
u
cas
and
Kana
de [
9
]
,
based o
n
t
h
e
raw co
ncent
r
a
t
i
on o
f
val
u
es
ha
ve
be
en
br
oadl
y
use
d
f
o
r t
r
ac
ki
n
g
[1
0]
. T
h
e L
u
c
a
s an
d
Kana
de
ap
pr
oac
h
[
1
1]
, d
o
n
o
t
t
a
ke
great
occurre
nce c
h
angea
b
ility into account and thus, do
not
succee
d due t
o
the
dissim
i
l
a
rity of the c
h
rom
a
tic
p
r
op
erties
o
f
a targ
et
o
b
j
ect.
Matth
ews et al. [12
]
estab
lish
e
d a tem
p
late
up
d
a
te m
e
th
od
b
y
exp
l
o
iting
th
e
inform
ation of the first fram
e
to accurate points.
Towa
rds
better acc
ount
for c
h
anges i
n
the
presence
of t
h
e
ob
ject
, s
u
bs
pa
ce-base
d t
r
a
c
k
i
ng a
p
pr
oach
[7
,1
3]
ha
ve
b
een
pr
op
ose
d
.
In
[
14]
,
Ha
g
e
r an
d B
e
l
h
u
m
eur
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
502
-47
52
I
ndo
n
e
sian
J Elec Eng
& Com
p
Sci, V
o
l. 10
,
No
.
3
,
Jun
e
2
018
:
98
9 – 99
9
99
0
recom
m
ended
a pr
o
f
i
c
i
e
nt
L
u
cas a
n
d
Ka
n
a
de al
g
o
r
i
t
h
m
an
d
used
l
o
w
di
m
e
nsi
onal
desi
g
n
s
f
o
r t
r
acki
n
g
ch
ang
i
ng
illu
min
a
tio
n
co
nditio
n
s
. Sev
e
ral
tech
n
i
qu
es h
a
v
e
b
e
en
an
ticip
ated
fo
r
v
i
d
e
o
ob
j
ect seg
m
en
tatio
n.
C
o
m
m
onl
y
,
t
h
ese m
e
t
hods c
a
n
be a
p
p
r
o
x
i
m
at
ely
cl
assi
fi
ed i
n
t
o
t
w
o t
y
pes
[1
5,
1
6
]
:
back
g
r
o
u
n
d
st
r
u
ct
u
r
e
base
d vi
de
o ob
ject
se
gm
ent
a
t
i
on
an
d f
o
re
gr
o
u
n
d
e
x
t
r
act
i
o
n base
d vi
de
o ob
ject
se
gm
ent
a
t
i
on
.
In
recent tim
es
, a num
ber of t
r
acki
ng m
e
thods bas
e
d
on s
p
arse re
prese
n
ta
tions ha
ve
bee
n
propose
d
[1
7-
2
1
]
.
To ex
pan
d
t
r
ac
ki
n
g
ro
b
u
st
ness
, a l
o
cal
spars
e
pre
s
ence m
odel
was pr
op
ose
d
i
n
[22]
usi
ng t
h
e
m
ean
sh
ift algo
rith
m
to
d
e
tect obj
ects. By th
e p
r
esu
m
p
t
u
o
u
s rep
r
esen
tatio
n
o
f
particles as to
g
e
th
er,
H. Lu
et al [17
]
ex
pressed
o
b
j
e
ct track
ing
as
a
m
u
lti-task
sparse learn
i
ng
prob
lem
.
J. Fiscu
s
, J. Garo
fo
lo, T. Rose, an
d M.
M
i
chel
, [2
3]
p
r
o
v
i
d
es a
n
e
x
p
e
ri
m
e
nt
al
evaluat
i
on
o
f
seve
r
a
l
t
r
acki
n
g al
g
o
ri
t
h
m
s
on t
h
e
adva
nce
d
vi
d
e
o a
n
d
sig
n
a
l
b
a
sed
surv
eillan
ce, in
a con
f
eren
ce on adv
a
n
c
ed
v
i
d
e
o
an
d sign
al
b
a
sed
su
rv
eillan
c
e, using
d
a
taset for
fo
llowing
v
a
ri
o
u
s
p
e
op
le. Th
e effort of su
ch
rev
i
ews
i
s
o
f
b
e
tter-quality st
ill, as i
n
[2
4
]
wh
ich
d
i
scu
ss
tracking speci
fic targets only, for exam
pl
e, sports players.
The survey of J.
C. McCall
a
n
d
M. M. Tr
ived
i [
2
5
]
is on
track
i
ng
p
a
th
s fo
r driv
er aid.
In this pa
pe
r,
initially, a
m
o
ving object is
se
lected by fram
e
differenc
ing m
e
thod and extracte
d
ob
ject
by
seg
m
ent
t
h
resh
ol
d
i
ng. T
h
e b
u
bbl
e sort
al
go
ri
t
h
m
(B
SA) ar
ra
nge
s t
h
e regi
o
n
s (l
ar
ge t
o
s
m
al
l
)
t
o
make sure that
there is at leas
t one
big re
gion (object
).
To
track
the obj
ect, a m
o
tio
n
m
o
d
e
l is con
s
tru
c
t
e
d
t
o
set th
e syste
m
m
o
d
e
ls o
f
th
e
α
-
β
filter, LKF an
d
EKF. Man
y
exp
e
rim
e
n
t
s h
a
v
e
b
e
en
con
d
u
c
ted
on
b
a
l
l
s with
diffe
re
nt sizes
in im
age sequences a
n
d com
p
are
d
thei
r
t
r
a
c
ki
n
g
per
f
o
r
m
a
nce i
n
n
o
rm
al
l
i
ght
an
d
bad
l
i
ght
co
nd
itio
ns. The p
a
ram
e
ters o
b
t
ain
e
d
are RMSE, AE, OTE, TDR
,
and p
eak
si
g
n
a
l
-
to
-no
i
se ratio
(PSNR
)
.
Ex
peri
m
e
nt
al
resul
t
s
i
ndi
cat
e
t
h
at
t
h
e pr
op
ose
d
al
go
ri
t
h
m
s
per
f
o
r
m
wel
l
for t
h
e det
ect
i
on an
d t
r
ac
ki
ng
of
di
ffe
re
nt
si
zes
of
bal
l
s
i
n
vi
de
o se
q
u
ences
ca
pt
u
r
ed
by
usi
n
g a
st
at
i
c
i
p
h
o
n
e6
cam
era. T
h
e
or
ga
ni
zat
i
on
of
t
h
e
pape
r i
s
as fol
l
ows
.
The m
ovi
ng
ob
ject
det
e
ct
i
on an
d t
r
ac
k
i
ng m
e
t
hods ar
e pr
op
ose
d
i
n
Sect
i
on
2. Sect
i
on 3
prese
n
t
s
e
x
peri
m
e
nt
al
resul
t
s
and
di
sc
ussi
on
s. Fi
n
a
l
l
y
, t
h
e c
oncl
u
si
o
n
i
s
gi
ven
i
n
Sect
i
o
n
4.
2.
THE PROPOSED
METHOD
2.
1.
Sim
p
le Back
ground
Su
btr
a
ction
Sim
p
l
e
B
ackg
r
o
u
nd
Su
bt
ract
i
on
(SB
S
) i
s
t
h
e
base f
o
r se
veral
devel
o
p
m
ent
s
i
n
o
b
j
e
c
t
det
ect
i
on
[2
6-
2
7
]
w
h
i
c
h
det
ect
s m
ovi
ng
o
b
j
ect
s thro
ugh
calcu
latin
g th
e ab
so
lu
t
e
differe
n
ce
between t
h
e re
ference
im
age
,
[
2
8]
a
n
d i
n
com
i
ng
vi
d
e
o
fram
e
,
. Notice that t
h
e
refe
rence
im
age
,
mu
s
t
h
a
v
e
o
n
l
y static b
a
ck
gro
und
; th
e in
co
m
i
n
g
f
r
a
me f
eatu
r
es po
ssib
l
e m
o
v
i
ng
o
b
j
ects al
on
g w
ith b
a
ckg
r
ound
inf
o
rm
ation. A
f
ter the
refe
re
nce im
age
,
an
d t
h
e i
n
c
o
m
i
ng vi
deo
fram
e
,
are taken
from a
vi
de
o se
q
u
ence
, t
h
e
det
ect
ed
b
i
nary
m
o
t
i
on
d
e
t
ect
i
on m
a
sk
,
is calcu
lated
as
fo
llows:
,
1;
if|
,
,
|
0;
if|
,
,
|
τ
(1)
whe
r
e,
τ
i
s
an e
m
pi
ri
call
y
a sel
ect
i
on o
f
t
h
res
hol
d,
whi
c
h i
s
use
d
t
o
di
st
i
n
g
u
i
s
h
pi
xel
s
of
m
ovi
ng o
b
j
ect
s fr
om
th
o
s
e of
th
e back
gro
und
in
an im
age frame. The
existe
nce of m
oving
objects is i
ndi
cated if the
absolute
diffe
re
nce bet
w
een t
h
e re
fe
rence im
age
,
and t
h
e i
n
c
o
m
i
ng
vi
de
o
fram
e
,
go
b
e
y
o
nd
τ
these
pi
xel
s
of
t
h
e de
t
ect
i
on
m
a
sk
,
are th
en
lab
e
led
with
1
an
d tho
s
e th
at
are
non-active are la
beled
with
0.
2.
2.
Bubble S
o
rt Algorithm
Th
e
bu
bb
le sort algo
rith
m
is b
a
sed
on
a co
m
p
ariso
n
of
consecutive
ele
m
ents of a li
st. It
uses
a
n
ob
ject
det
ect
i
on p
r
ocess. B
u
b
b
l
e
so
rt
alg
o
rit
h
m
arran
g
es th
e regi
on
s (lar
g
e
to
sm
all) to
mak
e
sure th
at
th
ere is
at
l
east
one bi
g re
gi
o
n
(t
ar
g
e
t
ob
ject
) i
n
o
b
ject
det
ect
i
o
n
pr
ocess.
It
pe
rf
orm
s
a seri
es of
passes o
v
e
r t
h
e
sequence:
a.
Com
p
are each ele
m
ent (not
includi
ng the l
a
st one
)
with
i
t
s neighbour to the
ri
ght-if they are
out of
order, swap t
h
e
m
. Once t
h
e
largest
elem
ent is reache
d
, it
kee
p
s on
swa
ppi
ng until it goe
s to t
h
e last
p
o
s
ition
.
b.
Com
p
are each ele
m
ent (not includi
ng t
h
e last two)
with
its neighbour to the ri
ght-i
f
they are out of
order, swap t
h
e
m
. Once the
second
largest e
l
e
m
ent is reached, it ke
e
p
s
on swappi
ng
until it goes to the
secon
d
last po
sitio
n
.
Th
is
p
r
o
c
ess con
tinu
e
s
un
til th
ere are
no
u
n
s
orted
elemen
ts are left.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
d
onesi
a
n
J
E
l
ec En
g &
C
o
m
p
Sci
ISS
N
:
2
5
0
2
-
47
52
A
n
e
w method
fo
r b
a
ll
tra
cki
ng
b
a
s
ed
on
α
-
β
, l
i
n
ea
r K
a
l
m
a
n
an
d
Ext
e
n
d
e
d
…
(
H
at
hi
r
a
m
N
e
nav
at
h)
99
1
2.
3
Mathematica
l
Mo
delling
Th
e m
a
in
o
b
j
e
ctiv
e o
f
obj
ect track
ing
is t
o
esti
m
a
te
the state trajectories
of
an
o
b
ject
a
m
ovi
ng
o
r
m
ovabl
e o
b
jec
t
. It
i
s
wel
l
-
kn
o
w
n t
h
at
a
poi
nt
m
ovi
ng i
n
o
u
r
2D
space
can
be
descri
bed
b
y
i
t
s
2D
posi
t
i
on
a
n
d
v
e
lo
city v
ect
o
r
s.
X
k
x
,
y
,
x
,
y
,ω
(2)
For instance, “
E
quation (2)” can be us
ed as
a state vector of suc
h
a poi
nt
in
th
e Cartesian co
ord
i
n
a
te syste
m
,
w
h
er
e
′
de
not
e
s
m
a
t
r
i
x
t
r
a
n
s
pos
e,
x
and
y
a
r
e t
h
e
posi
t
i
o
n
coo
r
di
nat
e
s
al
on
g
x,
y
axes
,
res
p
ectively, a
n
d
x
and
y
are the
ve
locity vectors
,
and
ω
is th
e ang
u
l
ar
v
e
lo
city.
Th
e system
u
s
ed
in th
is
work is m
o
d
e
lled
u
s
ing
th
e “constan
t
v
e
lo
city” (CV) m
o
d
e
l, o
r
m
o
re p
r
ecisely “n
early-con
stan
t-v
e
lo
city
m
o
d
e
l”. Th
e
Process
state an
d d
y
n
a
mic
m
easu
r
emen
t m
o
d
e
ls are
assu
m
e
d
to
b
e
lin
ear in d
i
screte ti
m
e
. It can
b
e
written
as,
x
A
x
w
(3)
z
H
x
v
(4)
whe
r
e
x
is th
e
N
1
vector state at discrete tim
e
k
,
z
is th
e
M
1
vector m
easure
m
ent at discrete-time
k
,
H
is the consta
nt m
easurem
en
t
m
a
trix of size
M
N
,
A
is th
e con
s
tan
t
state tran
sitio
n
m
a
trix
o
f
size
M
N
,
w
is an
N
1
zero
m
ean Gaussian distri
bute
d
proces
sing(or s
y
ste
m
) noise vector at tim
e
h
a
vi
ng
constant c
o
vari
ance m
a
trix
of
, and
is an
1
z
e
ro m
ean ra
ndom
m
easurem
e
n
t
noise
vect
or at tim
e
h
a
v
i
ng
a co
nstan
t
po
sitiv
e
defin
ite co
v
a
rian
ce m
a
trix
of
.
,
are
ass
u
m
e
d to
be
indepe
ndent
of eac
h
ot
he
r f
o
r al
l
and
.
It
i
s
ass
u
m
e
d t
h
at
t
h
e
peri
od
bet
w
een ea
ch
observation
is a constant
∆
s
econds
.
2.
4
Trac
king Al
gori
t
hms
2.
4.
1
α
-
β
filtering
The
α
-
β
t
r
ac
ki
ng
[
2
9]
al
go
ri
t
h
m
can be e
x
pr
essed a
s
f
o
l
l
o
w
s
. P
r
edi
c
t
i
o
n st
ep
(t
im
e up
dat
e
):
x
x
∆
s
(5)
s
s
(6)
whe
r
e t
h
e
ve
ctor
x
and
s
=
x
are th
e
p
r
ed
icted
p
o
s
ition
m
a
trix
an
d
pred
icted
v
e
lo
city matrix
,
respectively. Correction st
ep (measurem
ent update
):
x
x
α
z
x
(7)
s
s
∆
z
x
(8)
whe
r
e the
vect
or
,
̂
are th
e esti
m
a
ted
p
o
s
itio
n
m
a
trix
and
es
ti
m
a
ted
v
e
lo
city
matrix
, resp
ectiv
ely,
α
and
β
are tun
i
ng
co
nstan
t
s b
e
t
w
een
nu
m
b
er 0
and
n
u
m
b
e
r 1
to
sm
o
o
t
h
t
h
e po
sition
and
v
e
l
o
city esti
mates,
respectively.
2.
4.
2
Linea
r
Ka
lman Filter
Th
e LKF
is
a ti
m
e
-d
o
m
ain
recu
rsi
v
e filter with
th
e cap
a
bilit
y to
estim
a
t
e th
e state
o
f
a d
y
n
a
m
i
c
sy
st
em
by
usi
ng a se
ri
es o
f
m
easurem
ents. C
o
nsi
d
e
r
i
n
g
t
h
e sy
st
em
model
e
quat
i
o
ns
“Equat
i
on
(3
)
”
and
“Equ
atio
n (4
)”; th
e Lin
e
ar
Kal
m
an
filter equatio
n
s
are
g
i
v
e
n
b
y
[30
]
,
P
A
P
A
Q
,
(9)
K
P
H
HP
H
R
(10)
x
A
x
(11)
x
x
K
z
H
x
(12)
P
(13)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
502
-47
52
I
ndo
n
e
sian
J Elec Eng
& Com
p
Sci, V
o
l. 10
,
No
.
3
,
Jun
e
2
018
:
98
9 – 99
9
99
2
fo
r
1
,
2
,
…
,
wh
er
e
is th
e i
d
en
tity m
a
t
r
ix
.
is th
e
a prio
ri
estim
a
t
e of the
state
gi
v
e
n m
easurem
ent
s
u
p
to and
i
n
clud
ing
tim
e
1
.
is
th
e
a post
e
ri
or
i
estim
a
te of the state
specifi
ed m
easurem
e
n
ts
up t
o
and
i
n
cl
udi
ng
t
i
m
e
.
is t
h
e
Kalm
an
g
a
in
,
is t
h
e c
o
varia
n
ce
of t
h
e
a prio
ri
estim
ation e
r
ror
,
and
is th
e
co
v
a
rian
ce
o
f
th
e
a po
steri
o
ri
estim
ation error
.
Th
e
Kalm
an
filter
is
in
itialized
b
y
:
x
E
x
(14)
P
E
0
0
0
0
(15)
whe
r
e
.
is the e
xpectation
ope
r
ator.
2.
4.
3
Extended Kal
m
an filter
The
EKF
31
is
t
he
n
onlinea
r
v
ersion
o
f
t
h
e
Ka
lma
n
f
ilt
e
r.
T
he
n
o
n
l
i
n
e
a
r
m
o
d
e
l
s
f
o
r
E
K
F
are
gi
ve
n by
,
Process M
odel
:
x
f
x
w
(16)
Me
asurement Model:
(17)
Th
e al
g
o
rith
m for t
h
e ex
tend
ed Kalm
an
fi
lterin
g
is esse
n
tially si
milar to
th
at
of Linear
Kalm
an
filterin
g
,
except
f
o
l
l
o
wi
ng
m
odi
fi
cat
i
ons,
F
|
x
(18)
H
|
x
(19)
3.
EX
P
E
R
I
M
E
NT
S
3.
1.
Impleme
ntation de
tails
In
th
is
wo
rk
,
we i
m
p
l
e
m
en
ted
o
u
r track
e
r i
n
MATLAB R
2
0
1
3
a
PC m
ach
in
e with
In
tel i7-377
0
CPU
(3.4GHz)
with 2GB m
e
m
o
ry, wh
ich
ru
n
s
29fp
s
on
t
h
is
p
l
atform
. In
add
itio
n, th
e self-m
a
d
e
v
i
d
e
o
con
s
istin
g
of J
P
EG im
age sequence
wi
th
720
x128
0
pi
xel
s
per
fram
e
i
s
t
a
k
e
n usi
ng t
h
e i
P
h
one
6 cam
era. Th
e
envi
ro
nm
ent
for t
h
e ba
d l
i
g
ht
con
d
i
t
i
on i
s
creat
ed by
swi
t
c
hi
n
g
o
ff a
l
l
t
h
e li
ght
s i
n
t
h
e room
(im
p
r
o
per
illu
m
i
n
a
tio
n
)
,
wh
ile th
e env
i
ron
m
en
t is co
n
s
id
ered
as a
n
o
rm
al
lig
h
t
co
nd
itio
n
wh
en th
e roo
m
is p
r
op
erly
illu
m
i
n
a
ted
.
3.
1.
1
Experimental Setup
The dista
n
ce c
onsi
d
ere
d
for tracki
ng t
h
e target obj
ect is a
p
proxim
a
tely
6m
eters with target objects
bei
n
g bal
l
s
of
di
ffe
rent
si
zes (sm
a
l
l
,
m
e
dium
and bi
g)
wh
ose ra
di
i
are
3.
32
5c
ms, 4.
9cms an
d
8
.
5cms
,
respect
i
v
el
y
.
T
h
e
vi
de
o ca
pt
u
r
ed
usi
n
g i
P
ho
ne6
as
f
o
llo
wi
n
g
sp
ecification
s
: Im
ag
e Ty
pe: RGB co
lou
r
sp
ace,
Im
age Dim
e
nsi
on:
720x1280
pi
xel
s
(
H
ei
g
h
t
=
1
2
8
0
and
Wi
dt
h= 7
2
0
)
, F
r
am
e R
a
t
e
:
29 f
p
s,
Nu
m
b
er of
fram
e
s: Fo
r No
rm
al Lig
h
t
co
n
d
ition
(For small s
i
ze b
a
ll
=5
4
fram
e
s, Fo
r m
e
d
i
u
m
siz
e
b
a
ll= 6
0
frames, For
bi
g si
ze bal
l
=
5
8
f
r
am
es) and F
o
r B
a
d Li
ght
c
o
n
d
i
t
i
on
(O
nl
y
for sm
al
l
si
ze
bal
l
=
70
fram
e
s). Tabl
e
1 sh
o
w
s t
h
e
in
itial v
a
lu
es
of th
e b
a
sic
p
a
rameters for th
e t
r
ack
i
n
g algo
rit
h
m
s
.
3.
2
Perfor
mance Measures
a.
Ab
so
lu
te erro
r (AE):
AE is th
e m
a
g
n
itu
d
e
o
f
t
h
e d
i
fference b
e
tween
th
e tru
e
v
a
lu
e and th
e track
ed
v
a
lue
of the
object.
ϵ
|
x
x
|
(20)
whe
r
e
x
t
h
e
t
r
ue val
u
e o
f
ob
ject
pa
ram
e
t
e
rs
and
x
is th
e track
e
d
v
a
lu
e
of th
e
o
b
j
ect
p
a
ram
e
t
e
rs.
b.
Ro
o
t
Mean
Squ
a
re Error(RM
S
E): RMSE is
o
n
e
of th
e
m
o
st wid
e
ly u
s
ed
fu
ll-referen
ce
qu
ality assess
men
t
m
e
t
r
i
c
, whi
c
h i
s
com
put
ed by
squ
a
re r
oot
o
f
t
h
e avera
g
e o
f
squ
a
re
d i
n
t
e
nsi
t
y
di
fferen
ces bet
w
ee
n t
r
acke
d
(
x
) an
d t
r
ue i
m
age
pi
xel
s
(
x
).
Evaluation Warning : The document was created with Spire.PDF for Python.
In
d
onesi
a
n
J
E
l
ec En
g &
C
o
m
p
Sci
ISS
N
:
2
5
0
2
-
47
52
A
n
e
w method
fo
r b
a
ll
tra
cki
ng
b
a
s
ed
on
α
-
β
, l
i
n
ea
r K
a
l
m
a
n
an
d
Ext
e
n
d
e
d
…
(
H
at
hi
r
a
m
N
e
nav
at
h)
99
3
RMSE
∑∑
x
x
(21)
whe
r
e
N a
n
d
M are the
im
age dim
e
nsions.
c.
Peak
Si
g
n
al
- t
o
-
N
o
i
s
e R
a
t
i
o
(PS
N
R
)
:
The
val
u
e
o
f
Pea
k
si
gnal
-
t
o
-
n
oi
se rat
i
o
ca
n easi
l
y
be
obt
ai
ne
d
by
usi
n
g t
h
e m
ean sq
uare
d e
r
r
o
r.
PSNR
1
0
∗
l
og
(22)
M
ean s
qua
re e
r
r
o
r:
MSE
∑∑
x
x
(23)
whe
r
e, L is
the dynam
i
c range
of the im
age (8
bits pe
r
pixel gray scale i
m
ages L =
255).
d.
Tracki
n
g
d
e
t
ect
i
on
rat
e
(T
DR
):
Trac
ki
n
g
de
t
ect
i
on rat
e
is
th
e ratio of a
n
u
m
b
e
r
o
f
frames in
wh
ich
th
e
o
b
j
ect is
d
e
tected
to th
e t
o
tal
n
u
m
b
e
r
o
f
frames in
wh
ich th
e object pres
e
n
t.
TDR
∗
100
(24)
e.
Object trac
king error
(OTE): Object
track
i
n
g
error is the no
rm
al in
co
nsisten
c
y in
th
e cen
tro
i
d
of t
h
e
t
r
acke
d
ob
ject
fr
om
i
t
s
t
r
ue va
l
u
e.
It
i
s
gi
ve
n
by
,
OTE
∑
(25)
whe
r
e,
x
and
y
are the act
ual 2D coordinates
of the
object,
x
and
y
are t
h
e trac
ked
2D
coordinates of the
object.
Tabl
e 1. Si
m
u
lat
i
on param
e
t
e
rs
f
o
r
T
r
acke
r
s
α
-
β
LKF
EKF
Para
m
e
ters
Valu
es
Para
m
e
ters
Valu
es
Para
m
e
ters
Valu
es
α
0.
9
dt
1
dt
1
β
0.
005
1
0
0
0
0
1
0
0
0
1
0
0
0
1
dt
dt
1
0
0
0
0
1
0
0
0
1
0
0
0
1
dt
dt
∆
1
0
0
1
0
0
0
0
1
0
0
1
0
0
0
0
1
--
--
0055
.
0
055
.
0
0143
.
0
1524
.
0
0045
.
0
0045
.
0
0045
.
0
2845
.
0
--
--
)
4
(
01
.
0
eye
)
4
(
01
.
0
eye
--
--
)
4
(
100
eye
)
4
(
100
eye
3.
3
Experimental results
3.
3.
1
Tra
c
k
i
ng
res
ult
s
of
s
m
a
ll s
i
ze
ba
ll in no
rma
l
lig
ht co
nditio
n
The RMSE
s
of the
LKF,
α
-
β
filter an
d EKF
for sm
all size b
a
ll d
a
taset are
g
i
v
e
n
in Tab
l
e
2
.
As
obs
er
ved i
n
Ta
bl
e 2, t
h
e a
v
er
age val
u
e of t
h
e R
M
SE was r
e
duce
d
aft
e
r appl
y
i
n
g
t
h
e L
K
F t
o
sm
al
l bal
l
dat
a
.
On a
n
a
v
erage
,
the m
i
nim
u
m
RMSE reducti
on
was
0.
03
30 an
d th
e m
a
x
i
m
u
m
w
a
s 0
.
447
8
fo
r
sm
all si
ze b
a
ll
d
a
ta. Th
is
resu
lt ind
i
cates th
at th
e pred
icted
ou
tpu
t
of
t
h
e L
K
F is cl
oser t
o
the
desired output. The bes
t
avera
g
e values
are em
phasized in Ta
ble
2.
Sim
ilarly, the
PSNRs
of t
h
e LKF,
α
-
β
filter and
EKF fo
r s
m
all
si
ze bal
l
dat
a
s
e
t
are
gi
ve
n i
n
Tabl
e
2.
As
o
b
s
erve
d i
n
Ta
bl
e 2,
t
h
e
ave
r
ag
e val
u
e
of t
h
e
PSNR
was
i
n
c
r
ease
d
after app
l
yin
g
th
e LKF to
sm
a
ll b
a
ll
d
a
ta. On
a
n
average, the m
a
xim
u
m
PSN
R w
a
s 48
.4
4d
B
an
d
the
m
i
nim
u
m
was 33
.1
0
d
B
f
o
r s
m
al
l
si
ze bal
l
dat
a
. Fi
g
u
r
e1
s
h
ows t
h
e error of ce
ntroid
, er
ro
r o
f
ra
diu
s
,
RM
SE,
cent
r
oi
d a
n
d
r
a
di
us
of
x
-
y
c
o
o
r
di
nat
e
s,
AE
and
OT
E i
n
e
ach f
r
am
e for
sm
al
l
si
ze bal
l
base
d o
n
L
K
F u
nde
r
n
o
rm
al lig
h
t
co
nd
itio
ns. Fi
g
u
re 1 shows the erro
r of cen
t
ro
id
, error
o
f
rad
i
u
s
, RMSE, cen
t
ro
id
and
rad
i
u
s
of
x-y
c
o
or
di
nat
e
s, A
E
a
n
d
O
T
E
i
n
eac
h
fram
e
fo
r sm
al
l
si
ze bal
l
based
o
n
L
K
F
u
nde
r
n
o
r
m
al
l
i
ght
co
ndi
t
i
ons.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
502
-47
52
I
ndo
n
e
sian
J Elec Eng
& Com
p
Sci, V
o
l. 10
,
No
.
3
,
Jun
e
2
018
:
98
9 – 99
9
99
4
The OTEs of
t
h
e
L
K
F,
α
-
β
fi
l
t
e
r and E
K
F
f
o
r sm
al
l
si
ze bal
l
dat
a
set
are gi
ve
n i
n
Ta
bl
e
2.
As o
b
se
rv
e
d
i
n
Table
2, the
average
val
u
e of
the OT
E
was reduce
d
a
f
ter
ap
p
l
ying
th
e LKF t
o
sm
all b
a
ll d
a
ta. On
an
av
erag
e,
th
e m
i
n
i
m
u
m
OTE
red
u
c
tion was
0
.
04
76
an
d th
e m
a
x
i
mu
m
was 1
.
35
62
for sm
all siz
e
b
a
ll d
a
ta. This resu
l
t
in
d
i
cates th
at th
e pred
icted
ou
tpu
t
of th
e LKF is clo
s
er
to th
e d
e
sired
o
u
tp
u
t
. Sim
ilarly,
th
e AEs
of the LKF,
α
-
β
filter and
EKF fo
r sm
all
size b
a
ll
d
a
taset are
g
i
v
e
n
i
n
Tab
l
e
2
.
As observ
e
d
i
n
Tab
l
e 2
,
th
e av
erage v
a
l
u
e
of t
h
e A
E
w
a
s
red
u
ce
d aft
e
r
ap
pl
y
i
ng t
h
e
LKF t
o
sm
all
ball data.
On
an a
v
era
g
e, t
h
e minim
u
m
AE was
0.
12
7
6
a
n
d
t
h
e
m
a
xim
u
m
was 0.
9
8
9
8
f
o
r
sm
al
l
si
ze bal
l
dat
a
.
Fi
gu
re
1.
The
e
xpe
ri
m
e
nt
al
resul
t
s
o
f
sm
al
l
bal
l
t
r
acki
n
g
usi
n
g
L
K
F
3.
3.
2
Trackin
g
resu
l
ts of me
dium
siz
e
ba
ll in norma
l lig
ht co
nditio
n
The OT
Es of the LKF,
α
-
β
filter an
d
EKF for m
e
d
i
u
m
si
ze b
a
ll d
a
taset are g
i
v
e
n
in
Tab
l
e 2
.
As
obs
erved in Ta
ble
2, t
h
e a
v
erage
value
of t
h
e OT
E
wa
s
red
u
ced
aft
e
r a
ppl
y
i
ng t
h
e L
K
F
t
o
m
e
di
um
si
ze bal
l
dat
a
.
O
n
a
n
a
v
erage
,
t
h
e
m
i
nim
u
m
OTE re
d
u
ct
i
o
n
was
0.
0
5
9
1
a
n
d t
h
e m
a
xi
m
u
m
was 0
.
69
7
2
fo
r m
e
di
um
si
ze
ball data. The
best avera
g
e values are em
phasized
in
Tab
l
e 2
.
Similarly,
th
e AEs of th
e LKF,
α
-
β
filter an
d
EKF
f
o
r
m
e
di
um
si
ze bal
l
dat
a
set
are
gi
ve
n
i
n
Ta
bl
e 2
.
As
obs
er
ved
i
n
Ta
bl
e 2
,
t
h
e a
v
er
age
val
u
e
o
f
t
h
e A
E
was redu
ced
after app
l
yin
g
the LKF to
sm
all b
a
ll d
a
ta.
On
a
n
a
v
er
ag
e
,
th
e
min
i
mu
m A
E
w
a
s
0.
214
6
and
th
e
max
i
m
u
m
was 1
.
25
19
fo
r
small size b
a
ll d
a
ta. Th
is resu
lt
in
d
i
cates th
at t
h
e
p
r
ed
icted
ou
tpu
t
of t
h
e LKF i
s
cl
oser
t
o
t
h
e
de
si
red o
u
t
p
ut
.
Tabl
e
2.
A
v
era
g
e
param
e
t
e
rs com
p
ari
s
on
f
o
r
di
f
f
ere
n
t
si
ze
of
bal
l
s
usi
n
g
f
i
ve t
r
ac
kers
u
n
d
er
n
o
rm
al an
d
b
a
d
ligh
t
cond
ition
s
In Norm
a
l
light co
ndition
In bad light conditi
on
Para
m
e
te
rs
For s
m
all
size ball
For m
e
diu
m
size
ball
For big size ball
For s
m
all
size ball
α
-
β
LKF
EKF
α
-
β
LKF
EKF
α
-
β
LKF
EKF
α
-
β
LKF
EKF
AE
0.
9
8
0.
12
0.
76
0.
7
4
0.
21
1.
25
1.
31
0.
63
1.
34
0.
45
0.
11
1.
05
RMSE
0.
4
4
0.
03
0.
17
0.
2
8
0.
04
0.
28
0.
45
0.
21
0.
30
0.
09
0.
02
0.
17
TDR (%
)
96.
2
100
100
91.
6
100
91.
6
93.
1
100
93.
1
85.
7
100
85.
7
OTE
1.
3
5
0.
04
0.
16
0.
6
9
0.
05
0.
29
1.
01
0.
48
0.
29
0.
22
0.
02
0.
14
0
10
20
30
40
50
60
-1
-0
.
8
-0
.
6
-0
.
4
-0
.
2
0
0.
2
0.
4
F
r
am
e N
u
m
ber
D
i
s
t
anc
e
(
i
n
m
e
t
e
r
s
)
E
r
ror
of
k
a
l
m
an f
i
l
t
er c
ent
r
o
i
d
V
s
F
r
am
e
N
u
m
b
er
0
10
20
30
40
50
60
-0
.
4
-0
.
3
-0
.
2
-0
.
1
0
0.
1
0.
2
0.
3
0.
4
0.
5
F
r
am
e N
u
m
b
er
D
i
sta
n
ce
(
i
n
m
e
t
e
r
s
)
E
r
r
o
r
of
k
a
lm
an
f
i
l
t
er
r
a
d
i
u
s
V
s
F
r
am
e N
u
m
b
er
0
10
20
30
40
50
60
0
0.
05
0.
1
0.
15
0.
2
0.
25
0.
3
0.
35
F
r
am
e N
u
m
b
er
er
r
o
r
R
o
ot
M
e
an S
q
u
a
re err
o
r
RM
S
E
0
10
20
30
40
50
60
0
100
200
300
400
500
600
700
800
900
F
r
am
e N
u
m
b
e
r
c
e
nt
r
o
i
d
an
d rad
i
u
s
x
-
y
c
o
adi
na
t
e
s
A
c
t
ual
c
oordi
nat
es
ce
n
t
r
o
i
d
r
a
di
us
0
10
20
30
40
50
60
0
0.
1
0.
2
0.
3
0.
4
0.
5
0.
6
0.
7
0.
8
0.
9
F
r
am
e N
u
m
b
e
r
Er
r
o
r
A
b
s
o
l
u
t
e
e
rro
r (A
E
)
AE
0
10
20
30
40
50
60
0
0.
0
2
0.
0
4
0.
0
6
0.
0
8
0.
1
0.
1
2
0.
1
4
F
r
am
e
N
u
m
b
er
Er
r
o
r
O
b
j
e
ct T
r
a
c
ki
n
g
E
r
r
o
r
(
O
T
E
)
OT
E
Evaluation Warning : The document was created with Spire.PDF for Python.
In
d
onesi
a
n
J
E
l
ec En
g &
C
o
m
p
Sci
ISS
N
:
2
5
0
2
-
47
52
A
n
e
w method
fo
r b
a
ll
tra
cki
ng
b
a
s
ed
on
α
-
β
, l
i
n
ea
r K
a
l
m
a
n
an
d
Ext
e
n
d
e
d
…
(
H
at
hi
r
a
m
N
e
nav
at
h)
99
5
The RMSEs
of the LKF,
α
-
β
filter an
d
EKF
fo
r m
e
d
i
u
m
siz
e
b
a
ll d
a
taset are g
i
v
e
n
i
n
Table 2
.
On
an
avera
g
e, t
h
e m
i
nim
u
m
R
M
SE red
u
ct
i
o
n was
0.
04
8
9
an
d t
h
e
m
a
xim
u
m
was 0.
28
6
9
f
o
r m
e
di
um
si
ze bal
l
dat
a
.
Si
m
ilarly,
th
e PSNRs
o
f
th
e
LKF,
α
-
β
filter an
d
EKF for med
i
u
m
s
i
ze b
a
ll d
a
taset are g
i
v
e
n
in
Tab
l
e 2
.
As
obs
erved i
n
Ta
ble 2, the a
v
erage va
lue of
t
h
e
PSNR was increase
d
afte
r
appl
y
i
n
g
t
h
e
LKF t
o
m
e
di
um
bal
l
data. On an average
,
the m
a
x
i
m
u
m
PSNR
was 45
.1
6
d
B
an
d t
h
e m
i
nim
u
m
was 28.
8
2
d
B
for m
e
di
um
si
ze bal
l
d
a
ta.
Fi
gu
r
e
2 sh
ow
s
th
e err
o
r
o
f
cen
tro
i
d,
er
ro
r
of
r
a
d
i
u
s
, RMSE,
cen
t
ro
id
an
d
r
a
d
i
us o
f
x-
y
coor
d
i
n
a
tes, A
E
an
d OTE in each
fram
e
fo
r med
i
u
m
size b
a
ll b
a
sed
on
LKF un
d
e
r
no
rm
al l
i
g
h
t
co
nd
ition
s
.
Fi
gu
re
2.
The
e
xpe
ri
m
e
nt
al
resul
t
s
o
f
m
e
di
u
m
si
ze bal
l
t
r
acki
n
g
usi
n
g L
K
F
3.
3.
3
Trackin
g
resu
l
ts of bi
g siz
e
ball in n
o
rm
al
light condition
The OTE
s
of the
L
K
F,
α
-
β
filter and
EKF fo
r b
i
g
size
ball d
a
taset are g
i
v
e
n in
Tab
l
e 2
.
On an
avera
g
e, t
h
e
m
i
nim
u
m
OTE red
u
ct
i
o
n
w
a
s 0.
2
9
9
7
a
n
d
t
h
e m
a
xim
u
m
was 1.
0
1
0
2
fo
r bi
g si
ze
bal
l
dat
a
.
Si
m
ilarly, th
e
AEs of th
e LKF,
α
-
β
filter and
EKF fo
r
b
i
g
size b
a
ll d
a
taset are g
i
v
e
n
in
Tab
l
e 2
.
As
o
b
serv
ed
in
Tab
l
e 2
,
the av
erag
e v
a
lue o
f
th
e AE
was red
u
c
ed
after app
l
yin
g
th
e LKF to
sm
all b
a
ll d
a
ta.
On
an
avera
g
e,
t
h
e m
i
ni
m
u
m
AE wa
s 0.
6
3
9
9
9
a
n
d t
h
e m
a
xim
u
m
was
1.
34
1
2
fo
r
bi
g
si
ze
bal
l
d
a
t
a
.
The RMSEs
of the L
K
F,
α
-
β
filter an
d
EKF for
b
i
g
size
b
a
ll d
a
taset are g
i
v
e
n
in
Table 2
.
On
an
avera
g
e, t
h
e m
i
nim
u
m
R
M
SE red
u
ct
i
o
n
was
0.
2
1
1
2
a
n
d
t
h
e m
a
xim
u
m
w
a
s 0.
4
5
4
5
fo
r
b
i
g si
ze bal
l
dat
a
. Th
e
best avera
g
e
values are
em
phasized in
Tabl
e 2
.
Sim
i
lar
l
y, t
h
e PSN
R
s
of
t
h
e LK
F,
α
-
β
fi
lter
an
d
E
K
F f
o
r big
si
ze bal
l
dat
a
s
e
t
are
gi
ve
n i
n
Tabl
e
2.
As
o
b
s
erve
d i
n
Ta
bl
e 2,
t
h
e
ave
r
ag
e val
u
e
of t
h
e
PSNR
was
i
n
c
r
ease
d
aft
e
r ap
pl
y
i
ng
t
h
e LKF t
o
bi
g
bal
l
dat
a
. O
n
an ave
r
ag
e, t
h
e
m
a
xim
u
m
PSNR
was
45
.0
0
d
B
an
d t
h
e m
i
ni
m
u
m
was 2
7
.
9
1dB
f
o
r
bi
g si
ze bal
l
dat
a
. Fi
g
u
re
3
sho
w
s t
h
e err
o
r o
f
cent
r
oi
d
,
e
r
r
o
r
of
radi
us,
R
M
SE, cent
r
oi
d an
d
radi
us of x-y coordinates,
AE and
OTE in each fram
e
for big size
ball based
on
LKF unde
r norm
al light
co
nd
itio
ns.
3.
3.
4
Trackin
g
resu
l
ts of sm
all siz
e
ball in
bad light c
o
nditi
on
The OT
Es of t
h
e LKF
,
α
-
β
filter an
d
EKF fo
r sm
all
size b
a
ll d
a
taset in
b
a
d
lig
h
t
con
d
itio
n
are
g
i
ven
in Table
2. As
observe
d
in
T
a
ble 2, the a
v
e
r
age
va
lue
of
th
e
O
T
E
w
a
s
re
d
u
c
ed
af
te
r
ap
p
l
ying
th
e LK
F
to
sm
al
l
bal
l
dat
a
. O
n
an a
v
e
r
ag
e, t
h
e m
i
nim
u
m
OTE re
duct
i
on
was
0.
02
2
7
an
d t
h
e
m
a
xim
u
m
was 0.
2
2
75
f
o
r
sm
a
ll size b
a
ll d
a
ta. Th
is result in
d
i
cates th
at th
e pred
ic
t
e
d
out
put
of
t
h
e L
K
F i
s
cl
oser
t
o
t
h
e
desi
re
d o
u
t
put
.
The best
a
v
e
r
a
g
e values
are
e
m
phasized
in
Tab
l
e 2.
Sim
i
larly,
th
e AEs o
f
th
e LKF,
α
-
β
filter and
EKF
fo
r
sm
a
ll size
b
a
ll
d
a
taset are g
i
ven
in
Tab
l
e 2
.
As ob
serv
ed
in Table 2, the a
v
era
g
e val
u
e of the AE
was reduc
e
d
after app
l
yin
g
th
e LKF to
sm
all b
a
ll d
a
ta. On
av
erag
e, th
e
min
i
m
u
m
AE
was 0.119
4
and
th
e m
a
x
i
m
u
m was
1
.
0
528
for sm
all size b
a
ll d
a
ta.
The RMSEs of the LKF,
α
-
β
filter an
d
EKF fo
r sm
all s
i
ze b
a
ll d
a
taset in
b
a
d
lig
h
t
co
nd
itio
n
are
gi
ve
n i
n
Tabl
e
2. As
ob
ser
v
ed
i
n
Tabl
e 2. O
n
an avera
g
e, t
h
e
m
i
nim
u
m
RM
SE red
u
ct
i
o
n
was 0.
02
1
8
an
d t
h
e
0
10
20
30
40
50
60
-1
-0
.
5
0
0.
5
1
1.
5
2
F
r
am
e N
u
m
ber
D
i
st
a
n
ce
(
i
n
m
e
te
r
s
)
E
r
ror
of
k
a
l
m
an
f
i
l
t
er
c
ent
roi
d
V
s
f
r
am
e i
ndex
0
10
20
30
40
50
60
-0
.
4
-0
.
3
-0
.
2
-0
.
1
0
0.
1
0.
2
0.
3
0.
4
F
r
a
m
e N
u
m
b
er
Di
s
t
a
n
c
e
(
i
n
m
e
t
e
rs
)
E
r
ror of
k
a
l
m
an f
i
l
t
er
rad
i
u
s
V
s
f
r
a
m
e i
n
dex
0
10
20
30
40
50
60
0
0.
05
0.
1
0.
15
0.
2
0.
25
0.
3
0.
35
F
r
am
e N
u
m
ber
err
o
r
R
oot
M
e
an S
qua
r
e
err
o
r
RM
S
E
0
10
20
30
40
50
60
0
100
200
300
400
500
600
700
800
F
r
am
e N
u
m
b
e
r
c
e
nt
roi
d
an
d radi
us
x
-
y
c
oad
i
n
a
t
es
A
c
t
u
a
l
c
oor
di
nat
es
c
ent
r
o
i
d
ra
d
i
u
s
0
10
20
30
40
50
60
0
0.
2
0.
4
0.
6
0.
8
1
1.
2
1.
4
1.
6
1.
8
F
r
am
e N
u
m
b
e
r
E
rror
A
bs
ol
u
t
e
error
(
A
E
)
AE
0
10
20
30
40
50
60
0
0.
02
0.
04
0.
06
0.
08
0.
1
0.
12
F
r
am
e
N
u
m
b
er
E
rror
Obj
e
c
t
T
r
ac
k
i
ng
E
r
r
o
r
(
OT
E
)
OT
E
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
502
-47
52
I
ndo
n
e
sian
J Elec Eng
& Com
p
Sci, V
o
l. 10
,
No
.
3
,
Jun
e
2
018
:
98
9 – 99
9
99
6
max
i
m
u
m
was 0
.
1
767
fo
r sm
all size b
a
ll d
a
ta. Similarly, th
e PSNRs
o
f
t
h
e
LKF,
α
-
β
filter an
d
EKF
for sma
l
l
si
ze bal
l
dat
a
s
e
t
are
gi
ve
n i
n
Tabl
e
2.
As
o
b
s
erve
d i
n
Ta
bl
e 2,
t
h
e
ave
r
ag
e val
u
e
of t
h
e
PSNR
was
i
n
c
r
ease
d
after app
l
yin
g
th
e LKF to
sm
a
ll b
a
ll
d
a
ta. On
a
n
average, the m
a
xim
u
m
PSN
R w
a
s 45
.1
7d
B
an
d
the
m
i
nim
u
m
was 27
.8
9
d
B
f
o
r s
m
al
l
si
ze bal
l
dat
a
.
Fi
gu
re
3.
The
e
xpe
ri
m
e
nt
al
resul
t
s
o
f
bi
g
si
ze bal
l
t
r
ac
ki
n
g
usi
n
g L
K
F
3.
3.
5
Objec
t
trac
king
fr
ame res
u
l
ts o
f
sm
a
ll size ba
ll in
no
rma
l
lig
ht co
nditio
n
Fig
u
re 4
sho
w
s so
m
e
o
f
th
e test resu
lts for s
m
all s
i
ze b
a
ll u
n
d
e
r
n
o
rm
al
lig
h
t
con
d
ition
.
We h
a
v
e
evaluate
d
our
approach using sel
f-m
ade Database,
with
th
e f
r
a
m
e
size
b
e
ing
72
0
x 12
80
. W
e
h
a
v
e
selected
so
m
e
fram
e
s fro
m
th
e v
i
d
e
o
wh
ile track
i
ng
th
e ob
j
ect co
n
t
in
uo
usly d
u
ring
its
m
o
v
e
m
e
n
t
. Th
e targ
et ob
j
ect is
m
a
rked
by
t
h
e
bo
u
ndi
ng
b
o
x
(
r
ed ci
r
c
l
e
). Fi
g
u
re
4 s
h
o
w
s
4
f
r
am
es out
of t
h
e t
r
acki
n
g
resu
l
t
s
based
o
n
L
K
F i
n
Figure 4(a),
α
-
β
filter in
Figure 4(b) and
EKF in
Figu
re
4
(
c).
In
fram
es 2
0
and
30
sh
own
in
Figure 4(a), th
e
target object tracked
using L
K
F is
d
o
m
i
nated by
t
h
e re
d b
o
u
n
d
i
n
g b
o
x
as
the cam
era is
static and the size of
t
h
e ob
ject
i
s
reduce
d
wi
t
h
di
st
ance. The sa
m
e
phen
o
m
e
na i
s
obser
ved i
n
fram
e
s 30, 45
and
20
, 2
6
sh
o
w
n i
n
Fi
gu
re
4(
b) a
n
d Fi
g
u
r
e 4
(
c)
,
usi
n
g
α
-
β
filter an
d
EKF,
respectiv
ely. Ev
en
th
ou
gh
the target o
b
j
ect is small, th
e
pr
o
pose
d
t
r
ack
i
ng al
go
ri
t
h
m
s
can
d
e
tect
th
e o
b
j
ect (sm
a
ll
b
a
ll).
(a)
(
b
)
(c
)
Fig
u
re
4
.
Obj
e
ct track
ing
resu
lt of th
e sm
all
size b
a
ll: (a)
based
on
LKF
(7
,
9
,
20
and
3
0
fram
e
s),
(b)
based
on
α
-
β
filter (8, 15
, 30
and
4
5
fram
e
s) and
(c)
b
a
sed on
EKF (7
,
10
,
20
and
2
6
fram
es)
3.
3.
6
Trackin
g
fr
ame results
of
medium si
ze ba
ll in no
rma
l
lig
ht condition
T
h
e
Figure
5
s
h
ows
each
4
fram
e
s
out
of the tracking re
sults ba
s
e
d on L
K
F in
Figure 5(a),
α
-
β
filter
in
Figu
re
5(b)
an
d
EKF in
Fi
g
u
re
5
(
c).
In
spite o
f
th
e
reflectio
n
s
fro
m
th
e targ
et obj
ect on
th
e tiles of the floo
r
wh
ich
is con
s
id
er as
n
o
i
se, the LKF,
α
-
β
filter, and
EKF al
g
o
rith
m
s
are
effectiv
e in trackin
g
t
h
e
o
b
j
ect.
0
10
20
30
40
50
60
-4
-2
0
2
4
6
8
10
F
r
am
e
N
u
m
b
er
Di
s
t
a
n
c
e
(
i
n
m
e
t
e
rs
)
E
r
ror of
k
a
l
m
an
f
i
l
t
e
r
c
e
nt
roi
d
V
s
f
r
a
m
e i
nde
x
0
10
20
30
40
50
60
-3
-2.
5
-2
-1.
5
-1
-0.
5
0
0.
5
F
r
am
e N
u
m
b
e
r
D
i
s
t
anc
e
(
i
n
m
e
t
e
r
s
)
E
r
r
o
r
of
k
a
l
m
a
n
f
i
l
t
er
r
a
d
i
us
V
s
f
r
am
e
i
ndex
0
10
20
30
40
50
60
0
0.
5
1
1.
5
2
2.
5
3
3.
5
F
r
am
e
N
u
m
b
er
e
r
ror)
R
o
ot
M
ean
S
qua
r
e
err
o
r
RM
S
E
0
10
20
30
40
50
60
0
20
0
40
0
60
0
80
0
10
00
12
00
F
r
a
m
e N
u
m
b
er
c
e
n
t
r
o
i
d
an
d r
adi
us
x
-
y
c
o
a
d
i
n
at
e
s
A
c
t
ua
l
c
o
o
r
di
na
t
e
s
ce
n
t
r
o
i
d
r
a
di
us
0
10
20
30
40
50
60
0
1
2
3
4
5
6
7
8
9
10
F
r
a
m
e N
u
m
b
er
E
rror
A
b
s
o
l
u
t
e
e
r
r
o
r (A
E
)
AE
0
10
20
30
40
50
60
0
0.
2
0.
4
0.
6
0.
8
1
1.
2
1.
4
F
r
am
e N
u
m
ber
Er
r
o
r
O
b
j
e
c
t
T
r
a
c
k
i
n
g
E
rro
r(O
T
E
)
OT
E
Evaluation Warning : The document was created with Spire.PDF for Python.
In
d
onesi
a
n
J
E
l
ec En
g &
C
o
m
p
Sci
ISS
N
:
2
5
0
2
-
47
52
A
n
e
w method
fo
r b
a
ll
tra
cki
ng
b
a
s
ed
on
α
-
β
, l
i
n
ea
r K
a
l
m
a
n
an
d
Ext
e
n
d
e
d
…
(
H
at
hi
r
a
m
N
e
nav
at
h)
99
7
(a)
(
b
)
(c
)
Fig
u
re
5
.
Obj
e
ct track
ing
resu
lt of th
e m
e
d
i
u
m
size b
a
ll: (a)
b
a
sed on
LKF (1
0,
1
5
,
35
an
d 50
fram
e
s),
(b)
base
d on
α
-
β
filter
(9, 1
2
, 3
4
a
n
d
52
fram
e
s) an
d
(c) b
a
sed
on
EKF (8
, 18
, 3
8
a
nd
4
8
frames)
3.
3.
7
Trackin
g
fr
ame results
of
big siz
e
ba
ll in no
rma
l
lig
ht
co
ndi
tio
n
The t
e
st
re
sul
t
s f
o
r
bi
g
si
ze
bal
l
u
nde
r
no
r
m
al
l
i
ght
con
d
i
t
i
on are
sh
o
w
n i
n
Fi
g
u
re
6
.
W
e
ha
ve
sel
ect
ed som
e
fram
e
s from
t
h
e vi
de
o
w
h
i
l
e
t
r
acki
ng t
h
e
ob
j
ect
cont
i
n
u
o
u
s
l
y
du
ri
n
g
i
t
s
m
ovem
e
nt
, w
h
e
r
e t
h
e
target object is
m
a
rked by a bou
ndi
ng
box (re
d circle). Fi
gure 6 shows
each four fram
e
s out of the tracking
r
e
su
lts b
a
sed
on
LK
F
i
n
Figu
re
6(
a)
,
α
-
β
filter in Figu
re 6(b
)
and
EKF in
Fig
u
re
6
(
c).
The values
of cova
riance
m
a
trices
and
p
l
ays an
im
p
o
r
tan
t
ro
le in th
e
esti
m
a
t
i
o
n
o
f
LKF
and
EKF p
a
ram
e
te
rs.
In
th
is p
a
per, t
h
e
v
a
lu
es
o
f
and
were d
e
term
in
ed
usin
g trial error wh
ich fo
rm
s th
e
delicacy of the
propose
d
m
e
thod. C
h
oosing incorrect values of Q a
n
d R
may
lead to incr
ease in RMSE and
decrease
in t
h
e
accuracy.
(a)
(
b
)
(c
)
Fig
u
re
6
.
Obj
e
ct track
ing
resu
lt of th
e b
i
g
si
ze b
a
ll: (a
)
b
a
sed
o
n
LKF
(15, 25
, 45
an
d 54fram
e
s), (b)
b
a
sed
o
n
α
-
β
filter
(16
,
2
8
,
4
0
a
nd
5
2
fra
m
e
s) and
(c)
b
a
sed
o
n
EKF
(18
,
26
, 48
and 53
fram
es)
3.
3.
8
Trackin
g
fr
ame results
of s
m
all siz
e
ball in b
a
d light
condition
The res
u
lts for s
m
all size ball unde
r ba
d light conditio
n are shown i
n
Figure
7. The target object is
m
a
rked
by
t
h
e
bo
u
ndi
ng
b
ox
(
r
ed ci
rcl
e
). Fi
g
u
re
7 sh
o
w
s 4
f
ram
e
s out
of t
h
e t
r
acki
n
g res
u
l
t
s
based o
n
L
K
F i
n
Figure 7(a),
α
-
β
filter in
Figure 7(b) and
EKF in
Figu
re
7
(
c).
In
fram
es 4
3
and
50
sh
own
in
Figure 7(a), th
e
target object tracked
using L
K
F is
d
o
m
i
nated by
t
h
e re
d b
o
u
n
d
i
n
g b
o
x
as
the cam
era is
static and the size of
t
h
e ob
ject
i
s
reduce
d
wi
t
h
di
st
ance. The sa
m
e
phen
o
m
e
na i
s
obser
ved i
n
fram
e
s 34, 48
and
45
, 5
0
sh
o
w
n i
n
Fi
gu
re 7(
b
)
a
n
d
Fi
gu
re 7(c
)
,
usi
n
g
α
-
β
filter and
EKF, respectiv
ely.
(a)
(
b
)
(c
)
Fi
gu
re
7.
O
b
je
ct
t
r
acki
n
g
res
u
l
t
o
f
t
h
e
sm
al
l si
ze bal
l
un
de
r
ba
d l
i
g
ht
c
o
n
d
i
t
i
on:
(a)
b
a
sed
o
n
L
K
F
(
2
5
,
3
5
,
4
3
,
an
d 50
f
r
a
m
e
s)
, (b
) b
a
sed on
α
-
β
filter (19
,
2
4
, 3
4
and
48
fram
es)
an
d (c)
b
a
sed
o
n
EKF
(26
,
36
, 45
, and
50 fram
es)
4.
CO
NCL
USI
O
N
In t
h
i
s
pape
r, a
new m
e
t
hod i
s
devel
ope
d f
o
r t
r
acki
ng
of
di
ffe
rent
si
zes o
f
bal
l
s
usi
n
g
α
-
β
, LKF a
n
d
EKF.
In
pr
op
o
s
ed t
r
acki
ng
fr
am
ewor
k, t
h
re
e di
ffe
rent
t
r
ac
ki
n
g
al
go
ri
t
h
m
s
were us
ed
to
track the target object
u
n
d
e
r th
e
n
o
rmal lig
h
t
co
nditio
n
fo
r
d
i
fferen
t sizes of b
a
l
l
s an
d
also
u
nder b
a
d
ligh
t
con
d
ition
s
for the sm
a
l
l
si
ze of t
h
e
bal
l
.
Ex
peri
m
e
nt
al resul
t
s
d
e
m
o
nstrate th
at LKF alg
o
r
ith
m
is
m
o
re effective and e
fficient than
α
-
β
and E
K
F al
gorith
m
s
for linea
r system
m
odel applications.
In sp
ite
o
f
th
e
reflection
s
from
th
e targ
et obj
ect
o
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
502
-47
52
I
ndo
n
e
sian
J Elec Eng
& Com
p
Sci, V
o
l. 10
,
No
.
3
,
Jun
e
2
018
:
98
9 – 99
9
99
8
th
e tiles o
f
th
e flo
o
r
wh
ich
is co
n
s
i
d
ered
as n
o
i
se, t
h
e LKF,
α
-
β
filter, an
d
EKF al
g
o
ri
th
m
s
are effectiv
e in
t
r
acki
n
g t
h
e o
b
ject
. F
o
r fut
u
re w
o
rk
, an
opt
i
m
i
zati
on m
e
t
hod co
ul
d
be consi
d
ere
d
i
n
det
e
rm
i
n
i
ng t
h
e
cova
riance
m
a
trices Q and R
,
to obtain a
n
optim
a
l
LKF a
n
d
EKF
pa
ram
e
ter
estim
a
tor m
o
re accurately.
REFERE
NC
ES
[1]
K. Cannons., “A review of v
i
sual tr
acking”, in Tech.
R
e
p.
CSE-20
08-07,
York Univ., Ontario
,
C
a
n
a
da, 2008.
[2]
A. Yilmaz, O. Javed,
and M. Shah
, “Object tr
acking: A survey
”, in
ACM Comput. Surv.,
vol. 38
, no. 4, pp. 1–45,
2006.
[3]
B. Babenko, M.-H. Yang, a
nd S. Belongie, “Robust object tr
acking with onli
n
e m
u
ltiple inst
ance l
earn
i
ng”,
in
IEEE
T
r
ans. Pat
t
ern Ana
l
.
Mach
.
Inte
ll.
, vo
l.
33
, no
. 7
,
pp
. 1619
–1632, Aug 201
1.
[4]
D. Com
a
niciu,
V. Ram
e
s
h
, and
P
.
M
eer
, “
K
ernel-based ob
jec
t
tracking
”
,
in
I
EEE T
r
ans. Pat
t
ern Anal. Ma
c
h
.
Intell
.
, vol. 25
, n
o
. 5
,
pp
. 564–57
7, May
2003.
[5]
S.
Ha
re
,
A.
Sa
ffa
r
i,
a
nd P.
H.
S.
T
o
rr,
Struc
k
:
“
S
tructured outpu
t tracking with kernels”,
in P
r
oc.
IEEE Int
.
Conf.
Comput. Vis., p
p
. 263–270
, 201
1.
[6]
X. Mei and H. Ling, “Robust visual
tracking an
d vehicle classif
i
cation via spars
e
representation
”
, in
IEEE Tran
s.
Pattern
Anal
. M
a
ch. In
te
ll.
, vo
l.
33, no
. 11
, pp
. 2
259–2272, Nov
2011.
[7]
D. Ross, J. Lim, R.-S. Lin and
M.-H.
Yang, “I
ncremental learn
i
ng for robust vi
sual tracking
”,
in Int. J. Comput.
Vis., vol. 77
, no
.
1, pp. 125–141,
2008.
[8]
X.
Li,
W.
Hu,
C.
Shen,
Z.
Zhang,
A.
Dick,
and A.
He
ngel, “A
survey
of appearan
ce models in visual object
track
ing”,
in
AC
M Trans. Intell.
Syst.
Technol.
, v
o
l. 4
,
no
. 4
,
pp
.
A: 1-A: 42
, 201
3.
[9]
B. D. Lucas
and
T. Kanade
,
“An iterative imag
e registration technique
with an
application to s
t
ereo vision”,
in
Proc. 7th
Int
.
Joi
n
t Conf.
Artif
. In
tell
.
, pp. 674–67
9, 1981
.
[10]
N.
Alt,
S.
Hinterst
oisser, and N. Navab,
“Rapid selection of relia
b
le
templa
tes fo
r visual tracking
”
, in P
r
oc. IEE
E
Conf. Comput.
Vis. Pattern R
e
cognit., pp
. 1355–
1362, 2010
.
[11]
S. Baker and I.
Matthews,
Lu
cas-Kanade 20
y
ears on:
“A unif
y
ing framework”,
in
Int.
J. Comput. Vis.
, vol. 56, n
o
.
3, pp
. 221–255
,
2004.
[12]
I. Matthews, T. Ishikawa,
and
S. Baker, “The
temp
late update problem”, in
I
EEE T
r
ans. Pat
t
ern Anal. Ma
c
h
.
Intell
.
, vol. 26
, n
o
. 6
,
pp
. 810–81
5, Jun 2004
.
[13]
M. J. Black
and
A. D. Jepson
, “Eigen
Track
ing:
Robust ma
tchin
g
and
tracking o
f
arti
culated objects using
a view-
based rep
r
esentation”, in
Int.
J. C
o
mput. Vis.
, vol.
26, no
. 1
,
pp
. 63
–84, 1998
.
[14]
G.
D.
Hager
and P.
N.
Belhumeur,
“Efficient region
track
i
ng with par
a
metric models of
geometr
y
and
illum
i
nat
i
on”
, in
IEEE Trans. Pa
ttern
Anal. Mach
. Intell.
, vol. 20
,
no. 10
, pp
. 1025
–1039, Oct 1998
.
[15]
G. Zhang, W. Zhu,
“Automatic video object
segmentation by in
tegrating object registration
and background
constructing technology”,
in Pr
oceed
ings of th
e
2006 Int
e
rnat
io
nal Conf
erenc
e
on Com
m
unicati
ons, Circu
its an
d
S
y
stems, pp. 43
7–441, 2006
.
[16]
W.-C. Hu, “R
eal-
time on-lin
e video object
segmenta
tion b
a
sed on motion detection without backgroun
d
construction”, in
Int. J. Inno
vat.
Comput.,
Inform. Control,
7
(4),
pp. 1845–1860
,
2011.
[17]
Jing Cheng, Sucheng Kang
,
“
R
obust Visual Track
ing
with
Im
proved Subspace Repr
esent
a
tion Mode
l”,
i
n
TE
LKOMNIKA
Te
le
c
o
mmunic
a
tion,
Com
puting, Electronics
and
Control,
Vol 15, no. 1, March
20
17.
[18]
Indah Agustien Siradjuddin
,
M. Rahmat
Wid
y
anto, T. B
a
sarud
d
in, “Particle
Filter with Gaussian Weighting fo
r
Hum
a
n Tracking
”
,
in Indonesian
Journal of Electrical E
ngin
eerin
g and Computer
Scien
ce (
I
JEECS
)
, Vol 10, no. 6,
Oct 2012.
[19]
D. Wang, H.
Lu,
and M.-H.
Yang, “
O
nline
objec
t tr
acking
with sparse pro
t
ot
ypes”
,
in
IEEE Trans. Ima
g
e
Proc
e
ss.
, vol. 22
, no
. 1
,
pp
. 314–
325, Jan
2013.
[20]
B. Liu, J. Huan
g, L. Yang, and
C. Kulikowsk,
“Robust trackin
g
using local sp
arse appearance model and K-
selec
tion”
,
in Pr
oc. I
E
EE Conf
.
Comput. Vis.
Pattern
Recogn
it.,
pp. 1313–1320
,
2011.
[21]
B. Ris
t
ic
and M
.
L.
Hernand
e
z
,
“
T
rack
ing s
y
s
t
em
s
”
, in
P
r
oc.
IE
EE
RA
DA
R
, Rome, I
t
aly
,
pp. 1-2,
2008.
[22]
W. Zhong, H. Lu, and M.-H
. Yang,
“Robust object
tracking via
sparse coll
aborative
app
ear
ance model”, in
IE
EE
Trans. Image Pr
ocess.
, vol. 23, n
o
. 5
,
pp
. 2356–2
368, May
2014.
[23]
J.
Fisc
us,
J.
Garofolo,
T.
Rose, a
nd M.
M
i
ch
el
, “
AVSS mul
tipl
e
camer
a
p
e
rson tracking
challenge
evalua
tion
over
vi
ew”
,
in
Pr
oc
. 6
t
h I
EEE A
V
SS, Genova, I
t
aly
,
2009.
[24]
C
.
B
.
S
a
n
t
i
a
g
o
,
A
.
S
o
u
s
a
,
M
.
L
.
E
s
t
r
i
g
a
,
L
.
P
.
R
e
i
s
,
and M. Lames, “Survey
on team
tracking tech
niques applied to
sports”, in
Proc. AIS, Povoa
de
Varzim
, Portugal,
pp. 1–6
, 2010
.
[25]
J
.
C. M
c
Call and M
.
M
.
Trivedi, “
V
ideo-bas
ed lane es
tim
ati
on and tracking for driver as
s
i
s
t
anc
e
:
S
u
rve
y
, s
y
s
t
em
,
and ev
alu
a
tion
”
,
in
IEEE Trans. Inte
ll. Transp. S
y
st
., vo
l. 7, no. 1,
pp. 20–37
, Mar
2006.
[26]
C. R. Wren, A. Azarbay
e
h
a
ni, T. Da
rrell, and A. P. Pentland,
Pfinder: “Real-time tracki
ng of the
human body
”, in
IEEE
T
r
ans. Pat
t
ern Ana
l
.
Machi
n
e Int
e
ll
., vol. 1
9
, no
. 7
,
pp
. 780
–785, Jul 1997
.
[27]
L. Maddalena and A. Petrosi
no, “A
self-organizing approach to bac
kground subtraction for visual surveillance
applications”, in
IEEE Trans. Image Process.
, vol. 17, no. 7, pp. 11
68–1177, Jul 20
08.
[28]
M. Oral and U.
Deniz, “Cent
r
e
of
m
a
ss
m
odel—A novel appr
oach to b
ackgro
und m
odelling f
o
r segm
entation
of
moving objects”,
in
Image Vis.
Comput
.,
vol. 2
5
, no
. 8
,
pp
. 136
5–1376, Aug 20
07.
[29]
Paul R. Kal
a
ta
,
“
T
he track
ing
i
ndex: A gen
e
ralized par
a
meter for
α
-
β
and
α
-
β
-
γ
target tr
ackers
"
, in
IEEE
Transactions on
Aerospace and
Electronic Systems
, AES Vol. 20
, no
. 2
,
pp
.174–1
81, mar 1984
.
Evaluation Warning : The document was created with Spire.PDF for Python.