TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 10, Octobe
r 20
14, pp. 7463
~ 747
0
DOI: 10.115
9
1
/telkomni
ka.
v
12i8.558
1
7463
Re
cei
v
ed
Jan
uary 5, 2014;
Re
vised July
19, 2014; Accepted Augu
st
15, 2014
A Tracking Algorithm of Moving Target in Sports Video
Zhang Hai Tao
Dep
a
rtment of Ph
y
s
ica
l
Educ
ation, Qufu Nor
m
al Univ
ersit
y
,
Yantai R
o
a
d
8
0
, Rizha
o Cit
y
Shan
do
ng Pro
v
ince
email: zh
ang
ht
_20
14@
16
3.co
m
A
b
st
r
a
ct
Moving target tracking is
a co
re
sub
j
ect
in
the fi
el
d
of co
mp
uter v
i
sion,
its cor
e
id
ea
i
s
compre
hens
ive
utili
z
a
t
i
on
of i
m
a
ge
process
i
ng, vid
eo
ana
l
ysis techn
o
lo
g
y
, quickly
an
d
accurate
ly ca
pture
the
movi
ng tar
gets. In ord
e
r
to obtai
n a
ll k
i
nds of te
c
hnic
a
l p
a
ra
met
e
rs
in the
train
i
ng
of athl
etes, th
e
traditio
nal w
a
y is throug
h vari
ous sens
ors us
ed in the
ath
l
et
es. T
he shortcomin
gs of this meth
od is n
o
t put
the i
n
flue
nce
o
f
sensor t
e
chn
i
ques
for ath
l
et
es in
co
ns
id
er
ation, s
o
it is
urge
nt to h
a
ve
a n
e
w
techn
i
c
a
l
me
ans t
o
g
e
t
better
mov
e
ment p
a
ra
meter
s
of the
athl
etes, the tra
d
itio
nal t
a
r
get trac
king
al
gorith
m
i
s
difficult to ac
hi
eve g
ood r
e
sul
t
s, in this stud
y, bas
ed
on th
e char
acteristi
cs of
sports vi
deo, a
nd it
ma
kes
some transfor
m
to the traditi
ona
l tracking a
l
gorit
hm, a
nd it propos
es a ne
w
hybrid tracking al
gorith
m
. T
h
i
s
alg
o
rith
m not o
n
ly can solv
e the difficu
lt pro
b
le
ms of
sport
s
video targ
et trackin
g
, and b
u
t also can re
d
u
ce
the co
mp
lex
i
ty of the a
l
gor
ith
m
, thro
ugh
the
relate
d a
nalys
i
s
and
exp
e
ri
ments in
sports v
i
de
o softw
are, th
e
alg
o
rith
m can
achi
eve g
ood e
ffect.
Ke
y
w
ords
: tracking a
l
gor
ith
m
, movin
g
target
, sports video, researc
h
Co
p
y
rig
h
t
©
2014 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
In the spo
r
ts video, moving sp
eed of
target
is fa
ster, and relat
ed ba
ckgro
u
nd is very
compl
e
x, in
gene
rally, it i
s
ve
ry difficul
t
to obtai
n g
ood t
r
a
cki
ng
effect throug
h commo
n ta
rge
t
tracking
algo
rithm. So in this pa
per, it prop
os
es a
n
e
w glo
bal mo
del tra
cki
ng
algorith
m
, wh
ich
can
solve
rel
a
ted p
r
oble
m
s. Firstly, it adopts th
e
pa
rticle filter al
g
o
rithm to
pre
d
ict the m
o
ving
targets at th
e begi
nning,
throug
h si
mil
a
rity com
pari
s
on
of the p
r
edicte
d
po
siti
on of the ta
rge
t
with the targ
et model, when the
simil
a
rity in
less
than a value
,
it can be t
houg
ht that the
fundame
n
tal
cha
nge
s hav
e taken
place in the ta
rg
et movement
model, it needs to u
s
e a
new
motion model
, when the si
milarity is gre
a
ter than a
certain value, it can be
thought that there is
no big
ch
ang
es in
the ta
rg
et motion m
o
del, it doe
s n
o
t need
to u
s
e ne
w motio
n
model. An
d then
use the m
e
a
n
shift algorit
hm based on
nucle
ar
itera
t
ion on the p
o
sition p
r
edi
ction and thu
s
it
c
an obtain the acc
u
rate target loc
a
tion.
At pre
s
ent, th
is alg
o
rithm
h
a
s
been
in th
e prac
ti
cal
ap
plicatio
n of ef
fective in
spe
c
tion, it
has
re
ceived
the goo
d effe
ct. The p
u
rp
o
s
e of
spo
r
ts
v
i
deo target tracking i
s
mai
n
ly to obtain t
he
track of the
moving targ
e
t
s and motio
n
informatio
n
,
at present in this pa
per,
the algo
rithm
has
been
widely u
s
ed in o
u
r sp
orts trai
ning v
i
deo an
alysis
system.
The gl
obal
m
o
vement i
s
mainly cau
s
e
d
by the
ca
mera
movem
ent in g
ene
ra
lly. If the
came
ra i
s
in the pro
c
e
ss of move
ment, the object
s
in the
images al
so has thei
r own
movement
s, then the ba
ckgro
und a
n
d
foreg
r
ou
nd
in the video seq
uen
ce
has thei
r o
w
n
movement
s, the ba
ckgro
u
nd of the mo
vement is
ca
use
d
by the
came
ra m
o
vement, whi
c
h
is
kno
w
n a
s
th
e global mov
e
ment. The
purp
o
se of global motion
estimation i
s
to find out the
came
ra move
ment law fro
m
the video seque
nce ca
u
s
ed by the gl
obal movem
e
nt.
Global
motio
n
e
s
timation
can
be
u
s
ed
i
n
the vid
eo
m
o
tion o
b
je
ct segmentatio
n
su
ch
as
pano
ram
a
ge
neratio
n a
nd
Psrti
codi
ng fi
eld, etc.
In
th
e video
motio
n
obj
ect
seg
m
entation, gl
obal
motion e
s
tim
a
tion i
s
the fi
rst
step, a
nd
then to m
a
ke
estimatio
n
of
the
came
ra
motion b
e
twe
en
frame
s
a
nd t
hen
com
pen
sate the f
r
ame
alignm
ent
be
tween
the
ba
ckgro
und,
an
d then
a
c
cord
ing
to the move
ment, it can
sep
a
rate th
e foreg
r
o
und
obje
c
ts an
d
backg
rou
nd.
The pa
nora
m
a
gene
ration i
s
obtained th
rough e
s
timat
i
on cal
c
ul
atio
n of global
motion bet
ween fram
es
and
pixels, then a
c
cordi
ng to th
e movement
para
m
et
ers a
nd thro
ugh joi
n
ing togeth
e
r of the adja
c
e
n
t
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 10, Octobe
r 2014: 746
3
– 7470
7464
frame
s
pan
orama imag
e can be obtai
n
ed. The Ps
rti code of MP
EG4 is the u
s
ing of pa
norama
for pre
d
ictio
n
and compe
n
sation, thus it c
an g
r
eatly improve the
compressio
n ratio.
Therefore, n
o
matter for th
e analysi
s
of l
a
w of
camera
motion dire
ct
ly, or analysi
s
of th
e
foreg
r
ou
nd
o
b
ject m
o
tion,
the gl
obal
m
o
tion e
s
timati
on i
s
the
ba
sis of th
ese a
nalyse
s
. Gl
o
b
a
l
motion pa
ra
meters estim
a
tion method
con
s
ist
s
of differential met
hod an
d the method of fe
ature
point corre
s
p
onde
nce, the
y
respe
c
tively according t
o
the velo
cit
y
field or
on
the imag
e pi
xel
feature poi
nts to calcul
ate the relatio
n
be
tween the gl
o
bal motion pa
ramete
rs.
In this
study, it use
s
the
six pa
ramete
rs m
odel
in
the mod
e
ling
of interfa
c
e
scene
cha
nge
s
cau
s
ed
by came
ra motion, a
n
d
ado
pts
differential m
e
tho
d
in
solving t
he glo
bal m
o
tion
para
m
eters. Whe
n
the ch
ange
s of rela
tive depth
of
obje
c
ts in the
sce
ne and t
he cam
e
ra
zoom
rang
e i
s
not l
a
rge,
the
six
para
m
eters o
f
the mo
del
can d
e
scribe
the
rotation
of
the
cam
e
ra, l
ens
and di
spla
ce
ment movem
ent well. Be
cau
s
e in
the
acqui
sition
of video fra
m
e, the adja
c
ent
frame
s
can
meet the a
b
o
ve co
ndition
s, so thi
s
m
odel
can
rea
s
on
ably de
scribe the
ca
m
e
ra
motion betwe
en adja
c
e
n
t frames.
2. Camera M
odel and
Ca
m
e
ra Calibra
tion
The glo
bal m
o
tion cau
s
ed
by came
ra m
o
vement on t
he ba
ckgroun
d ca
n be exp
r
esse
d
with 6 param
eters m
o
tion
model ju
st sh
own a
s
(1
):
,,
,,
ii
i
ii
i
x
ax
b
y
e
yc
x
d
y
f
(1)
Whe
r
e
，
()
ii
px
y
the
curre
n
t coord
i
nates of
k
I
,
,,
,
()
ii
px
y
is the a
d
jacent f
r
ame
s
,
they are
co
rresp
ondi
ng p
o
i
nts of
,
k
I
and P,
,
,
(,
,
,
,
,
)
kk
ab
c
d
e
f
is the gl
obal
p
a
ram
e
ters
，
,,
,
ab
c
d
rep
r
e
s
ent rot
a
tion
a
nd sca
ling,
re
sp
ecti
vely,
,
ef
represe
n
t the di
spla
cement. If in t
he formula i
s
corre
c
t, there
shoul
d be
,
,
()
(
)
k
k
Ip
I
p
,
so the o
b
jecti
v
e function. i
s
,,
,
,
()
(
)
(
)
k
kk
k
p
RI
p
I
p
.
Global m
o
tio
n
estimatio
n
is to solve u
n
determi
ned p
a
ram
e
ters
,
,
kk
and get the sm
a
llest value of
,
,
()
kk
R
. We
ca
n u
s
e
Gau
s
s - Ne
wton
or
Leve
ngerg-
Ma
rqu
a
rdet
nonli
n
e
a
r ite
r
ation
al
gorithm.
The ba
sic p
r
i
n
cipl
e of Gau
ss
- Ne
wton it
erative metho
d
is:
Make
the
a
s
sumptio
n
s th
at for
param
eters
,
,
k
tt
A
, a
nd
make
the
Ta
ylor expa
nsi
o
n to
obje
c
tive function R at cu
rrent poi
nt, the type (2) can b
e
obtaine
d.
,,
,
,
,
2
,,
,
,
,
1
(
)
(
)
()
()
(
)
2
Tk
k
T
k
kk
tt
tt
t
t
t
t
tt
RA
RA
g
A
A
H
A
(2)
Whe
r
e
k
g
an
d
k
H
rep
r
e
s
e
n
ts
the gradie
n
t and
Hessia
n matrix
of
,
,
()
tt
RA
at
,
,
k
tt
A
r
e
spec
tively.
T
kk
k
g
JW
(3)
T
kk
k
i
i
i
k
i
H
JW
J
H
(4)
k
rep
r
e
s
ent th
e re
sidu
als o
f
,
,
k
tt
A
,
,
,
tt
J
kA
,
W
is a dia
g
onal mat
r
ix and
ii
Wi
,
ik
H
is
the Hess
ian matrix of
i
，
if the value of
is sm
all
，
we can think
that:
T
kJ
k
H
Jk
W
(5)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Tracking Al
gorithm
of Moving Ta
rg
et in Sports Video
(Zhan
g Hai T
ao)
7465
,,
,,
()
tt
t
t
RA
A
(6)
And the in the further
we can get:
,
,
()
Tk
T
kk
k
k
tt
JW
J
A
JW
(7)
And we
ca
n
get the re
sult
ing in
cre
m
ent
al
,
,
k
tt
A
of
,
,
k
tt
A
and th
en we can find out the
next param
eter.
,,
,
1
,,
,
kk
k
t
t
tt
tt
A
AA
(8)
Thro
ugh
such iteration, it
ca
n g
r
a
duall
y
re
d
u
ce
the
obje
c
tive fu
nction, th
us they can
obtain the be
st estimatio
n
of the globa
l motion
para
m
eters , the throug
hout iterative proce
ss
adopt
s the level of the three
-
tier pyramid struct
u
r
e in orde
r to improve t
he efficien
cy
of
cal
c
ulatio
n, the pyramid
with [1/4, 1/2, 1/4] filter.
The
came
ra model
i
s
the simplification and
a
pproximation of opt
ical ima
g
ing
geomet
ry
relationshi
p
. Camera m
o
del usually consist
s
of
a set of para
m
eters, these pa
ra
meters are ca
lled
Came
ra
pa
rameters, a
n
d
the
solvin
g process i
s
called
ca
mera
calibra
tion of cam
e
ra
para
m
eters.
Pinhole
Mod
e
l is a
ki
nd
of ideal
M
o
d
e
l of th
e
ca
mera. It
de
scribe
s
as a
central
perspe
c
tive p
r
oje
c
tion ima
g
ing p
r
ocess.
Persp
e
ct
ive
proje
c
tion h
a
s
the c
haract
e
risti
cs of
sm
all
depe
nding
almost zero. P
a
rallel
strai
g
h
t
lines in the
spa
c
e
will intersect in the
proje
c
ted im
a
ge;
intersectio
n
p
o
int is
called
as the
dire
cti
on of
the van
i
shin
g point. In addition, in
the pro
c
e
s
s
of
projection poi
n
t of in the line will keep invariance ratio.
Figure 1. The
World
Coo
r
di
nates of the
Point
Proje
c
tion to the Image Pro
c
e
s
s
Figure 2. Pinhole Came
ra
Model
Figure 1 pre
s
ent
s the si
mulation proj
ection
p
r
o
c
e
s
s of the cam
e
ra thro
ugh
comp
uter
grap
hics. We
call the scen
e coo
r
din
a
te system a
s
th
e cam
e
ra
co
ordin
a
te syst
em, and take
the
came
ra
a
s
th
e center,
an
d coordinate
system
e
s
tab
lishe
d a
c
cord
ing to th
e o
r
i
entation
of the
came
ra i
s
cal
l
ed ca
mera coordi
nate. An
d while
th
e image
coo
r
di
nate syste
m
is esta
blished
in
two dim
e
n
s
io
nal imag
e
co
ordin
a
te
syst
em. The
co
ordinate
syste
m
and th
e co
nventional
ca
mera
coo
r
din
a
te a
r
e alig
nment.
Figure 2
is a
pinh
ole
ca
m
e
ra
mod
e
l di
agra
m
. Fig
u
re 4
presents
out
comp
uter graphi
cs sim
u
l
a
tion of
cam
e
ra projectio
n
process.
T
he scene co
ordin
a
te syst
em
kno
w
n
as the
cam
e
ra
coordinate
sy
ste
m
, establi
s
h
e
d
in a
c
corda
n
ce
with th
e
orientatio
n of
the
came
ra co
ordinate syste
m
is calle
d camera
co
o
r
di
nates
and
th
e imag
e coordinate
syste
m
is
establi
s
h
ed in
2-D ima
ge coordi
nate sy
stem, Figure 2
diagra
m
is a
pinhol
e cam
e
ra mod
e
l.
If we
kno
w
i
n
the
scene
point
w
X
to th
e world
coordinate
syste
m
of coo
r
din
a
tes
(,
,
)
T
ww
w
XY
Z
, calculate th
e point p
r
oje
c
tion also nee
d to
kn
ow th
e
relative po
siti
on of the
cam
e
ra
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 10, Octobe
r 2014: 746
3
– 7470
7466
coo
r
din
a
te sy
stem a
nd
world co
ordi
nate
system
(cam
era po
sition). The
relative
p
o
sition ca
n
u
s
e
a rotation m
a
trix R and a t
r
an
slation ve
ctor t to express
w
X
.
The
proj
ection coo
r
di
nates can use
the formula b
e
low:
,0
,
0
,0
0,
,
0
,
0
1
10
,
0
,
1
,
0
1
c
c
c
t
c
X
xf
Rt
Y
zy
f
Z
o
(9)
So the whole
proje
c
tion proce
s
s is dete
r
mine
d by the proje
c
tion
3×4 mat
r
ix of M, and
matrix M ca
n be de
com
p
o
s
ed into came
ra external
pa
ramete
rs
(,
)
ex
M
Rt
whi
c
h i
s
rel
a
ted to the
came
ra
po
siti
on a
nd
00
(,
,
,
)
mx
y
M
ff
u
v
only
related
with
the
came
ra
to
the inte
rnal
stru
cture of
came
ra
intri
n
sic
pa
ram
e
ter
s
. C
a
me
ra
c
a
libration
is in
orde
r to
dete
r
mine
the
pa
rameter matri
x
,
su
ch a
s
the rotation matrix
R and rel
a
te
d para
m
eters of the displa
ceme
nt vecto
r
t.
Above the ca
mera
calib
rati
on algo
rithm
is not o
n
ly very co
mplexit
y
, but also n
eed
s to
solve the
ca
mera i
n
trin
sic para
m
et
ers and
external para
m
et
ers,
more
over a
s
a re
sult of th
e real
came
ra e
s
pe
cially wide a
ngle len
s
ca
mera
w
ill have a certai
n amount of dist
ortion compa
r
e
d
with the pi
nh
ole came
ra
model. So yo
u nee
d to do
to make
co
rrespon
ding
calibratio
n
of the
image. Th
ere
f
ore, in thi
s
p
aper, fo
r con
v
enient, it
do
n't use the o
p
timization te
chniqu
e, and
a
l
so
don't introdu
ce nonlin
ea
r p
r
og
ram, an
d
only adopt lin
ear e
quatio
n
(group
) an
d
matrix cal
c
ul
a
t
io
n
method. It ca
n make the a
l
gorithm m
o
re quickly,
an
d the followin
g
cam
e
ra
cali
bration
algo
rithm
adopte
d
in this pap
er is a
s
follows.
First,
we
sele
ct fou
r
featu
r
e poi
nts i
n
th
e ima
ge
and
on the
field
m
odel
re
spe
c
ti
vely (any
of the three feature points
not collin
ear),
just as shown in Figure 3:
Figure 3. Fea
t
ure Points o
n
the Model a
nd Image
In the Figure 3, we set
the coordi
n
a
tes of the four points on the image are
(,
)
(
0
,
1
,
2
,
3
)
ii
xy
i
, and in the field co
ordi
nat
es of the four
points mo
del
are:
,
11
12
13
,
21
22
23
31
32
33
(0
,
1
,
2
,
3
)
11
i
i
ii
x
aa
a
x
ya
a
a
y
i
aa
a
(10)
In the proce
ss of came
ra
calibration, we can u
s
e a set of field and image
estimate
equatio
n of correspon
ding
points a
ji
. While in p
r
a
c
tice, there
nee
d
s
some
re
gul
arization m
e
thod
to re
solve
the
non
uniq
uene
ss p
r
oble
m
of
type pa
ra
m
e
ter in
(10).
Usually ma
ke
el
ement a
33
=
1,
of course, it a
l
so can u
s
e o
t
her form
s of regul
ari
z
ation
:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Tracking Al
gorithm
of Moving Ta
rg
et in Sports Video
(Zhan
g Hai T
ao)
7467
From type (1
0), we
can ge
t:
,
11
12
13
31
32
33
,
21
22
23
31
32
33
i
i
ax
a
y
a
x
ax
a
x
a
ax
a
y
a
y
ax
a
x
a
(11)
Thro
ugh p
r
o
per al
geb
rai
c
pro
c
e
ssi
ng, let the a33
=l, throug
h ea
ch pair of ima
ge and
corre
s
p
ondin
g
point on th
e grou
nd mo
del, the of linear e
quatio
ns contain
s
two pa
ram
e
te
rs,
expre
s
sed in
matrix form is as belo
w
.
11
12
13
,,
,
21
,,
,
,
22
23
31
32
,
,
1
,
0
,
0,
0,
(
)
,
(
)
0,
0
,
0,
,
,
1
,
(
)
,
(
)
i
i
ii
ii
i
ii
i
i
i
i
i
a
a
a
x
yx
x
y
x
a
x
a
x
yy
x
y
y
y
a
a
a
(12)
3. Sports Video Multi-targ
et Trajec
tory
Tracking Al
gorithm
Multiple targ
e
t
tracki
ng i
s
the key of the curre
n
t re
se
arch of comp
uter visio
n
, e
s
pe
cially
in the hum
a
n
body tra
cking field, it is a h
o
t re
se
arch poi
nt. The multiple t
a
rget
s tra
c
ki
ng
algorith
m
is
roughly divid
e
d
into two
categori
e
s:
the
first o
ne i
s
ba
sed
on the
m
odel of m
u
ltip
le
targets tracki
ng syste
m
; and the se
co
n
d
one is b
a
se
d on the mult
i-so
urce info
rmation fusio
n
o
f
mult
iple t
a
rg
et
s t
r
ac
kin
g
sy
st
em
. M
u
ltiple targets tracking
alg
o
ri
thm which
is ba
sed
on
the
model, an
d it is mainly use the relatio
n
s
hip
s
bet
wee
n
the tracke
d
targets, the
movement of
the
multi-obj
ectiv
e
model i
s
e
s
tabli
s
he
d. And
then reu
s
e the co
rre
spondi
ng stat
e spa
c
e
sea
r
ch
target tra
c
kin
g
algo
rithm, i
t
can
achieve
the targ
et m
o
tion, the alg
o
rithm i
s
mai
n
ly use
d
in t
he
human b
ody tracking. Th
e
multi-sou
r
ce informati
o
n
fusion alg
o
rit
h
m of target tracking i
s
often
throug
h multi
p
le se
nsor to
reali
z
e the inf
o
rmat
io
n fusi
on. And then
throug
h u
s
ing
neural netwo
rk
and hi
dden
Markov mod
e
l
to reali
z
e inf
o
rmatio
n fusi
on.
This
kin
d
of algorith
m
i
s
mainly u
s
e
d
in
rada
r si
gnal p
r
ocessin
g
, etc.
In the study, the motion e
s
timation algo
ri
thm
base
d
on
particle filter
algorith
m
and
video
came
ra
s i
s
a
dopted i
n
the
analysi
s
of fo
otball, ho
ckey video. The
ai
m is to
track athletes ru
nni
ng
route
so
as t
o
obtain th
e
movement e
s
timation of
th
e of athlete
s
and
spo
r
ts inf
o
rmatio
n such as
spe
ed, whi
c
h
can p
r
ovide t
he help for
co
ach in ta
ctical
analysi
s
.
As the tra
d
itional m
u
ltiple
target tracki
ng
alg
o
rithm,
it is often
b
a
se
d on
the
static
backg
rou
nd, and only ca
n
get
the speed
an
d
traj
ecto
ry if th
e target
whi
c
h
ha
s
relati
ve
movement to
the came
ra,
it can't
get real move
m
e
nt informatio
n of targ
et. so the alg
o
rith
ms
can
not provid
e useful info
rmation for the
coa
c
h.
In orde
r to get the motion estimation of
the pl
ayers
on the pitch,
the first step
is to get
the traje
c
tory
of players i
n
the video,
and the
n
reuse the tra
j
ectori
es of
came
ra
moti
on
estimation
algorithm fo
r camera, throu
gh the video
image
spa
c
e
coo
r
din
a
tes
with the map
p
ing
relation
shi
p
b
e
twee
n
real
space
coo
r
din
a
tes, fina
lly
g
e
t the m
o
tion
of pl
ayers o
n
the
pitch.
T
h
e
algorith
m
pro
c
e
ss al
gorith
m
is as follo
ws:
1. Throug
h u
s
ing
of
came
ra
calib
ration
algo
rithm, it
ca
n g
e
t the
mappin
g
rela
tionshi
p
betwe
en the
model an
d the first video image
s:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 10, Octobe
r 2014: 746
3
– 7470
7468
,
11
12
13
31
32
33
,
21
22
23
31
32
33
ax
a
y
a
x
ax
a
x
a
ax
a
y
a
y
ax
a
x
a
(13)
Whe
r
e the p
o
int
,,
(,
)
x
y
is the
co
ordin
a
tes
of the field mo
del, the poin
t
x (y) is the
coo
r
din
a
tes o
f
video image
in first frame.
2. Throu
gh h
y
brid pa
rticle
filter and the mean
shift tracki
ng algo
rith
m, the coordi
nates of
the player on
the curre
n
t video fram
e
,,
,,
(,
)
kk
x
y
ca
n be obtain
e
d
.
3. Ma
ke th
e
assumptio
n
t
hat
(,
)
T
ii
x
y
is a
pix
e
l po
sition
in
the
cu
rre
nt f
r
ame
imag
e,
(,
)
T
ii
x
y
is the point of
the image lo
cation, the rel
a
tionship bet
wee
n
them can be expressed a
s
:
,
02
3
,
14
5
()
()
ii
i
ii
i
x
aa
x
a
y
ya
a
x
a
y
(14)
,
,
ii
i
i
xx
A
T
y
y
(15)
Among the
m
,
23
45
aa
A
aa
represe
n
t zoom, ro
tate and st
retchin
g
,
01
(,
)
T
Ta
a
rep
r
e
s
ent tra
n
slatio
nal mo
tion. And thro
ugh ap
plyi
ng
the global m
o
tion estimatio
n
algo
rithm, we
can o
b
tain th
e came
ra mot
i
on paramete
r
s
23
14
5
,,
,,
ko
k
k
kk
k
aa
a
aa
a
.
4. Make the
solutio
n
of
,,
,
,
(,
)
kk
x
y
coordi
nate
of tracke
d target
point o
n
the
curre
n
t frame
,
whi
c
h is correspon
ding
to the first frame image
co
ordin
a
tes
(,
)
kk
x
y
.
02
3
,,
,
,
1
14
5
1
,,
(,
)
,,
k
jj
j
T
kk
k
j
jj
j
k
aa
a
x
yx
aa
a
y
(16)
,,
,
,
1
(,
)
T
kk
k
k
x
yA
x
y
(17)
Whe
r
e,
02
3
0,
2
3
1
14
5
14
5
,,
,
,,
,,
k
jj
j
j
jj
j
aa
a
aa
a
A
aa
a
aa
a
(18)
02
3
14
5
,,
,,
jj
j
jj
j
aa
a
aa
a
is the global
motion estim
a
tion paramet
ers of the vid
eo from fram
e j-1 to frame
j .
,,
,,
51
3
0
5
3
25
43
,,
,
,
20
4
1
2
4
25
43
()
(
)
()
()
(
)
()
kk
k
kk
k
ax
a
a
a
a
a
y
x
aa
aa
ay
a
a
a
a
a
x
y
aa
a
a
(19)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Tracking Al
gorithm
of Moving Ta
rg
et in Sports Video
(Zhan
g Hai T
ao)
7469
5. Throu
gh the mappi
ng relation
shi
p
bet
we
en the grou
nd mod
e
l
and the first frame
video ima
ge,
we
can
obt
ain the
coor
dinate
s
of t
r
acked
targ
et point
s
,,
,
,
(,
)
kk
x
y
, co
rresp
ondi
ng
coo
r
din
a
tes o
f
the point on the model
,,
,
,
(,
)
mm
x
y
,,
11
12
13
3
1
32
33
,,
21
22
23
31
32
33
kk
m
kk
kk
m
kk
ax
a
y
a
x
ax
a
y
a
ax
a
y
a
y
ax
a
y
a
(20)
Below is an
experim
ental
result of B
r
a
z
il vs
En
glan
d
200
2 Wo
rld Cup
football
game.
T
hey are
1293
2 to 132
34 frame
s
of the match vid
eo.
Figure 4. The
Image of the Football Mat
c
h
Figure 5. The
Traje
c
tory Tracking
Re
sult
s
Figure 6. Th
e Image of the Football Ma
tch
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 10, Octobe
r 2014: 746
3
– 7470
7470
Figure 7. The Track
i
ng Res
u
lts
of Red
Team
4. Conclusio
n
The algo
rith
m combin
s p
a
rticle filter a
nd global mo
tion estimatio
n
algorithm,
and it can
solve the
problem of m
u
ltiple targ
et trac
king al
gorithm whi
c
h
only
can receive
lo
cat
i
on
coo
r
din
a
tes relative to the
came
ra
coordinate
sy
ste
m
, and
the
shortcomin
g of
una
ble to
ge
t the
real motio
n
p
o
sition al
so
can be waked.
The arti
cle al
so is aim at t
he player p
o
s
it
ion overl
a
p
p
ing ph
enom
enon in multi
p
le target
tracking, the
Bayesian
cla
ssifie
r
is add
e
d
, and thr
oug
h using the ta
rget moveme
nt information
,
it
can p
r
elimin
ary solve the
athletes ove
r
lappi
ng ca
se, the trackin
g
probl
em of
athletes. From
Figures 4-7, it
can be se
e
tha
t, the a
l
gorithm
prop
ose
d
in thi
s
pape
r, the
camera an
gle
is
changed, will
still be abl
e to accuratel
y
track ta
rget position on
the pitch,
so
as
to solve the
multiple targe
t
tracking
alg
o
rithm b
e
fore
only ca
n tra
ck
moving ta
rget rel
a
tive to the po
sition
of
the video
an
d ca
n't g
e
t the a
c
tual l
o
cation. The
ca
mera
motion
estimatio
n
a
l
gorithm
ca
n
be
applie
d to the algorith
m
in target tra
c
king, thus
ma
ke the multiple
target tra
cki
ng algo
rithm
can
be appli
ed to
the dynamic backg
rou
nd
and solve t
he dynami
c
b
a
ckgroun
d, target p
o
sitio
n
ing
probl
em at sa
me time.
Ackn
o
w
l
e
dg
ements
The
re
sea
r
ch is su
ppo
rt
ed by Soft
sci
en
ce
re
se
arch p
r
oje
c
t
of Shand
ong
provin
ce
gene
ral p
r
oje
c
t (201
3 RKB
0102
5)
Referen
ces
[1]
C W
r
en, Az
arb
a
y
ej
ani, T
Darr
ell, A
Pentl
a
n
d
,
Pfinder:
Rea
l
-time T
r
acking
of the
Hum
a
n
B
od
y.
IEEE
Trans. PAMI,
1
997; 19(
7): 780
-785.
[2]
T
Olson, F Brill.
Movi
ng Ob
j
e
ct Detecti
on
and Ev
ent R
e
cogn
ition
Alg
o
r
i
thms
for S
m
a
r
t Ca
meras
.
Proc.DARPA Image U
nderst
and
ing W
o
rksh
op. 199
7.
[3]
T
N
T
an, GD Sulliv
an, KD B
a
ker. Mode
l bas
ed l
o
cal
i
zatio
n
and r
e
cog
n
iti
o
n of roa
d
ve
hi
cles
. Int. J.
Com
p
ut. Vis
. 1998; 29(
1): 22-
25.
[4]
RT
Collins, AJ Lipto
n
, T Kanade.
A Syste
m
for Video S
u
r
v
eill
anc
ean
d M
onitor
i
ng
. Proc.Am.Nuclear
Soc. (ANS) English Int. T
opical Meet
i
ngR
ob
otic and R
e
mot
e
S
y
stems.1
9
9
9
[5]
MJ S
w
a
i
n, DH
Ball
ard. Col
o
r i
nde
xing,
Intern
ation
a
l Jo
urna
l of Computer Vi
sion
. 19
91; 11-
32.
[6]
R Kie
l
ds
en, JK
end
er.
Fi
n
d
i
ng ski
n i
n
co
lo
r im
ag
e
s
.
2n
d i
n
ternati
ona
l c
onf
erenc
e o
n
A
u
tomatic F
a
c
e
and Gesture R
e
cog
n
itio
n, K
ill
i
ngton,Verm
ont
, USA. 1996.
[7]
P F
i
eguth, D
T
e
rzopoulos.
Color-
bas
ed T
r
ackin
g
of he
ad
s and
other M
obil
e
o
b
jects
a
t
video fra
m
e
rates.
Proc.Of I
EEE CVPR. 1997: 21-27.
[8]
M Hunke, A W
a
ibel.
F
a
ce l
o
catio
n
an
d trackin
g
for hu
man-co
mputer i
n
teractio
n.
Pro
c
of the 28t
h
Asilomar C
onf.
ON Signals, S
y
s. an
d Comp.,
1994: 1
277-
12
81.
[9]
K Sobottka, I Pitas.
Seg
m
e
n
tation
an
d tr
ackin
g
of fac
e
s in C
o
l
o
r Images.
Pr
oc of
the Sec
o
n
d
Intl.Conf.On Auto
Face and Gesture Rec
o
gn
ition. 1996: 236-241.
[10]
Y Bar-Shal
om, T
F
o
rtmann.
T
r
ackin
g
an
d Dat
a
Associati
on. Academ
ic Pres
s. 1998.
[11]
S Julier, J U
h
l
m
ann.
A New
Extensio
n of the Ka
lman F
ilt
er to Nol
i
ne
ar
Systems.
Proc
. SPIE. 3068.
199
7; 182-
193.
[12]
Micha
e
l Isar
d, Andre
w
Bl
ak
e Co
nd
ensati
o
n-Co
nditi
on
al
Densit
y Pr
op
a
gatio
n for V
i
s
ual T
r
ackin
g
.
Internatio
na
l Joun
al of Co
mp
uter Visio
n
.
19
98; 29(1): 5-2
8
.
Evaluation Warning : The document was created with Spire.PDF for Python.