Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
Vol.
4, No. 6, Decem
ber
2014, pp. 979~
988
I
S
SN
: 208
8-8
7
0
8
9
79
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Goal Detection in Soccer Video:
Ro
le
-Ba
s
ed Ev
e
n
ts Detec
t
ion
Approach
Fa
rsha
d Bayat
*,
M
o
ha
mmad Sha
h
ra
m
Mo
in*, Fa
rha
d
Ba
ya
t**
* Depart
em
ent o
f
El
ectr
i
c
a
l
and
Com
puter Engin
eering
,
Is
l
a
m
i
c
Azad
Unive
r
s
i
t
y
, Qa
zvin Br
anch
,
Qazvin
, Ir
an
** Depart
em
ent
of El
ectr
i
c
a
l
Eng
i
neer
ing,
Univer
s
i
t
y
of
Zanj
an,
Z
a
njan
, Ir
an.
Article Info
A
B
STRAC
T
Article histo
r
y:
Received J
u
n
7, 2014
R
e
vi
sed Oct
9,
2
0
1
4
Accepted Oct 25, 2014
S
o
ccer video p
r
oces
s
i
ng and anal
ys
is
to fin
d
criti
cal
event
s
s
u
ch as
occurren
ces
of g
o
al ev
ent hav
e
b
een one
of
the
important issues and topics of
act
ive res
e
arche
s
in recent
years
.
In this
paper, a
new role-bas
ed fram
e
work
is proposed for g
o
al even
t dete
cti
on in which the sem
a
ntic structu
r
e of soccer
game is used. Usually
after a go
al scen
e, the
aud
i
ences’ and repo
rters’ sound
intensit
y
is incr
eased, b
a
ll
is
s
e
nt back to
the c
e
nter
and the
ca
m
e
ra m
a
y:
zoom on
Play
er, show audiences
’ delightin
g
,
rep
eat the goa
l s
cen
e or dis
p
la
y
a com
b
inat
ion o
f
them
. Thus
,
th
e occurr
enc
e
of
goal ev
ent wil
l
b
e
det
ect
abl
e
b
y
analy
s
is of s
e
quences of
abo
v
e ro
les.
The pr
oposed framework in
this
paper
consists of four main procedures:
1- detection of game
’s
cri
tic
al
event
s
b
y
using aud
i
o channel, 2-
dete
ction of
shot boundary
and shots
classification, 3-
selection of
can
d
idat
e ev
ents according to
the ty
pe of shot
and existence of
goalmouth in the shot,
4- detection of r
e
starting the game
from the center
of the field
.
A new me
thod fo
r shot classification is als
o
presented in this
framework. Finally
,
b
y
apply
i
ng
the proposed method it was
shown that th
e g
o
al
even
ts d
e
tection has
a good
accuracy
and th
e
percen
tag
e
of detection f
a
ilure is
also v
e
r
y
low.
Keyword:
Ev
en
t d
e
tection
Field cente
r
de
tection
Field ext
r
action
Sho
t
bou
nd
ar
y
d
e
tectio
n
Shot classification
Socce
r vi
de
o p
r
oces
si
n
g
Copyright ©
201
4 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
:
Far
s
h
a
d Bayat,
Depa
rt
m
e
nt
of
El
ect
ri
cal
and
C
o
m
put
er E
ngi
neeri
n
g
,
Islamic Azad
Uni
v
ers
ity, Qazv
in
B
r
an
ch
,
No
k
hbe
ga
n
B
l
vd
. Qazvi
n
. Ira
n.
Em
a
il: farsh
a
d
.
b
a
yat@q
i
au
.ac.ir
1.
INTRODUCTION
In
rece
nt years, the
devel
opmen
t
of m
u
l
t
i
m
e
di
a appl
i
cat
i
ons i
n
di
ffe
re
nt
fi
el
ds
, i
n
c
r
ease of
t
h
e
i
n
t
e
rnet
i
m
pact
and
devel
o
p
m
ent
of c
o
m
m
uni
cat
i
ons, l
e
d t
o
t
h
e
de
vel
o
pm
ent
of
m
u
l
t
i
m
e
di
a
st
ora
g
e
t
echn
o
l
o
gy
. T
h
ere
f
o
r
e, p
r
od
uci
n
g i
n
f
o
rm
ati
on m
a
nagem
e
nt
and ret
r
i
e
va
l
t
ool
s for h
a
n
d
l
i
ng t
h
ese l
a
r
g
e dat
a
becam
e necessary. In t
h
is
regard, sp
ort c
o
mpetitions a
r
e
one
of the m
o
st
popular m
u
lti
media and
on
top
of
t
h
em
, wat
c
hi
ng socce
r i
s
t
h
e m
o
st
exci
ti
ng and
po
p
u
l
a
r sp
o
r
t
vi
de
os. T
hus
, hi
g
h
v
o
l
u
m
e
s of socce
r vi
de
os are
pr
o
duce
d
an
d t
h
i
s
po
p
u
l
a
ri
t
y
has l
e
d resea
r
c
h
ers t
o
be
in
terested
in
creatin
g
n
e
w id
eas an
d
ways to
m
a
n
a
g
e
and
retrieve inform
ation from soccer
vide
os [1,
2 and 3]
. Socce
r vi
deo pr
ocessing a
n
d inform
ation retrieva
l
t
echni
q
u
es are
di
vi
de
d i
n
t
o
t
w
o onl
i
n
e and
of
fl
i
n
e g
e
neral
st
ru
ct
ur
es. Onl
i
n
e t
e
chni
que
s ext
r
act
t
h
e
i
n
f
o
rm
at
i
on of
a li
ve gam
e
and t
h
e
r
eby
e
nha
nce vi
si
t
o
r
s
’ i
n
f
o
rm
at
i
on of a gam
e
bei
ng di
s
p
l
a
y
e
d
.
For
exam
ple, the
num
ber of team
’s offside, the distance tr
aveled by a player, the tim
e
of ball possessi
on by
tea
m
s. In [4] s
o
m
e
ca
m
e
ras
with hi
gh
fra
me rate (e.g
.200fs
p)
were
pl
aced in t
h
e stadium
and used to fi
nd
online
goal. T
h
e purpose of
using cam
eras
with
h
i
gh
frame rate is
th
at i
n
t
hose s
hot
s t
h
at
goal
eve
n
t
occu
rs
t
h
e bal
l
spee
d
i
s
oft
e
n
very
h
i
gh;
t
h
ere
f
ore
,
usi
n
g t
h
e
vi
de
o o
b
t
a
i
n
e
d
fr
o
m
conve
nt
i
o
na
l
sy
st
em
s wi
t
h
abo
u
t
30 fram
es per second recording spee
d, makes detec
tion of goal occurre
nce very di
fficult and pe
rha
p
s
im
possi
bl
e.
In
[4]
t
w
o t
ech
ni
que
s ha
ve
bee
n
used
f
o
r
det
e
ct
i
ng
goal
e
v
e
n
t
.
I
n
t
h
e fi
rst
m
e
t
hod,
t
h
e al
go
ri
t
h
m
of ci
rcl
e
det
ect
i
on
(ci
r
cul
a
r H
o
u
g
h
t
r
a
n
sf
or
m
al
gori
t
h
m
)
i
s
use
d
t
o
fi
n
d
candi
dat
e
s.
In
t
h
e seco
n
d
t
echni
qu
e
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJECE Vol. 4, No. 6, D
ecem
ber 2014
:
979 – 988
98
0
t
h
e pr
o
duce
d
candi
dat
e
s are
eval
uat
e
d a
n
d com
p
ared
u
s
i
ng a s
upe
rvi
s
ed ne
ur
al
net
w
o
r
k
,
an
d t
h
e
pr
ope
r
can
d
i
d
a
te is selected
. By track
ing
th
e
p
o
s
itio
n
o
f
t
h
e b
a
ll,
th
e o
c
cu
rren
ce o
f
g
o
a
l ev
en
t
is d
e
term
in
ed
. Sin
ce
watching t
h
e
whole s
o
ccer ga
m
e
video is time cons
um
i
ng so, offli
n
e inform
ation retrie
val and m
a
nage
m
e
nt
tools
help s
o
cc
er ent
h
usiasts and resea
r
ch
ers
in
th
is fiel
d
.
Th
e
o
f
flin
e ex
tractio
n
an
d
an
aly
s
is to
o
l
s allow
u
s
t
o
extract the ne
cessary information
fro
m
t
h
e g
a
m
e
s in
t
h
e lo
west ti
me
. An e
x
am
ple of suc
h
i
m
po
rt
ant
in
fo
rm
atio
n
is d
e
tection
o
f
g
o
a
l ev
en
ts in so
ccer
g
a
m
e
s. Utilizin
g these m
a
n
a
g
e
m
e
n
t
and
i
n
formation
retriev
a
l too
l
s
su
mm
arize th
e
v
i
d
e
o
g
a
m
e
an
d
resu
lt in
sav
i
ng
th
e so
ccer exp
e
rts’ tim
e
wh
en
an
alyzin
g
t
h
e
t
act
i
c
s
of
t
h
e g
a
m
e
.
Goal
det
ect
i
on based o
n
ci
nem
a
ti
c
feat
ures
i
s
p
r
esent
e
d
i
n
[5]
.
Thi
s
pape
r poi
nt
s
o
u
t
t
h
at
ju
st
usi
n
g
vi
de
o p
r
ocessi
n
g
t
e
chni
que
s f
o
r
g
o
al
det
ect
i
o
n i
s
di
ffi
c
u
l
t
.
T
h
er
efo
r
e, t
h
ey
ha
v
e
use
d
t
h
e a
d
v
a
nt
age
of t
h
e nat
u
ral
event
s
t
h
at
occ
u
r a
f
t
e
r a
goal
event
.
The a
u
t
h
o
r
s
have
pr
o
p
o
se
d f
o
l
l
o
wi
ng
pr
oce
d
u
r
es f
o
r g
o
al
d
e
tectio
n
:
con
s
id
ering
th
e in
terru
p
tion
s
d
u
ring
th
e g
a
m
e
,
co
n
s
id
ering
th
e ex
isten
ce of at least o
n
e
sh
o
t
ou
t of
the field (e
.g. a
udie
n
ce s
h
ot)
or a clos
e-
up
s
hot
,
co
nsi
d
eri
n
g t
h
e
prese
n
ce
of
at least one
replay s
hot. In
[6,
7]
t
h
e sc
ore
b
oar
d
i
n
f
o
rm
at
i
on i
s
use
d
t
o
fi
nd
a
goal
occ
u
r
r
e
n
c
e
i
n
s
o
cce
r
vi
d
e
o.
I
n
t
h
i
s
m
e
tho
d
,
g
o
al
occ
u
rre
nces
are d
e
t
e
rm
i
n
ed by
det
ect
i
on
of
sco
r
eb
oa
rd
l
o
cat
i
on
d
u
ri
n
g
t
h
e
vi
deo
di
spl
a
y
an
d t
r
ac
ki
n
g
c
h
an
ges i
n
i
t
.
A
di
sad
v
a
n
t
a
ge
o
f
t
h
i
s
m
e
t
hod i
s
t
h
at
t
h
e t
e
xt
u
r
e o
f
m
o
st
of
t
h
e sc
ore
b
oar
d
s
i
s
h
o
m
ogene
o
u
s
wi
t
h
t
h
e
t
e
x
t
ure
of
th
e aud
i
en
ce, wh
ich
lead
s to
d
i
fficu
lty in
sco
r
ebo
a
rd
d
e
tectio
n
an
d m
a
tch
resu
lt find
ing
.
Ano
t
h
e
r
di
sad
v
a
n
t
a
ge
o
f
t
h
i
s
m
e
t
hod i
s
t
h
e
occu
rre
n
ce of
i
m
age i
n
t
e
rfere
nce
o
n
t
h
e l
o
cat
i
o
n
of
score
b
oar
d
(e.
g
. t
h
e
TV c
h
a
nnel
s
’ l
o
g
o
s,
a
dve
rt
i
s
i
n
g
et
c.)
,
t
h
at
m
a
ke t
h
e
det
e
ct
i
on
of
cha
n
g
e
s o
n
t
h
e sc
or
eboa
r
d
di
ffi
c
u
l
t
.
Usi
n
g
m
u
l
ti-clu
e
s is an
o
t
h
e
r st
rategy to
find
th
e
go
al o
c
cu
rren
ce [8
].
In
t
h
is m
e
th
od
th
e ca
ndidate shots are
chosen
by
usi
n
g t
h
e cl
ues suc
h
as i
n
crease of s
o
un
d i
n
t
e
nsi
t
y
and
occur
r
e
n
ce of
goal
bar i
n
t
h
e shot
. Ne
xt
, t
h
e g
o
al
event
s
are
det
e
rm
i
n
ed by
revi
ewi
n
g t
h
e ne
xt
shot
s an
d i
d
e
n
t
i
f
y
i
ng t
h
e vi
s
u
al
cl
ues of g
o
al
scori
n
g s
u
ch a
s
t
h
e
audi
e
n
ces’
s
h
o
t
and
re
pl
ay
sh
ot
. T
h
i
s
pape
r
pr
o
v
i
d
es a
fra
m
ework t
o
detect the goal e
v
ents in
a s
o
cce
r gam
e
vi
de
o. F
o
r t
h
i
s
pu
rp
ose,
we
ext
r
act
vi
de
o e
v
ent
s
usi
n
g b
o
t
h vi
sual
an
d
audi
o cha
n
nel
s
. Thi
s
f
r
am
ewor
k i
s
com
posed
of
several m
a
in stages th
at in
each stage
we
relate a number
of feature
s
to hi
ghe
r se
mantic
conce
p
t
s
. I
n
t
h
e fi
rst
st
ep we use basi
c feat
u
r
es o
f
so
un
d a
nd
det
e
rm
i
n
e the im
port
a
nt
event
s
of vi
deo
.
In t
h
e
next
st
ag
e, we
use di
ssi
m
i
l
a
ri
ty
feat
ures
fo
r s
hot
bo
u
nda
ry
det
ect
i
on a
nd t
h
en cl
assi
fy
t
h
ese sh
ot
s wi
t
h
basi
c
visual
feature
s
and in the
ne
xt stage
we fi
nd t
h
e m
o
m
e
nts whe
n
the
a
udie
n
ces’ shot
and close s
hot exist
whi
c
h i
s
t
h
e i
n
di
cat
or
of i
m
port
a
nt
scenes
o
f
vi
de
o.
I
n
the
final step
we use the
created meaningful uni
ts
and
det
ect
g
o
al
m
o
ut
h a
n
d
gam
e
st
art
fram
e
wor
k
. B
y
appl
y
i
ng these steps we
can fi
nd
goal e
v
ents a
n
d s
u
mmarize
th
e v
i
d
e
o. Th
is fram
e
work is illu
strated
i
n
Fi
g
u
re
1
.
Fi
gu
re
1.
B
l
oc
k
di
ag
ram
of t
h
e p
r
o
p
o
sed
m
e
tho
d
.
2.
E
X
T
R
ACTI
N
G THE E
X
IS
TING E
V
EN
TS FR
OM
SO
CCE
R
VI
DEO
Undo
ub
ted
l
y th
e m
o
st p
o
p
u
l
ar spo
r
t in
th
e world
is so
ccer. The m
a
in
id
ea is th
at as an
im
p
o
r
tan
t
event occ
u
rs in a soccer ga
me, the audiences’ a
nd re
porters’ sound
intensity increases. Thus, the audi
o
chan
nel
has
hi
gh sem
a
nt
i
c
i
n
fo
rm
ati
on an
d
t
h
i
s
feat
ure ca
n be wi
del
y
us
ed f
o
r
det
ect
i
on o
f
im
port
a
nt
event
s
.
Usi
n
g t
h
i
s
feat
ure l
e
a
d
s t
o
i
n
crease
of
det
e
ct
i
on acc
uracy
and
nat
u
ral
l
y
resul
t
s
i
n
del
e
t
i
on
of
uni
m
p
o
r
t
a
nt
part
s an
d dec
r
ease o
f
pr
oc
essi
ng c
o
m
p
l
e
xi
t
y
of cal
cul
a
t
i
ons.
I
n m
o
st
of t
h
e exi
s
t
i
ng i
m
port
a
nt
event
s
extraction approache
s
, the
bas
i
c audio si
gnal energy level fe
ature is used. In
[9], the a
b
ove feature along wit
h
t
h
e det
ect
i
on
of s
o
m
e
key
w
or
ds i
n
cl
u
d
i
n
g
t
h
e wor
d
s s
u
ch as "goal
"
and "
p
enal
t
y
" i
s
used fo
r f
i
ndi
n
g
im
port
a
nt
eve
n
t
s
of socce
r ga
m
e
. Som
e
t
i
m
e
s i
n
t
h
e gam
e
,
t
h
e audi
e
n
ces l
o
u
d
l
y
and c
ont
i
n
u
o
u
s
l
y
enco
ura
g
e
their fa
vorite t
e
a
m
, so i
n
the
s
e cases t
h
e a
u
dio si
gnal
ene
r
gy level
feature won’t ha
ve
the
require
d
ac
curacy
for th
e d
e
tectio
n
o
f
im
p
o
r
tant ev
en
ts. In
[10
]
, in
ad
d
ition
to
th
e au
d
i
o
sig
n
a
l en
erg
y
lev
e
l featu
r
e, th
e b
a
sic
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Goa
l
Detectio
n
in
S
o
ccer
Vi
d
e
o
:
Ro
le-Based Even
ts Detectio
n
Ap
p
r
oa
ch
(F
a
r
sh
a
d
B
a
ya
t)
98
1
feat
ure
of ze
r
o
crossi
ng
rat
e
i
s
us
ed to e
nhance the acc
uracy of im
porta
n
t
ev
en
ts d
e
tectio
n
and
th
e
resu
lts
show t
h
at the
accuracy
has i
n
crease
d
t
o
a l
a
rge
exte
nt.
In this
pape
r,
we
ha
ve
use
d
the
proposed m
e
thod i
n
[10
]
; and
after
id
en
tifying
th
e ex
citing
p
a
rts
o
f
th
e m
a
tch
,
w
e
seg
m
en
t th
e v
i
d
e
o in
t
h
e
p
e
ri
o
d
s
of
50
seconds
bef
o
re
an
d
85
seco
nds
aft
e
r t
h
e
det
ect
ed
poi
nt
s, i
n
o
r
de
r t
o
per
f
o
rm
subse
que
nt
pr
ocessi
ng
. It
i
s
w
o
rt
h
not
i
n
g
t
h
at
t
h
e se
gm
ent
a
t
i
on l
e
n
g
t
h
has
bee
n
obt
ai
ned
aft
e
r
re
pea
t
ed ex
pe
ri
m
e
nts an
d si
m
u
l
a
t
i
ons.
3.
SHOT BOUNDARY
DETECTION
The m
o
st
im
port
a
nt
st
ep i
n
t
h
e seri
es of vi
de
os p
r
oce
ssi
n
g
and
ret
r
i
e
vi
ng i
s
shot
b
o
u
n
d
a
r
y
det
ect
i
on.
By sh
o
t
b
oun
dary d
e
tectio
n,
th
e b
a
sic sem
a
n
tic stru
ct
u
r
e
o
f
v
i
d
e
o
will be read
y
for th
e n
e
x
t
step
s
o
f
v
i
d
e
o
pr
ocessi
ng
. S
h
ot
b
o
u
n
d
ary
i
s
det
e
rm
i
n
ed by
si
gni
fi
ca
nt
cha
nge
s i
n
t
h
e
co
m
posi
t
i
on o
f
c
o
l
o
r o
r
pi
xel
l
o
cat
i
o
n
[1
1]
. The
dom
i
n
ant
col
o
r i
n
t
h
e socce
r vi
de
o i
s
gree
n;
ho
weve
r i
t
s
am
ount
i
s
cha
nge
d
i
n
som
e
part
s of t
h
e
vi
de
o. F
o
r e
x
a
m
pl
e,
t
h
e am
ount
o
f
g
r
een c
o
l
o
r w
h
ere t
h
e
cam
e
ra i
s
l
o
cat
e
d i
n
t
h
e m
i
ddl
e o
f
t
h
e fi
el
d o
r
far
v
i
ew is sign
ifican
tly d
i
fferent fro
m
close or audience s
h
ot. Thes
e cha
nge
s can be
used as a m
eas
ure
of
d
i
ssimilarity to
sep
a
rate th
e sh
o
t
bo
und
aries. An
o
t
h
e
r d
i
ssi
milarit
y
criterio
n
th
at is also
u
s
ed
in
t
h
is p
a
p
e
r i
s
th
e d
i
fferen
c
e in
th
e co
l
o
r in
ten
s
ity.
3.1. The Fe
ature of Gree
n Color Rati
o Difference
In
t
h
i
s
pape
r
w
e
ha
ve
use
d
t
h
e
feat
u
r
e
of
g
r
ee
n c
o
l
o
r rat
i
o
di
ffe
rence
t
h
at
i
s
speci
fi
e
d
as
f
o
l
l
o
ws.
|
1
|
In t
h
is f
o
rm
ula,
sh
o
w
s t
h
e
t
o
t
a
l
num
ber o
f
pi
xel
s
an
d
Pg
i
sho
w
s t
h
e n
u
m
ber of
gre
e
n
p
i
x
e
ls i
n
th
e ith
fram
e
th
at is ch
aracterized
as fo
llo
ws:
,
(3
)
,
1,
T
E
D
,
i
0,
O.
W
(4
)
Whe
r
e
ED
,
i
1
,
,
,
.
Th
e P
(x
, y)
variab
le rep
r
esen
ts th
e
v
a
lu
es
o
f
R
G
B co
l
o
r
in
th
e
p
o
s
ition o
f
(x, y) i
n
the ith
fram
e
.
The
ED
,
i
va
riable
shows t
h
e E
u
c
lidean
distan
ce
bet
w
ee
n t
h
e
c
o
l
o
rs o
f
t
h
e pi
xel
,
in
th
e ith
frame
rel
a
t
i
v
e t
o
t
h
e
pu
re
gree
n c
o
l
o
r
.
C
e
rt
ai
nl
y
,
t
h
e c
hoi
ce
of
t
h
e t
h
res
h
ol
d
val
u
e i
n
di
f
f
ere
n
t
peri
ods
an
d l
i
g
ht
i
n
g
co
nd
itio
ns of stad
iu
m
wo
u
l
d
b
e
d
i
fferen
t. To
find
th
e ap
pro
p
riate th
resh
old
v
a
lu
e, we use th
e fo
llowing
fo
ur
steps:
1.
Sel
ect
num
ber
of
f
r
am
es from
vi
de
o
ra
nd
om
ly
and
cal
l
t
h
es
e fram
e
s
.
2.
For each
fram
e
, calculate the Euclidea
n
distance
relative to green
color a
nd a
rra
nge them
in
desce
ndi
ng
o
r
d
e
r.
3.
To
fi
n
d
t
h
e
ap
pr
o
p
ri
at
e t
h
res
hol
d
val
u
e
,
fra
m
es are selecte
d
from
the
beginning of t
h
e a
rra
nge
d set
and from
each
؛
1
; calculate Euclidea
n
distance
for eac
h
pixel, a
n
d t
h
e
n
a
rra
nge
thes
e
val
u
es
i
n
asce
n
d
i
n
g
or
der
.
4.
R
e
m
ove
of the end elem
ents and
of the
first elem
ents
of t
h
e
set and avera
g
e the
rest of
t
h
e
v
a
lu
es of th
e set an
d con
s
i
d
er it as th
e t
h
resho
l
d
v
a
lu
e.
R
e
m
oval
of
t
h
e
of the
end
values
of the
s
e
t leads
to
del
e
tion
of i
n
accurate inform
atio
n a
n
d al
s
o
rem
oval
of t
h
e
of t
h
e
dat
a
fr
om
t
h
e begi
n
n
i
ng
of t
h
e set
m
a
kes t
h
e E
u
cl
i
d
ean
di
st
ance
f
r
om
t
h
e pu
re g
r
ee
n
g
r
eater and
prov
id
es cond
itio
ns to
cho
o
s
e d
a
rk
er co
l
o
rs
. Th
resho
l
d
v
a
lu
e is u
s
u
a
lly in
th
e
[0.6,
0
.
8
]
ran
g
e
.
Th
e cr
iter
i
on
of
gr
een
co
lo
r ratio
ch
an
g
e
s,
deter
m
in
es th
e
pr
op
er bou
nd
ary f
o
r
th
e sho
t
fr
o
m
th
e sho
t
b
oun
d
a
ry in
the fram
e
s th
at
g
r
een
co
l
o
r com
b
in
atio
n
s
chan
g
e
d
r
am
atic
ally. Bu
t in
th
i
s
criterio
n
, th
e sp
atial
inform
ation of
pixels is not c
onsi
d
ere
d
.
To
boost the
detec
tion accuracy
of s
h
ot bounda
r
y, we
use a c
r
iterion
th
at con
s
id
ers t
h
e sp
atial in
fo
rmatio
n
of
p
i
x
e
l
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJECE Vol. 4, No. 6, D
ecem
ber 2014
:
979 – 988
98
2
3.
2. T
h
e Fe
at
u
re o
f
Co
l
o
r Intensity
Differe
nce
If
,
is cth co
l
o
r ch
ann
e
l in
tensity
∈
,
,
in
,
po
sition
and
ith frame,
criterio
n
of
d
i
ssi
m
ilarit
y
, d
i
ffer
en
ce of sp
atial in
ten
s
ity of
co
lors, is d
e
fined
as fo
llo
ws:
,
,
∀
,
β
∈
1,
0
,
1
(5
)
,
,
1
,
,
1
,
,
,
,
/
,
q
∈
1,
∞
(6
)
In
w
h
i
c
h
q
de
fi
nes t
h
e t
y
pe
of
di
st
an
ce, s
o
t
h
at
q=
1 an
d
q=2 t
h
at
s
h
ows
t
h
e ci
t
y
di
st
ance a
n
d
Euclidea
n distance, res
p
ectively.
Param
e
ters
of
α
an
d
β
d
e
t
e
rm
i
n
e t
h
e sl
i
d
i
n
g
am
ount
of t
h
e
fi
rst
fra
m
e
ove
r
th
e
s
e
con
d
f
r
a
me
in
th
e
x and
y ax
e
s
, r
e
sp
ec
tiv
e
l
y.
T
h
e
e
quat
i
o
n
det
e
rm
i
n
es t
h
e m
i
nim
u
m
di
st
ance
b
e
tween
two
co
n
s
ecu
tiv
e frames in
th
e
9
di
rections
s
p
ecifi
ed in Figure
2.
Fi
gu
re 2.
Di
f
f
e
r
ent
val
u
es
o
f
α
and
β
i
n
9 di
r
ect
i
ons
The GR
D feat
ure
produces
values in
[0, 1] range
a
nd S
P
D
feat
ure pr
od
uces val
u
es i
n
[0
,
Im
ax]
rang
e th
at Imax
is th
e
m
a
x
i
m
u
m
lig
h
tin
g in
ten
s
ity in
the re
d and blue cha
nnels
. For ease of work and
classificatio
n
pro
cess, it is b
e
t
t
er to
n
o
r
m
a
liz
e th
e v
a
lu
es in the range
of z
e
ro and one and we should notice
th
at th
e
GR
D
v
a
lu
es are usually in
th
e
rang
e
o
f
[0
, 0.
8
6
]
be
fo
re n
o
rm
alization. Follo
win
g
fo
rm
ulas
can
b
e
use
d
f
o
r
no
rm
al
i
zat
i
on:
(7
)
(8
)
In Figure
3-a
you ca
n
see the norm
alized SPD cha
n
g
e
r
a
te an
d in Figu
re 3-
b th
e
no
r
m
alized
G
R
D
chan
ge
rat
e
i
n
t
e
rm
s of t
h
e fram
e
num
ber.
We use t
w
o
NSP
D
a
nd
NGR
D feat
ure
s
t
o
det
e
rm
i
n
e sh
ot
bounda
ry.
So,
feature
s
p
ace i
s
as follows:
,
|0
(9
)
The be
st
can
di
dat
e
fo
r s
hot
b
o
u
n
d
ary
i
s
t
h
e
fram
e
s
in
wh
ich
th
e
v
a
lu
es
o
f
NSPD or NGRD features
or b
o
t
h
val
u
es
are i
n
l
o
cal
m
a
xim
u
m
.
In or
der t
o
fi
nd t
h
e l
o
cal
m
a
xi
m
u
m
,
t
h
e val
u
e of feat
u
r
e sh
oul
d be
m
a
xim
u
m
i
n
com
p
ari
s
on
wi
t
h
i
t
’
s be
fo
re an
d aft
e
r
.
The
r
ef
ore
,
we
use f
o
l
l
owi
n
g e
quat
i
o
ns i
n
or
der t
o
f
i
nd t
h
e
diffe
re
nce between
val
u
es:
,
1
(1
0)
,
1
(1
1)
In t
h
e above e
quation
are the sam
e
features that are s
p
ecified
in
equ
a
tion
(9
). In
o
r
d
e
r
to
find
lo
cal
m
a
xim
u
m
,
we
use t
h
e f
o
l
l
o
wi
ng
eq
uat
i
o
n:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Goa
l
Detectio
n
in
S
o
ccer
Vi
d
e
o
:
Ro
le-Based Even
ts Detectio
n
Ap
p
r
oa
ch
(F
a
r
sh
a
d
B
a
ya
t)
98
3
,
,
,
,
,
0,
.
0.01
(1
2)
The
shows t
h
e
differe
n
ce c
o
m
p
ared
to
n
e
igh
bors who
s
e valu
es
are selected em
pirically.
Fi
gu
re 3.
Fram
e-base
d feat
u
r
e
s
di
a
g
ram
:
(a)
The
feature
of
gree
n c
o
lor
ratio c
h
anges
in te
rm
s of fram
e
num
ber,
(
b
)
Th
e feat
u
r
e
of
spa
t
i
a
l
chan
ges
of
pi
xel
s
’
col
o
r
i
n
t
e
nsi
t
y
i
n
t
e
rm
s o
f
f
r
am
e num
ber
An
d
gi
ve
n t
h
a
t
t
h
e sh
ot
l
e
n
g
t
h
i
s
at
l
east
40
fram
e
s, w
e
use
wi
n
d
o
w
i
ng
ope
rat
o
r a
nd
f
o
r eac
h
wind
ow select
a lo
cal m
a
x
i
mu
m
.
To
d
o
th
is, th
e
fo
llowing
eq
u
a
tion
is u
s
ed
to id
en
tify the sho
t
b
oun
d
a
ry:
,
0∃
j
∈
i1
,
i
l
|
,
,
1,
o.
w.
(1
3)
In
w
h
ich
in
d
i
cates
th
e
leng
th o
f
th
e windo
w,
h
e
re we h
a
v
e
co
nsid
ered
40
fram
e
s.
Wh
en
t
h
e SB
v
a
lu
e i
n
ith
frame eq
u
a
ls
1
,
t
h
e fram
e will
b
e
th
e
sho
t
bou
nd
ary, ind
eed in
th
is fram
e
t
h
e val
u
es
of
f
eat
ures a
r
e l
o
c
a
l
m
a
xim
u
m
.
In Fi
gu
re
4.a s
h
ot
b
o
u
n
d
ary
e
x
t
r
act
i
on
base
d
on
N
S
P
D
feat
u
r
e an
d
in
Figur
e
4
.
b
t
h
e sho
t
bou
nd
ar
y ex
tr
acti
o
n by N
G
R
D
is show
n.
Fi
gu
re
4.
(a
) S
hot
b
o
u
n
d
ary
e
x
t
r
act
i
o
n
by
N
SPD
feat
ur
e
(A)
NSPD feat
ure c
u
rve
(B) E
x
tracted boundaries by
l
o
cal
m
a
xim
u
m
.
(b) S
h
ot
b
o
u
nda
ry
e
x
t
r
act
i
o
n
by
N
G
RD
fe
ature (A) NGRD
feat
ure
curve (B) Ext
r
acted
bo
u
nda
ri
es by
l
o
cal
m
a
xim
u
m
.
In o
r
der t
o
fi
n
d
t
h
e fi
nal
shot
bo
u
nda
ri
es, we
use
t
h
e com
b
inat
i
on
of s
hot
bo
u
nda
ri
es fo
u
nd
based
o
n
NSPD a
n
d
NGRD feat
ures
, as
follows:
,
,
,
(1
4)
We ha
ve
use
d
wi
n
d
o
w
i
n
g t
e
c
hni
que i
n
o
r
de
r t
o
decrease e
r
r
o
r i
n
fi
n
d
i
n
g
t
h
e sh
ot
b
o
u
n
d
a
ry
. Gi
ven
that the shot lengt
h is at least 40
fram
e
s, we can use a
wi
n
d
o
w
wi
t
h
a l
e
n
g
t
h
of
40
fram
e
s;
and
we sel
ect
t
h
e
m
a
in shot bounda
ry from
a
m
ong the set
of s
hot bounda
ries with the m
o
st SPD
in each window. Figure 5
sh
ows th
e
result o
f
co
m
b
in
ing
th
e feat
u
r
es.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJECE Vol. 4, No. 6, D
ecem
ber 2014
:
979 – 988
98
4
Fi
gu
re
5.
C
o
m
b
i
n
at
i
o
n
of
sh
o
t
bo
u
nda
ri
es t
h
at are s
p
ecified by
NSPD a
n
d
NGR
D
feature
s
.
4.
CLAS
SIFI
C
A
T
ION OF SH
OTS
There a
r
e ge
nerally four different s
hots in
soccer
gam
e
s. These four s
h
ots are shown
in Figure 6.
The fa
r s
hot
i
s
usual
l
y
use
d
t
o
sh
o
w
t
h
e
ga
m
e
si
t
u
at
i
on i
n
whi
c
h t
h
e
nu
m
b
er of pl
ay
er
s and
g
r
eat
am
ou
nt
o
f
pl
ay
i
ng fi
el
d’s
grass i
s
i
n
t
h
e
im
age. The m
i
ddl
e sh
ot
i
s
u
s
ed t
o
s
h
o
w
t
h
e
m
ovem
e
nt
s
of t
h
e
pl
ay
ers
and t
h
e
ca
m
e
ra zo
o
m
s to
th
e po
sition
of th
e
p
l
ayer who
h
a
s t
h
e
b
a
ll so
th
at the n
u
m
b
e
r
o
f
players with
rel
a
tiv
ely
l
a
rge si
ze
i
s
i
n
cl
ude
d i
n
t
h
e
i
m
age. C
l
ose
s
hot
us
ual
l
y
occ
u
rs
w
h
e
n
st
op
has
occ
u
r
r
ed
d
u
ri
ng
t
h
e
gam
e
w
h
i
c
h
is u
s
u
a
lly associated
with
imp
o
rtan
t ev
en
ts
o
f
t
h
e
p
l
ay
; and
in
t
h
is m
o
d
e
, th
e cam
era will zo
o
m
o
n
a
p
l
ayer to
th
e fu
ll ex
ten
t
so
th
at th
e m
a
j
o
rity o
f
t
h
e imag
e is related
t
o
th
e
p
l
ayer.
Au
d
i
en
ce sh
o
t
is related
to
th
e
v
i
ew
that ca
m
e
ra images the audience pre
s
en
t
in
th
e stad
iu
m
th
at is u
s
u
a
lly after th
e g
o
al o
ccu
rren
ce. After
identifying the
shot
bounda
ri
es, a ke
y fram
e
is extracted from
each shot a
nd t
h
e algorithm which is shown i
n
Figure 7 is pe
rform
e
d as a hierarc
h
y. The features of
col
o
r ratio, holes’
area ratio and
edge
ratio are
use
d
for
shot
cl
assi
fi
cat
i
on.
Fi
gu
re
6.
S
hot
Ki
n
d
s,
res
p
ect
i
v
el
y
(
A
)
Far
sh
ot
(B
) M
i
d
d
l
e
Sh
ot
(C
) C
l
ose
Sh
ot
(D
)
Au
di
ence S
h
ot
.
To
calcu
late ho
les’ area
ratio, first we ex
tract g
r
a
ss area
from
the
main fra
m
e and detect
the blobs in
the grass
area a
n
d calculate th
e b
l
ob
area
ratio
as fo
llo
ws:
∑
(1
5)
In
w
h
ich
is th
e set of
h
o
l
es in
th
e ith
fram
e
an
d C sh
ow
s t
h
e nu
m
b
er
of
h
o
les.
To calc
u
late the edge
ratio
we
canny
edge
de
tection on
gray level fram
e.
(1
6)
We
use t
h
e foll
owi
n
g equation to calcu
late
Energy
of NSPD in each s
hot:
,
∑
|
|
(1
7)
Param
e
ters s and e s
p
ecify the be
ginnin
g a
n
d e
n
d
o
f
t
h
e s
h
ot
,
respect
i
v
el
y
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Goa
l
Detectio
n
in
S
o
ccer
Vi
d
e
o
:
Ro
le-Based Even
ts Detectio
n
Ap
p
r
oa
ch
(F
a
r
sh
a
d
B
a
ya
t)
98
5
Fi
gu
re
7.
Di
a
g
r
a
m
of
Sh
ot
C
l
assi
fi
cat
i
o
n
5.
DETECTION OF
GOAL E
V
ENT
We
d
e
tect go
al ev
en
ts
u
s
i
n
g
sem
a
n
tic u
n
its. Th
es
e un
its are ob
tain
ed
h
i
er
arc
h
i
cal
l
y
an
d
by
usi
ng
im
age processi
ng techniques.
The
high
-l
e
v
el
uni
t
s
a
r
e descri
bed
bel
o
w
.
5.
1 Detec
t
i
o
n
of
G
oal
m
o
u
t
h
Thi
s
pape
r has use
d
t
h
e p
r
ese
n
t
e
d way
s
i
n
[
12]
fo
r det
ect
i
o
n
o
f
goal
m
out
h whi
c
h
i
s
c
o
m
pos
ed o
f
t
h
e
follo
win
g
t
h
ree
stages:
1.
First,
we app
l
y to
p
h
a
t tran
sfo
r
m
a
tio
n
(THT)
o
n
th
e
frame. T
h
is tra
n
sform
a
tion leads t
o
inc
r
ease
of
wh
ite co
l
o
r in
t
e
n
s
ity an
d
th
erefore th
e edg
e
s o
f
g
o
a
lm
o
u
t
h
will b
e
m
o
re sp
ecific and
fo
r a b
i
n
a
ry
im
age, a self-a
daptive
thres
h
old can be
used.
2.
Th
e wh
ite v
e
rt
ical lin
es in
th
e
im
age are extracted as the
prope
r can
di
d
a
t
e
for t
h
e
g
o
a
lm
out
h as
fo
llows. V ind
i
cates v
e
rtical
wh
ite lin
es.
,
1 ,
∀
,
∈
,
,
1
0 ,
.
(1
8)
,
1,
∃
∈
1
0
,
1
0
,∀
∈
10,
10
,
,
0 ,
.
(1
9)
In
t
h
e a
b
ove
eq
uat
i
o
n
,
VL
x
indicat
es the
vertical l
i
nes.
3.
Tw
o l
i
n
es t
h
at
hav
e
t
h
e
fol
l
owi
ng t
e
rm
s are sel
ect
ed as
t
h
e l
i
n
es
of
g
o
al
m
out
h f
r
om
am
ong t
h
e
vertical lines e
x
tracted in ste
p
2.
20
0
.
3
∗
2
0
0
.
4
∗
0.5
,
2
0
0
|
|
2
0
1
0
.
3
∗
0.5
∗
(2
0)
H an
d
W, res
p
ect
i
v
el
y
,
det
e
r
m
i
n
e t
h
e wi
dt
h
and
hei
g
ht
pi
x
e
l
s
and Li
sh
o
w
s t
h
e am
ount
of pi
xel
s
i
n
each line
,
D s
h
ows
the
horiz
o
ntal
distance
between t
h
e t
w
o lines,
and
, res
p
ectively,
pres
ents the
hi
ghes
t
an
d l
o
w
e
st
positio
n
of
lin
e,
an
d
represen
ts th
e
o
v
e
rlap
o
f
th
e two lin
es in the
ve
rtical directi
o
n.
By
form
ing the
s
m
allest rectangle
b
y
th
e t
w
o
v
e
rtical lin
es, C
(
,
) is t
h
e
center
of t
h
e
rectangle t
h
at
represen
ts th
e wid
t
h
of th
e cen
ter of th
e rectang
l
e.
Th
e p
r
o
c
ess o
f
go
alm
o
u
t
h
d
e
tectio
n
is illu
strated
in
Fi
gu
re 8.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJECE Vol. 4, No. 6, D
ecem
ber 2014
:
979 – 988
98
6
5.2 Detec
t
ion of
Center of
the
Field
Sho
t
s related t
o
th
e cen
ter
o
f
th
e field
are
com
posed
of t
w
o m
a
in com
p
onents:
1.
Ex
isten
ce
of el
lip
se in
t
h
e im
a
g
e
2.
Ellip
se
cu
t b
y
a
lin
e
Th
ere are d
i
fferen
t
ways to
su
rv
ey ellip
se su
ch
as Least-Squ
a
res Fitting
(LSF) [13
,
14
], Inv
a
rian
t
Pattern
Filter (IPF)
[15
]
an
d
Ellip
se Hou
gh transfo
r
m
(EHT)
[16
]
. In
th
is p
a
p
e
r we h
a
v
e
u
s
ed
th
e Houg
h
t
r
ans
f
o
r
m
m
e
t
h
od
, a
n
d
ha
ve m
a
de
bi
na
ry
i
m
age as
f
o
l
l
o
w
s
, t
o
obt
ai
n
bet
t
e
r
resul
t
s
.
(E)
(D
)
(C)
(B)
(A
)
Fi
gu
re
8.
St
age
s
o
f
goal
ba
r e
x
t
r
act
i
o
n:
(
A
)
t
h
e m
a
i
n
im
age, (B
)
ap
pl
i
cat
i
on
of
EH
T,
(C
)
m
a
ki
ng i
m
age
bi
na
ry
, (
D
) ca
n
d
i
d
at
es
of
ve
r
tical lines, (E) goal ba
r
[12].
First, we calculate th
e Eu
clid
ean
distance of the correspondi
ng fram
e re
lativ
e to
g
r
een co
lo
r, th
en,
appl
y
ca
nny
e
d
ge
det
ect
i
o
n
al
go
ri
t
h
m
on i
t
.
(C)
(B)
(A
)
Fi
gu
re
9.
Pr
oce
ss o
f
pre
p
a
r
i
n
g
a bi
nary
i
m
age (
A
) t
h
e m
a
in im
age (B) E
u
cl
idean
d
i
stan
ce relativ
e
to
pu
re
gree
n
c
o
l
o
r (C
)
ap
pl
i
cat
i
on of
canny
e
d
ge det
ect
i
on.
El
im
i
n
at
e part
s wi
t
h
dense e
d
ge an
d ap
pl
y
EHT al
g
o
ri
t
h
m
on i
t
,
aft
e
r cre
a
t
i
ng a bi
na
ry
im
age. Thi
s
p
r
o
cess is sh
ow
n
i
n
Figu
r
e
9. A
p
p
l
y lin
ear
H
oug
h
tr
an
sf
o
r
m
a
ti
ons o
n
t
h
e bi
nary
i
m
age created in last stage
and fi
nd the lines’ candidates. If there is
a li
n
e
in
th
e cand
i
d
a
tes th
at its a
n
g
l
e with
th
e
ho
rizo
n
t
al ang
l
e o
f
the
ellip
se is ab
o
u
t 9
0
,
and
th
e
d
i
stan
ce b
e
t
w
een
th
e cen
ter
o
f
th
e ellip
se an
d
th
e
p
e
rp
en
d
i
cu
lar lin
e to
th
e
can
d
i
d
a
te lin
e is lo
w, th
e shot is co
n
s
id
ered as th
e
g
a
m
e
start.Wh
en
a goal o
ccu
rs, g
a
m
e
will b
e
in
terru
p
t
ed
for a
few m
i
n
u
tes. Th
is in
terru
p
tion
is
u
s
ed
to
show
t
h
e
hap
p
i
n
e
ss
of a
udi
e
n
ces a
n
d r
e
pl
ay
of
g
o
al
even
t
.
Thus, after a goal shot the
r
e is usually a close-up shot
, audi
e
n
ce sh
ot
and re
pl
ay
sh
ot. The states after goal
o
ccurren
ce are illu
strated
in
Fig
u
re 10
. To
d
e
tect g
o
a
l ev
en
t, we find
th
e p
a
rts th
at at le
ast o
n
e
o
f
th
e
ab
ov
e
m
odes has
bee
n
i
n
t
h
e s
hot
c
h
an
ge.
I
n
t
h
e
c
a
se t
h
e
goal
m
out
h i
s
det
ect
ed
i
n
t
h
e m
i
ddl
e s
hot
a
n
d gam
e
st
art
i
s
det
ect
ed a
f
t
e
r
m
odes i
n
Fi
gu
r
e
1
1
,
we
de
not
e t
h
at
m
i
ddl
e shot
a
s
t
h
e
g
o
al
event
.
6.
E
X
PERI
MEN
T
AL RES
U
L
T
S
We ha
ve us
ed
several s
o
ccer
gam
e
s’ videos
available
on internet to test this approac
h
. T
h
ese vi
deos
are
recorde
d
from
TV c
h
annel and
with t
h
e rate
of
25
f
r
a
m
e
s per
sec
o
nd
. T
h
e
ex
peri
m
e
nt
i
s
com
p
o
s
ed
of
three pa
rts. In
the first pa
rt, we ha
ve eval
uated det
ection of
s
hot bounda
ry. W
e
ha
ve s
e
parate
d vide
o pieces
wi
t
h
ap
p
r
o
x
i
m
at
el
y
3,37
5
fra
m
e
s
m
a
nual
l
y
and
ha
ve com
p
are
d
wi
t
h
ap
p
r
oac
h
es
prese
n
t
e
d i
n
t
h
e
pape
r. T
h
e
r
e
su
lts ar
e show
n in
Tab
l
e
1
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISS
N
:
2088-8708
Goa
l
Detectio
n
in
S
o
ccer
Vi
d
e
o
:
Ro
le-Based Even
ts Detectio
n
Ap
p
r
oa
ch
(F
a
r
sh
a
d
B
a
ya
t)
98
7
Fi
gu
re 1
1
. Di
f
f
e
rent
s
h
ot
t
y
pe st
at
es
aft
e
r go
a
l
occu
rre
nce
.
The e
v
aluation criteria are
de
fined a
s
follows
:
,
(2
1)
In
w
h
ich
shows t
h
e
num
ber of c
o
rrect case
s
,
t
h
e
num
ber
of
m
i
ssed cases an
d
indicates the
num
ber of false cases.As you can see, the recall criterion
and
precision criterion are
acceptable.The results of
evaluating shots’ classification int
o
four classes of fa
r view, middle
view, close and audience
view are
prese
n
t
e
d
i
n
T
a
bl
e
2. T
h
e
res
u
l
t
s
are
desi
rab
l
e an
d sat
i
s
fact
ory
.
The
fi
nal
resul
t
s
of
goal
det
ect
i
o
n
are
s
h
o
w
n
in Ta
ble 3.
As
it can be
see
n
,
the precision c
r
iteri
on is alm
o
st 92%, a
n
d the recall m
easure is 78%.
Tabl
e 1.
E
v
al
u
a
t
i
on of
sh
ot
b
o
u
n
d
ary
det
ect
i
on.
Precision
Recall
Clips
Match
N
O
0.
90
0.
94
31
17
287
21
1
0.
89
0.
94
26
12
215
19
2
0.
87
0.
92
26
14
183
17
3
0.
94
0.
89
10
21
167
16
4
0.
89
0.
91
19
15
161
16
5
0.
93
0.
91
11
14
143
15
6
Tabl
e
2. E
v
al
u
a
t
i
on
of
sh
ot
cl
assi
fi
cat
i
o
n
Precision
Recall
Shot T
y
pe
0.
82
0.
86
61
59
350
L
ong
0.
88
0.
87
34
36
242
Med
i
u
m
0.
87
0.
90
29
21
198
Close
0.
94
0.
88
5
11
84
Audience
Tabl
e 3.
E
v
al
u
a
t
i
on of
g
o
al
d
e
t
ect
i
o
n
6
5
4
3
2
1
Match
N
O
2
4
3
5
3
7
Total Goal
1
2
3
4
3
6
Correct
1
2
0
1
0
1
Miss
0
1
0
0
0
1
False
0.
5
0.
5
1
0.
8
1
0.
86
Recall
1
0.
67
1
1
1
0.
86
Precision
7.
CO
NCL
USI
O
N
A ne
w ap
pr
oa
ch f
o
r s
hot
b
o
u
n
d
a
r
y
separat
i
ng an
d cl
assi
fy
i
ng
was p
r
e
s
ent
e
d i
n
t
h
i
s
pape
r.
We
id
en
tify th
e goal sh
o
t
s b
a
sed
o
n
h
i
gh
-class se
m
a
n
tic u
n
its ex
tracted
fro
m
v
i
d
e
o. As it is sh
own
in
th
e resu
lt
eval
uat
i
o
n sect
i
on, t
h
e al
go
ri
t
h
m
based o
n
t
w
o
feat
u
r
es o
f
di
ssi
m
i
l
a
ri
t
y
of
gree
n col
o
r rat
i
o cha
nges a
n
d
col
o
r
in
ten
s
ity ch
anges and with th
e h
e
l
p
o
f
self-adap
tiv
e t
h
re
s
h
ol
d, t
h
e
res
u
l
t
s
o
f
s
hot
b
o
u
n
d
a
r
y
ha
ve t
h
e
a
d
e
quat
e
precision a
n
d recall.
W
e
can
also det
ect goa
l
events
with pr
o
p
er
p
r
eci
si
o
n
by
creat
i
n
g se
m
a
nt
i
c
uni
t
s
a
nd
by
using sem
a
ntic struct
ure
of the gam
e
. Th
e
goal of
obtained res
u
lts is acce
pted.
REFERE
NC
ES
[1]
S
.
F
.
d
e
S
ous
a
J
ú
nior,
A.
de
A
Araújo
, D.
M
e
no
tti
, "An
overv
iew
of au
tom
a
ti
c
eve
n
t d
e
te
ct
ion
in
s
o
ccer
m
a
t
c
hes
"
,
Applications of
Computer Visio
n
(WACV), 2011
IEEE Workshop on, pp. 31, 38
, 5-7
Jan. 2011.
[2]
T.T
.
Pham
, T.T
.
Trinh, V.H. Vo
, N.
Q. L
y
, D.A. Duong, "Event
Retriev
a
l
in Soccer Vid
e
o from
Coarse to Fine
Based on Multi-Modal Approach", Computing and Comm
uni
cation Technolo
g
ies, Research
, Innovation,
an
d
Vision for th
e F
u
ture (RIVF), 20
10 IEEE RIVF I
n
terna
tional Con
f
erence on
, pp.
1, 6
,
1-4
Nov. 20
10.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJECE Vol. 4, No. 6, D
ecem
ber 2014
:
979 – 988
98
8
[3]
J.
Yu,
Y.
He,
K.
Sun,
Z.
Wang,
X.
Wu,
"Semantic Analy
s
is and
Retrieval of Sports Video", Frontier of Computer
Science
and Technolog
y
,
2006
.
FCST '
06. Japan
-
Ch
ina Join
t Workshop on, pp
. 9
7
, 108
, Nov. 200
6.
[4]
T. D’Orazio
,
M. Leo, P. Spagnol
o, M.
Nitti, N. Mosca, A. Distante, “A visual
sy
stem
for real tim
e detection of goal
events dur
ing so
ccer
m
a
tch
e
s”, C
o
m
puter Visi
on
and Im
age
Unde
rstanding, 113(5
)
: 622–632, 200
9.
[5]
A. Ekin, A
.
M
.
Teka
lp, R
.
M
e
hr
otra, “
A
utom
ati
c
s
o
ccer vid
e
o an
al
y
s
is
and s
u
m
m
ariz
ation
”
, I
EEE
Trans
a
c
tions
on
Image Processin
g
, 12(7): 796–80
7, 2003
.
[6]
S. Yang, W. Xiangming,
S. Yong, Z. Liang
,
Y. Lelin
, L. Haitao,
"A Scoreboard Based Method for Goal Events
Detecting in Football Videos",
Dig
ital Media and Digital Content Manage
ment (DMDC
M
), 20
11 Workshop
o
n
,
pp. 248
, 251
, 15
-16 May
2011
.
[7]
M. Hung, C. Hsieh, "Ev
e
nt D
e
tection
of
Broad
c
ast Baseball Vi
d
e
os", Circuits and
S
y
st
ems for Video Technolog
y
,
IEEE Tr
ansactio
ns on , vol. 18,
n
o
. 12
, pp
. 1713
,
1726, Dec. 2008
.
[8]
P
.
S
h
i, Y. Xiao-qing, "Goal Eve
n
t Detect
ion in S
o
ccer
Videos Using Multi-Clues De
tect
ion Rules", Managem
e
nt
and Service Science, 2009. MASS '
0
9
.
Internatio
nal C
onf
eren
ce
on, vol., no., pp
.
1, 4
,
20-22
Sept. 2009.
[9]
L.H. Yong, H.
Tingting
,
"Integ
rating Multip
le
Feature
Fusion for Sem
a
ntic E
v
ent Dete
ction i
n
Soccer Video
"
,
Artificial Intellig
ence, 2009
. JCA
I
'
09. Intern
ation
a
l Join
t Conference on
, pp
. 128
,
131, 25-26
April 2009.
[10]
M
.
H. Kolekar
,
K. P
a
lan
i
appan
,
"A hierarch
ic
al f
r
am
ework for s
e
m
a
ntic s
c
ene
cla
s
s
i
fication
in s
o
c
cer s
ports
vid
e
o"
,
TENCON 2008
- 2008 IEEE Region 10 Conf
er
ence, pp. 1, 6, 19-
21 Nov. 2008
.
[11]
X. Gao, X. Tan
g
, "Unsupervised vide
o-shot
seg
m
entation and model-free an
ch
orperson detection for news video
stor
y
p
a
rsing",
Circuits
and S
y
s
t
ems for Video
Techno
lo
g
y
, I
E
EE Tr
ansactions
on , vol. 12
, no
. 9, pp
. 765
, 77
6,
Sep. 2002
.
[12]
Y.
Yang,
S.
Lin, Y.
Zhang,
S.
Tang,
"Highlights
extra
c
tion
in soc
cer vid
e
os
based
on goal-mouth
detection"
, Sign
al
Processing and
Its Applications,
2007.
ISSPA 2007. 9th
International S
y
mpos
ium on, pp. 1, 4,
12-15 Feb. 2007.
[13]
A. Fitzg
i
bbon,
M. Pilu, R.B
.
Fisher, "Direct
least squa
re fitting of
ellipses", Pa
ttern Analy
s
is and Machin
e
Intelligen
ce, IEEE Tr
ansa
ctions o
n
, vol. 21, no.
5, pp. 476, 480, May
1999.
[14]
W. Jianping, "
R
obust Real-T
i
m
e
Ellipse Det
ect
ion b
y
Di
rect L
east-Square-Fitting"
, Com
puter Sci
e
nce
an
d
Software Eng
i
neering, 2008 In
ter
n
ation
a
l Conf
erence on
, vol. 1
,
p
p
. 923
, 927
, 12-
14 Dec. 2008
.
[15]
T. J. Atherton,
D. J. Ke
rby
s
on
, “Size invarian
t circ
le detection”, Image and Vi
sion Computing, 17(11): 795–803,
1999.
[16]
X. Yu, H.W. Leon., C. Xu, Q. Tian
, "A robust
Hough-base
d algorithm for partial ellipse detection in broadcas
t
soccer video"
,
Multim
edia and
Expo, 2004. IC
ME '
04. 2004 I
EEE Int
e
rna
tion
a
l Conferen
ce o
n
, vol. 3, no
., p
p
.
1555, 1558
, Vol. 3, 30-30 June 2
004.
Evaluation Warning : The document was created with Spire.PDF for Python.