TELKOM
NIKA
, Vol. 11, No. 10, Octobe
r 2013, pp. 5
965 ~ 5
972
ISSN: 2302-4
046
5965
Re
cei
v
ed Ap
ril 28, 2013; Revi
sed
Jul
y
8, 2013; Accept
ed Jul
y
19, 2
013
Weighted Multi-Scale Image Matching Based on Harris-
Sift Descriptor
Can Sun, Jin-ge
Wang,
Zaixin Liu, Junmin Li*
Schoo
l of Mechan
ical En
gi
ne
erin
g and A
u
to
mation
Xi
hu
a Univ
ersit
y
, Ch
eng
du, C
h
in
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: lijunm
in
197
5
@
16
3.com
A
b
st
r
a
ct
Accordi
ng to t
he rotati
on
al
i
n
vari
ance
of
Harri
s c
o
rner
detector
and
the ro
bustn
ess
of Sift
descri
p
tor. An
i
m
pr
ove
d
H
a
rri
s-Sift corner
de
scriptor w
a
s
pr
opos
ed. At first
,
the a
l
gor
ith
m
give
n mu
lti-sca
le
strategy to H
a
r
r
is corn
er, i
m
p
r
oved c
o
rn
er c
ounti
ng
metho
d
an
d re
move
d red
u
n
dant
p
o
ints at th
e sa
me
time, th
en, th
e
corn
er w
a
s d
i
rectly a
ppl
ied
to low
-
p
a
ss Ga
ussia
n
filter
i
m
age. B
a
se
d o
n
the
histogr
a
m
of
Sift feature d
e
scriptor, ge
ner
ates a
new
12
8-di
me
nsi
ona
l
feature v
e
ctor
descriptor
by m
u
lti-scale
Gaus
s
w
e
ighted. T
h
ro
ugh th
e ab
ove,
Harris cor
ner
detector a
nd Si
ft descriptor w
a
s nor
mali
z
e
d i
n
the sca
le l
a
y
e
r
and gr
adi
ent features. T
he e
x
peri
m
e
n
t resu
lts indic
a
t
ed th
at, the impr
ov
ed corn
er des
criptor co
mpris
e
d
both a
d
va
ntag
e of Harr
is c
o
rner
detecti
o
n
an
d Sift fe
ature d
e
script
o
r. T
he
meth
od re
duc
ed t
h
e
computati
on ti
me
an
d the fa
lse
match r
a
te
, w
h
ich c
oul
d
be va
lid
ly a
ppl
ied to th
e ro
b
o
t stereo vis
i
o
n
match
i
n
g
an
d three-
di
me
nsio
nal rec
onstructi
on.
Ke
y
w
ords
: stereo visi
on, rob
o
t, corner det
e
c
tion, feature d
e
scriptor, scal
e
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Comp
uter
st
ereo vi
sion t
e
ch
niqu
e ma
inly includ
es five parts [1]: image a
c
quisitio
n
,
feature extra
c
tion, came
ra
calibration,
stereo
m
a
tchin
g
an
d 3
D
recon
s
tru
c
tion. In these
module
s
, the
co
rrect m
a
tch
of differe
nt image
fea
t
ureis ne
ce
ssary for ro
bot
co
gni
ze
spa
c
e
obje
c
ts, this p
r
ocess is
call
ed feature p
o
i
n
ts matc
hing
[2]. The impact of the light in the pairin
g
of
the feature
p
o
ints, the
correct d
e
scripti
on of t
he fe
ature p
o
ints i
n
different
scale spa
c
e an
d
image
rotatio
n
chan
ge
s b
a
s
e
co
ordi
nate
syste
m
a
r
e
a
ll as the i
nevi
t
able role i
n
i
m
age
matchi
ng
p
r
oc
es
s
.
For the i
n
flue
nce
of the ab
ove factors o
n
t
he matchi
ng, feature
p
o
ints d
e
scri
b
ed in the
pre
s
ent m
e
thod
comp
ri
ses: Diffe
renti
a
l-ba
se
d
de
scripto
r
[3], Descripto
r
b
a
sed on
mom
ent
invariant
s [4], Di
stributio
n-based d
e
scri
p
tor. Ko
e
nde
rink ma
de a
deep
re
se
arch on th
e ima
ge
local differen
t
ial prope
rtie
s, and puts
forwa
r
d the
con
c
e
p
t of scale pa
ram
e
ter [5]. Free
man
spe
c
ified
a controlla
ble di
rection
differe
ntial filter
[6], the filter effectively avoide
d the impa
ct
of
image rotatio
n
feature p
o
i
n
t matchin
g
. Mikolaj
c
zyk.
K made a
co
mpre
hen
sive
summ
ary of the
descri
p
tor al
gorithm [7] who foun
d that Lowe
'
s
Sift operator [8] with a strong sta
b
ility, the
algorith
m
bui
lt DOG scal
e spa
c
e
at first, got
ima
ge esse
ntial
feature
s
characteri
stics. T
hen
cal
c
ulate
d
th
e extrem
um
point by
usin
g Hessia
n m
a
trix, giveng
radient
attribu
t
e to extrem
um
point ma
de it
become ve
ctor feature p
o
ints, fina
lly, pre
c
ise
classi
fied histo
g
ra
m inform
atio
n of
each feature
point, it ensuring the uni
q
u
ene
ss of feat
ure poi
nt matchin
g
.
It is inconve
n
i
ent to cal
c
ula
t
e the He
ssia
n matrix in Sift point detecti
on, 128
-dime
n
sio
nal
corne
r
featu
r
e vecto
r
de
scriptor
nee
d a
long timeto j
udge. It re
sul
t
ing in the Sif
t
algorithm
do
es
not have the real
-time of determi
n
e
co
rner an
d endo
ws the ability
to the descri
p
tor. This pa
per
combi
ned the
Harris
corne
r
detect
o
r to the Sift
algorithm, first, expande
d scale
spa
c
e o
n
Ha
rris
corne
r
dete
c
tion, complet
ed co
nsi
s
ten
c
y of Ha
rris corner d
e
te
ction an
d Sift in the scale
para
m
eter,
b
e
ca
use the
Harri
s
co
rne
r
h
a
s
rotatio
n
in
varian
ce, it
can b
e
com
p
l
e
ted
repla
c
e
the
Sift extreme
points in the
detectio
n
p
r
o
c
e
ss,
by u
s
in
g the Sift op
e
r
ator give the
gra
d
ient fe
ature
to the Harris
corne
r
directly
, built a sta
b
l
e
feat
ure poi
nts which d
o
e
s n
o
t vary
with image
rota
tion,
scaling,
affine tra
n
sfo
r
ma
tion, illuminat
ion
ch
a
nge
s and
othe
r f
a
ctors. An
d
gene
rated
p
o
int
matchin
g
fused image
with image fusi
o
n
algorith
m
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : 596
5 –
5972
5966
2. Detect Interest Point
In ord
e
r to e
x
tract the
re
q
u
ired
poi
nts,
sele
cted
Harris
corne
r
d
e
tector ba
se
d
on g
r
ay
level ch
ang
e
s
. Detect ave
r
age
en
ergy
cha
nge
s in
the la
rge
r
ima
ge a
r
ea
whil
e the d
e
tecto
r
i
s
runni
ng. Ifit cha
nge
s at
x, y direction
,
con
s
ide
r
ed
to be the
corne
r
,if only
in one
dire
ct
ion
cha
nge
s gre
a
tly, conside
r
ed to be the edge featur
e, if
the cha
nge
s in both
directio
n is not
obviou
s
, for image
smo
o
thi
ng regi
on.
2.1. Impro
v
e
d
Harris Cor
n
er Scale Sp
ace
Harri
s
co
rne
r
is determi
ne
d only by the pr
a
c
tical ex
perie
nce of the thre
sh
old
to judge
wheth
e
r metri
c
functio
n
wa
s a co
rne
r
or
not.
Corn
er a
u
toco
rrelation
matrix [9] as
follows
:
2
2
2
,,
,,
,,
xD
x
y
D
ND
D
N
xy
D
Y
D
LL
L
Mg
LL
L
,
(1)
whe
r
e
is
a Ga
ussian win
d
o
w
po
sition,
,
xD
L
,
,
y
D
L
re
spe
c
tively are
hori
z
o
n
ta
l and
vertical
differential op
erato
r
at a
po
sitio
n
,
,
xy
D
LL
is
per-pix
el ope
ratio
n
s.
Type (1)
ha
d
alrea
d
y repl
ace the o
r
i
g
inal Harris Gauss
win
dow into di
screte 2
D
zero
-mea
n G
auss
function
22
2
ex
p
2
N
xy
g
,
, defining
0
n
N
,
0
is the initial scale value,
0
1.
0
,
n
is the ratio of
the adja
c
e
n
t scale
s
, norm
a
lly
1.4
,
n
for scal
e layers,gen
e
r
ally take
7
n
,
M
as theauto
c
o
rrel
a
tion matri
x
of target pointwithin the G
aussia
n
wind
ow.
By repla
c
in
g
multi-scal
e G
aussia
n
filter
N
g
,
, the
Ha
rri
s d
e
t
ector with
si
ngle
scale
become m
u
lti-scale
sp
atial dete
c
torwi
t
h linear
gro
w
th which i
n
com
p
a
r
iso
n
of the scale
s
p
ac
ea
sp
ec
t.
2.2. Define
d Harris
Corne
r
Resp
onse
Function
It can get the corn
er respo
n
sefu
nctio
n
[10]
by calculating the ch
a
r
acte
ri
stic val
ues
1
,
2
of the correlat
ion matrix
M
:
2
de
t
RM
k
t
r
a
c
e
M
(2)
coeffici
ent
0.04
,
0
.06
k
, th
e value di
re
ctly impact on
the re
sult of functio
n
re
sp
o
n
se. Ino
r
de
r
to stabili
zeco
rne
r
amo
unt
in each tim
e
se
rie
s
sa
mple pi
cture
s
, improved
corne
r
re
sp
o
n
se
algorith
m
as f
o
llow:
2
22
22
22
xy
xy
xy
LL
LL
R
k
trac
e
M
cim
k
t
r
ac
e
M
LL
(3)
R> 0, exp
r
e
s
sed
a
s
a
corner fe
ature,
whe
n
R
<0,
expre
s
sed
as edg
e featu
r
e
,
whe
n
R i
s
v
e
ry
small, sm
oot
h area that is within the wi
ndo
w.
trac
e
M
is the trace of
M
.
12
tra
c
e
M
as corne
r
det
ection pro
c
e
s
s
data whi
c
h descri
be
the
l
o
cal fe
ature
s
of an imag
e.
Figure 1 (a) i
s
a
test image, Fi
gure 1
(b) to t
he co
rrespon
ding tra
c
e im
age.
2.3. Remov
e
Redu
ndan
t
Points
Figure 1
(b) p
r
odu
ce
d du
pli
c
ate traces
when t
he
scale
spa
c
e i
n
cre
a
s
e, the
existe
nce
of
repe
at trace
s
reacte
d dupl
icate dete
c
tio
n
of co
rne
r
, optimizatio
n eliminate re
d
unda
nt features
point as follo
ws:
2
ma
x
,
|
2
ii
k
k
i
xx
x
x
x
(4)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Weig
hted Mul
t
i-Scale Im
age Matchin
g
Base
d on Harri
s
-Sift De
scrip
t
or (Junm
in Li)
5967
Figure 1(a
)
. T
e
st image
Figure 1(b
)
. T
r
ace imag
e
Equation (4) can avoid
im
age
min
u
tiae
detecti
o
n
fai
l
ure in l
a
rg
e-scaleeffe
ctively and
detect l
a
rge
range
of th
e i
m
age
featur
e
point
in th
e
small
scale,
which
i
x
,
k
x
corre
s
p
ond to
the
low scal
e and
high scale G
aussia
n
sp
ace.
The a
bove i
m
provem
ents Ha
rri
s alg
o
rithmgiven
th
e ch
aracte
ristics of m
u
lti-scale to
detecto
r, it provides a
c
curate corne
r
poi
nt
s to image f
eature extract
i
on and ima
g
e matchin
g
.
3. Impro
v
ed
Harris
-
Sift
Corner Descri
p
tor
Sift algorith
m
mainly i
n
clud
es DOG
layere
d
sca
l
e spa
c
e
co
nstru
c
tion,
image
pixel
extreme poi
n
t
detection, determi
ne th
e scale in
variant feature
points, remo
ve the unsta
ble
feature point
s,
dete
r
mine
the
interest p
r
inci
pal
dire
ct
ion of
reliabl
e point
s, ge
n
e
rating
feature
descri
p
tor. A
pply this fea
t
ure mat
c
hin
g
algo
rithm
can
effective
l
y overco
me
imaging
pla
ne
transl
a
tion, scalin
g, rotatio
n
and affine
transfo
rmatio
n in the image
s matching
3.1. Crea
te G
r
adient
Char
acte
r
istics
The Sift de
scriptor u
s
ed
n
on-m
a
xima
suppressi
o
n
al
gorithm
[11]
comp
ared all
extreme
points in the
same
scal
e l
a
yer a
nd th
e
upp
er-lower
scale
l
a
yer, resultin
g
in a large
am
ount
o
f
cal
c
ulatio
n. In orde
r to avoid the probl
em of real
-time lack in Sift algorithm, we take multi-scale
Harri
s
co
rne
r
instea
d of Sift extreme points, us
i
ng
sift key point
allocatio
n
method, cal
c
ul
ate
gradi
ent mag
n
itude an
d direction of ne
w Harris p
o
int:
22
-1
,1
,
1
,
,
1
,
1
,1
,
1
,t
a
n
1,
1,
m
x
y
L
x
y
Lx
y
L
x
y
Lx
y
Lx
y
L
x
y
xy
Lx
y
L
x
y
(5)
(5) i
s
the
co
rner g
r
a
d
ient
attribute. Noti
ce that
,
L
xy
for e
a
ch l
o
w-pa
ss Gau
ssi
an filter ima
g
e
of the key poi
nt, the improved Harris
corner
will be key point in the
cal
c
ulation, while the
scale
of
improve
d
Ha
rris poi
nt as the key point scale laye
r. Then sam
p
le in the feature p
o
ints within t
he
neigh
borhoo
d
wind
ow an
d
create
s
a
gradient di
re
cti
on hi
stog
ram
of
the featu
r
e point
s in th
is
wind
ow, we sele
cted imp
r
ove Ha
rri
s
corne
r
Gau
s
sian
wind
ow
N
g
,
, the windo
w size
wa
staken to
be 1.5
time
s
N
. In the
hi
stog
ram
whi
c
h
ex
ists
80%
rem
a
ining
of m
a
i
n
pe
ak g
r
ay
colum
n
, it can define the
di
rection of t
h
is gradient t
o
the auxilia
ry direction of
the feature point.
Each
feature
point
with o
n
l
y
one
main
d
i
rectio
n,
b
u
t there
i
s
m
o
re
than o
n
e
aux
iliary di
re
ctio
n.
Finally, make
each
key poi
nt get three chara
c
te
risti
c
s: scale, di
recti
on and p
o
siti
on.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : 596
5 –
5972
5968
3.2. Genera
te Ne
w
Harris
-
Sift Fe
atur
e Des
c
riptor
Ensu
re the
ro
tation invaria
n
ce
of the d
e
s
cripto
r at first. Select
22
Gau
ssi
an wind
ow
on the
poi
nt,
gradi
ent m
a
g
n
itude
and
di
rectio
n in
e
a
ch wi
ndo
w
wa
s
cal
c
ulate
d
by usi
ng fo
rm
ula
(5). Sift al
go
rithm
re
com
m
ende
d
44
Ga
ussian
wi
ndo
w o
n
the
poi
nt, we
ca
n i
n
trodu
ce
a
Gau
s
s di
strib
u
tion fun
c
tion
, weig
hteddi
stancefe
a
ture
points,the
n
, e
a
ch
of the
ke
y points in
clu
d
e
128 dime
nsi
o
n, normali
zed
the 128-dim
ensi
onal vect
or, whi
c
h ge
n
e
rate ne
w Ha
rri
s-Sift feature
point de
script
or.
3.3. The Matching Criterion of Image
Feature Poin
ts
Gene
rally the
coinci
den
ce
rate of differe
nt picture
s
in
same sce
n
e
is more tha
n
50%,
usin
g simpl
e
thre
shol
d-based matchi
ng:
22
p
q
mm
(6)
p
m
,
q
m
respe
c
tively are the
de
scripto
r
di
sta
n
ce
fo
r two f
eature
point
s in the
same
scene
obje
c
ts
whi
c
h
corre
s
p
ondin
g
to the set o
f
12
3
,
,
...
n
P
pp
p
p
,
12
3
,,
.
.
.
n
Qq
q
q
q
,
is de
script
ortwo
-
dimen
s
ion
a
l Euclide
an di
stance,
0.4
,
0.6
.
Thus,fe
ature
point de
script
or can be m
a
tched
wi
th different sets of
points at the
same
scale inva
ria
n
t and rotatio
n
invaria
n
t, and to distin
g
u
ish fal
s
e m
a
tchin
g
point
s effectively. The
algorith
m
blo
ck di
agram of
binocular visual f
eature p
o
i
nt descripto
r
is sh
own in Figure 2.
Figure 2. Fea
t
ure matching
algorithm di
a
g
ram
4. Experimenta
t
ion
Usi
ng MATL
AB program
ming environ
ment to
wo
rking the al
g
o
rithm in Co
re E550
0
2.8GHZ dual
-core CP
U, 2GB RAM co
mputer. Fig
u
re 3 is the bi
nocular
cam
e
ra a
s
the im
age
captu
r
e devi
c
e. Each pie
c
e
ofexperime
n
tal image
s pixel ratio is 64
0
×
48
0, bit depth is 8bit.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Weig
hted Mul
t
i-Scale Im
age Matchin
g
Base
d on Harri
s
-Sift De
scrip
t
or (Junm
in Li)
5969
Figure 3. Binocul
ar visio
n
system
4.1. Camera
Calibra
tion
Calib
rate the
cam
e
ra
at first, the p
a
ra
meters obtai
ned by pl
ane
calib
ration t
a
rget [12
]
cal
c
ulatio
n:
0
0
0
8
3
1
.88690
0
3
2
1
.20991
0
0
830.19
993
241.56118
00
1
0
0
1
x
y
fu
Ff
v
12
3
1.
95
1
.
78
-
0
.15
Rr
r
r
-
23.
24
-
2
7
.
42
225.
31
xy
z
Tt
t
t
12
-0
.
4
2
0
.
1
9
Kk
k
12
-
0
.000
3
-
0.00
05
Pp
p
12
0.
0004
0.
0003
F
for the matrix of c
a
me
ra intrins
i
c
parameters
,
x
f
,
y
f
respectively i
s
compon
ent of
a len
s
scale focal le
ngth abo
ut x-axis and y-ax
i
s
in the imag
e coo
r
din
a
te system,
0
u
,
0
v
for the cente
r
coo
r
din
a
te of the image.
R
,
T
are the exte
rnal ca
mera p
a
ram
e
ter mat
r
ix.
K
,
P
,
for three
dist
ort
i
o
n
co
e
f
f
i
cient
s of
ca
mera int
r
i
n
si
c
.
4.2. Diffe
ren
t
Attrib
ute
s
o
f
Single Image Matching
Comp
ared wi
th the SIFT algorithm, mat
c
hin
g
tw
o pi
ctures fe
ature
point which
d
i
fferent
in the b
r
ig
htn
e
ss u
nevenn
ess a
nd pixel
ratio
re
spe
c
ti
vely, taking
0.6
in (6
) to
en
su
re that i
s
able to dete
c
t a stable feat
ure poi
nt.
Test matchin
g
effect of two image
s for lu
minan
ce u
n
e
venne
ss in the sam
e
sce
ne, th
e
pro
c
e
ssi
ng result in Fig
u
re 4 (a), Fig
u
r
e 4 (b
) an
d Table 1
com
pare
d
two m
e
thod
s of ma
tch
time,
the
n
u
m
ber
of co
rn
er dete
c
tion, the
nu
mbe
r
o
f
feature
poi
n
t
s mat
c
hin
g
a
nd the
mat
c
h
i
ng
rate.
Test m
a
tchi
n
g
effect
of two imag
es for
diffe
rent pixel
ratio i
n
the
same
scene
u
s
ing
the
same
way, the pro
c
e
s
sin
g
result in Figure
5
(a),
Figure 5 (b
) and table 2
compa
r
e
d
two
method
s.
The co
mpa
r
ative data in Table 1
and Table
2 shows
matchin
g
st
ability
o
f
improve
d
feat
ure
point
s, redu
ced th
e
compl
e
xity
of com
puting t
i
me, improve
d
the d
e
tecti
o
n
pre
c
isi
on of t
he corner poi
nts an
d the
numbe
r of
co
rre
ct
ly
pai
red
,
des
cri
p
t
o
r h
a
s well st
abili
t
y
and re
al-time.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : 596
5 –
5972
5970
Figure 4 (a). Sift descri
p
tor
in illumination changes m
a
tch
Figure 4 (b). Sift descri
p
tor
in illumination changes m
a
tch
Table 1. Algo
rithm co
mpa
r
i
s
on
Descr
iptor T
i
me
(s
)
Corne
r
amount
Pair
ing
number
Matching
rate
Sift 9.8
1825
586
90.0%
Improve
4.2
1136
347
90.5%
Figure 5(a
)
. Sift descriptor i
n
different pixel ratio
Figure 5(b
)
. Improve d
e
scriptor in differe
nt pixel ratio
Table 2. Algo
rithm co
mpa
r
i
s
on
Descr
iptor T
i
me
(s
)
Corne
r
amount
Pair
ing
number
Matching
rate
Sift 12.5
580
423
96.4%
Improve
7.2
317
240
96.9%
4.3. Binocul
ar Visual Matching an
d Fusion
Then, the al
gorithm i
s
a
pplied tobi
no
cula
r ste
r
eo
scopi
c featu
r
e
point match
i
ng, the
results
sho
w
n in Figu
re
6
(a)
and Fi
gu
re 6
(b). In th
e image
mat
c
hin
g
, image
fusion
algo
rithm
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Weig
hted Mul
t
i-Scale Im
age Matchin
g
Base
d on Harri
s
-Sift De
scrip
t
or (Junm
in Li)
5971
[13] wa
s intro
duced, define
d
fusio
n
facto
r
mi
n
m
i
n
ma
x
m
ax
mi
n
m
i
n
,,
,,
xy
x
y
a
xy
x
y
, Computed
the ch
ang
e
pro
c
e
s
s of bo
unda
ry coordi
nates
mi
n
x
,
mi
n
y
to
max
x
,
max
y
, matche
d u
p
the repe
ating
feature
C
betwe
en ima
ge A and image B, C=A
∩
B.
,,
1
,
P
ij
a
j
A
i
j
a
j
B
ij
is the
mixing pointg
r
adatio
n of synthetic imag
e D, Figure
7 is the bino
cula
r visual f
eature
s
of the
image fusi
on
results.
Figure 6 (a
). Sift descri
p
tor with bino
cula
r visual
Figure 6 (b
). Improve de
scriptor
with bin
o
cul
a
r visu
al
Figure 7. Image fusio
n
It can be see
n
that the matching featu
r
e
line
of two image
s by almo
st hori
z
ontal
without
epipol
ar li
ne
con
s
trai
nts. It
ha
s g
ood
re
sults in
th
e
center re
gion
fusio
n
a
bout t
w
o
pictu
r
e
s
.
The
algorith
m
ca
n
be use
d
in the bino
cula
r st
ereo m
a
tchi
n
g
.
5. Conclusio
n
The meth
od t
a
ke
sthe
adva
n
tage of
Ha
rri
s cor
ner dete
c
tion in
qui
ck and
simpl
e
and Sift
feature de
scriptorin accu
rate
an
d
sta
b
l
e.Con
s
id
erin
g the im
pa
ct of scale
sp
ace
and
Ga
uss
weig
hted o
n
comp
utationa
l efficien
cy, e
liminat
e redu
ndant co
rne
r
,
rea
s
o
nabl
e evaluate
fe
ature
point matchi
n
g
crite
r
ion, calibrate the p
a
ram
e
ters of came
ra
so a
s
to corre
c
t imagepi
ncush
i
on
distortio
n
, ma
tch up the re
peating featu
r
e with
ima
g
e
fusion al
go
rithm. The ex
perim
entre
sul
t
s
sho
w
s that
th
e imp
r
oved
Harri
s-Sift feat
ure
poi
nt
de
scripto
r
co
uld effectively
avoid
the
condit
i
on
of false matching within ill
umination, scaling, tr
an
slat
ion, affine transformation
whi
c
h con
s
ist
in
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : 596
5 –
5972
5972
the matchin
g
procedu
re. It is an efficient and sta
b
le algorithm
can b
e
use
d
in robot visi
on
sy
st
em.
Ackn
o
w
l
e
dg
ements
We
woul
d like to app
re
cia
t
e the suppo
rted
bythe
Ed
ucatio
nal Co
mmission of
Sichua
n
Province of China (1
2ZB12
9
), the Ope
n
Re
se
a
r
ch Fu
nd of Key Laborato
r
y of Manufa
c
ture a
n
d
AutomationL
aboratory (X
ihua University, Szjj2012
-022
), Innov
ation Fun
d
of Postgradu
ate
(ycjj20
127
6)
of Xihua Univ
ersity for the first autho
r.
Referen
ces
[1]
Budi
harto W
i
d
odo, Sa
ntoso
Ari, Pur
w
a
n
to Dj
oko, Jazi
die Ac
hmad.
Multipl
e
mov
i
ng o
b
stacl
e
s
avoi
danc
e
of service r
o
b
o
t
usin
g stere
o
vision.
T
E
LKOMNIKA Indo
n
e
sia
n
Jo
urna
l
of Electric
al
Engi
neer
in
g.
2011; 9(3): 4
33-
444.
[2]
Radk
e RJ, An
dra S, Al-Kof
a
h
io, et. al. Ima
ge ch
an
ge d
e
t
e
ction
alg
o
rith
ms A s
y
stem
atic surve
y
.
IEEE
T
r
ansactio
n
on
Imag
e Proces
sing
. 20
05; 14(
3): 294-3
07.
[3]
Yu Da
o-li. Ima
ge Match
i
ng A
l
gorit
hm Base
d
on t
he L
o
ca
l Invari
ant F
eatu
r
es. Master T
h
esis. Anh
u
i:
Dep
a
rtment of computer sci
en
ce and e
n
g
i
ne
erin
g of Anhu
i Univers
i
t
y
; 2
0
1
2
.
[4]
M David. Visi
o
n
. San F
r
ancis
co: F
r
eeman P
ublis
hers. 19
82
.
[5]
JJ Koend
eri
n
k
,
AJ van Doo
r
n. Repres
ent
at
ion of l
o
cal
geom
etr
y
i
n
thevis
ual s
y
ste
m
.
Biolog
ical
Cyber
netics
. 1
987; 55: 3
67-3
75.
[6]
WT
Freeman, EH Adelson. T
he desi
gn and use of steer
able filt
ers. IEEE
T
r
ansactio
n
s on Pattern
Analys
is and M
a
chi
ne Intel
lig
e
n
ce
. 199
1; 13: 891-
906.
[7]
K Mikol
a
jcz
y
k,
T
T
u
y
t
el
aars,
C Schmi
d
, et.
al. A c
o
mp
aris
on
of affin
e
re
gio
n
d
e
tectors.
International
Journ
a
l of Co
mputer Visi
on
. 2
005; 65(
1/2): 4
3
-72.
[8]
DG Lo
w
e
. D
i
s
tinctive
imag
e featur
es fro
m
scale-
i
nv
ari
ant ke
yp
oi
nts.
Internati
o
n
a
l
Jour
nal
o
n
Co
mp
uter Visi
on
. 200
4; 60(2)
: 91-110.
[9]
C H
a
rris, M St
eph
ens.
A
co
mb
in
ed c
o
rn
er
an
d
ed
ge
det
ector
. Proce
e
d
i
ngs
ofthe
4th
Alve
y V
i
sio
n
Confer
ence. 1
988: 14
7-1
51.
[10]
Z
hen-Yi
ng
Xu,
W
en-Bin
Li,
Yun W
a
n
g
, Chu
n
Lu
o, S
hu-Yu
an Ga
o, Dan-D
an
Ca
o. T
r
inocular
calibr
a
tio
n
met
hod
base
d
o
n
binoc
ul
ar
cali
bratio
n.
T
E
LKOMNIKA Indo
nesi
an Jo
urn
a
l
of Electrical
Engi
neer
in
g
. 2012; 10(
6): 143
9-14
44.1
[11]
W
ang Jin
g
, W
ang H
a
i-l
i
a
ng,
Xi
ang Ma
o-s
hen
g,
W
e
i Li-
den
g, Liu Z
h
o
ng-sh
eng. Su
b
p
i
x
el
accurac
y
central loc
a
tio
n
of circle tar
get base
d
on
nonm
a
x
imum
suppr
essio
n
.
Chin
ese Jo
urn
a
l of Scientific
Instrum
e
nt
. 20
12; 33(7): 1
460
-146
8.
[12]
Z
hao
Xue-fen
g
, Z
hon
g
Xi
a
o
-min, L
an Y
i
-hua.
Improv
e
m
ent of T
s
ai'
s
camera
cal
i
b
ratio
n
u
n
d
e
r
copl
anar c
a
li
br
ation
boar
d n
e
a
rl
y par
al
lels to
imagi
ng
pla
ne.
Co
mput
er Eng
i
ne
erin
g an
d D
e
sig
n
. 201
1;
32(3): 10
19-
10
21.
[13]
T
ao Z
h
i-jian
g
, Hua
ng H
ua. F
und
us ima
ges
mosaic b
a
se
d
on impr
oved
harris a
nd SIF
T
algorithm
.
Co
mp
uter Engi
neer
ing a
nd D
e
sig
n
. 201
2; 33(9): 351
0-3
5
1
1
.
Evaluation Warning : The document was created with Spire.PDF for Python.