Inter
national
J
our
nal
of
Electrical
and
Computer
Engineering
(IJECE)
V
ol.
10,
No.
3,
June
2020,
pp.
2375
2382
ISSN:
2088-8708,
DOI:
10.11591/ijece.v10i3.pp2375-2382
r
2375
De
v
elopment
of
ster
eo
matching
algorithm
based
on
sum
of
absolute
RGB
color
differ
ences
and
gradient
matching
Rostam
Affendi
Hamzah
1
,
M.
G.
Y
eou
W
ei
2
,
N.
Syahrim
Nik
Anwar
3
1
F
akulti
T
eknologi
K
ejuruteraan
Elektrik
and
Elektronik
,
Uni
v
ersiti
T
eknikal
Malysia
Melaka,
Malaysi
a
2,3
F
akulti
K
ejuruteraan
Elektrik,
Uni
v
ersiti
T
eknikal
Malysia
Melaka,
Malaysia
Article
Inf
o
Article
history:
Recei
v
ed
Dec
19,
2017
Re
vised
Dec
11,
2019
Accepted
Dec
18,
2019
K
eyw
ords:
Bilateral
filtering
Computer
vision
Gradient
matching
Stereo
matching
Stereo
vision
ABSTRA
CT
This
article
presents
local-based
stereo
matching
algorithm
whic
h
comprises
a
de
v
el-
opment
of
an
algorithm
using
block
matching
and
tw
o
edge
preserving
filters
in
the
frame
w
ork.
Fundamentally
,
the
matching
process
consists
of
se
v
eral
stages
which
will
produce
the
disparity
or
depth
map.
The
problem
and
mos
t
challenging
w
ork
for
matching
process
is
to
get
an
accurate
corresponding
point
between
tw
o
images.
Hence,
this
article
proposes
an
algorithm
for
stereo
matching
using
impro
v
ed
Sum
of
Absolute
RGB
Dif
ferences
(SAD),
gradient
matching
and
edge
preserving
filters.
It
is
Bilateral
Filter
(B
F)
to
sur
ge
up
the
accurac
y
.
The
SAD
and
gradient
matching
will
be
im
plemented
at
the
first
sta
ge
to
get
the
preliminary
corresponding
result,
then
the
BF
w
orks
as
an
edge-preserving
filter
to
rem
o
v
e
the
noise
from
the
first
stage.
The
second
BF
is
used
at
the
last
stage
to
impro
v
e
final
disparity
map
and
increase
the
object
boundaries.
The
e
xperimental
analysis
and
v
alidation
are
using
the
Mid-
dleb
ury
standard
benchmarking
e
v
aluation
system.
Based
on
the
results,
the
proposed
w
ork
is
capable
to
increa
se
the
accurac
y
and
to
preserv
e
the
object
edges.
T
o
mak
e
the
proposed
w
ork
more
reliable
w
ith
current
a
v
ailable
methods,
the
quantitati
v
e
measure-
ment
has
been
made
to
compare
with
other
e
xisting
methods
and
it
sho
ws
the
proposed
w
ork
in
this
article
perform
much
better
.
Copyright
c
2020
Insitute
of
Advanced
Engineeering
and
Science
.
All
rights
r
eserved.
Corresponding
A
uthor:
Rostam
Af
fendi
Hamzah,
F
akulti
T
eknologi
K
ejuruteraan
Elektrik
and
Elektronik,
Uni
v
ersiti
T
eknikal
Malaysia
Melaka,
Malaysia.
Email:
rostamaf
fendi@utem.edu.my
1.
INTR
ODUCTION
Computer
vision
is
interdisciplinary
fiel
d
that
comprises
methods
for
acquiring,
proces
sing
and
analyzing
and
image
understanding
from
digital
images
or
videos.
It
is
artificial
intelligence
to
mimic
the
human
visual
system.
Stereo
vision
is
a
part
of
them
and
the
process
to
get
the
information
such
as
object
detection,
recognition
and
depth
es
timation
is
called
as
stereo
matching.
This
process
starts
with
corresponding
from
one
point
on
reference
image
to
another
point
on
the
tar
get
image.
These
images
can
be
tw
o
or
more.
In
this
article,
the
images
are
using
from
the
stereo
camera
input
which
is
also
kno
wn
as
stereo
images.
The
matching
algorithm
from
the
matching
process
produces
disparity
map.
This
map
consists
of
depth
information
which
is
v
aluable
for
man
y
applications
such
as
virtual
reality
[1],
3D
surf
ace
reconstruction
[2],
f
ace
recognition
[3]
and
robotics
automation
[4-5].
The
stereo
baseline
can
be
setup
in
a
wide
or
short
baseline
[6]
dista
nce
which
depends
on
the
applications.
T
o
determine
the
range
or
distance
estimation,
the
triangulation
function
is
appli
ed
to
each
of
the
pix
el
on
the
disparity
map.
Therefore,
to
get
an
accu-
rate
result,
the
matching
process
requires
com
ple
x
and
challenging
solution
for
depth
or
distance
estimation.
It
requires
precise
function
on
the
propose
frame
w
ork.
Fundamentally
,
matching
algorithm
consists
of
multiple
J
ournal
homepage:
http://ijece
.iaescor
e
.com/inde
x.php/IJECE
Evaluation Warning : The document was created with Spire.PDF for Python.
2376
r
ISSN:
2088-8708
stages
which
w
as
proposed
by
Szeliski
and
Scharstein
[7].
First
stage,
matching
cost
computes
the
preliminary
matching
point
of
stereo
image.
Second
stage,
the
filtering
is
utilized
to
reduce
the
preliminary
noise
of
the
first
stage.
Then,
disparity
selection
and
optimization
stage
normalizes
the
disparity
v
alue
each
pix
el
on
the
image.
Last
stage
is
to
refine
the
final
result
and
also
kno
wn
as
disparity
map
post-processing
step.
In
stereo
matching
de
v
elopment,
there
are
tw
o
major
approaches
a
v
ailable
in
de
v
eloping
the
a
lgorithm
frame
w
ork.
It
is
local
methods
as
published
in
[8-10]
and
global
method
[11].
Mostly
local
methods
use
local
properties
or
local
contents
using
windo
ws-based
technique
such
as
fix
ed
windo
ws
implemented
in
[12-13],
adapti
v
e
windo
w
[14],
con
v
olution
neural
netw
ork
[15]
and
multiple
windo
ws
[16].
In
common,
W
inner
-
T
ak
es-All
(WT
A)
strate
gy
is
applied
for
local
based
optimization.
It
is
lo
w
computational
comple
xity
and
f
ast
e
x
ecution
time
[17-19].
Local
method
such
i
mplemented
in
[20]
that
used
plane
fitting
technique
to
increase
the
accurac
y
at
the
final
stage.
This
method
also
kno
wn
as
RANSA
C
that
ef
ficiently
w
orks
on
the
lo
w
te
xtured
areas.
Ho
we
v
er
,
the
error
still
occurred
on
the
object
edges.
Their
method
requires
se
v
eral
iterations
for
plane
fitting
process.
If
wrong
iterations,
then
it
will
af
fect
the
results.
Commonly
,
local
methods
sho
w
f
ast
running
time,
b
ut
lo
w
accurac
y
on
the
edges
due
to
improper
sele
ction
of
windo
ws
sizes.
Hence,
to
get
an
accurate
result
for
the
local
approach
is
a
challenge
to
the
researchers.
Another
approach
in
stereo
matching
algorithm
to
produce
the
disparity
map
is
global
opti
mization
method.
Fundamentally
,
this
method
uses
ener
gy-based
function
which
is
kno
wn
as
Mark
o
v
Random
Field
(MRF).
The
method
in
global
optimization
such
as
Belief
Propag
ation
(BP)
[21]
and
Graph
Cut
(GC)
[22]
produce
accurate
result
s.
Each
pix
el
of
interest
calculation
requires
all
pix
el’
s
ener
gy
in
dispar
-
ity
map.
It
calculates
neighboring
or
nearby
pix
els
using
maximum
flo
w
and
the
selection
is
made
based
on
the
minimum
cut-of
f
ener
gy
on
the
disparity
map.
The
algorithms
implemented
using
global
optimization
approach
normally
in
v
olv
e
high
computational
requirement
due
to
all
pix
el’
s
ener
gy
calculation
and
absorp-
tion.
Global
methods
in
v
olv
e
iterations
which
increase
the
e
x
ecution
time
each
disparity
map
reconstruction.
This
article
aims
to
produce
accurate
results
and
competiti
v
e
with
some
established
methods.
The
first
function
or
stage
will
be
implemented
using
impro
v
ed
Sum
of
Absolute
Dif
ferences
(SAD)
[23]
with
gradient
match-
ing.
Then,
the
second
stage
utilizes
the
edge
preserving
filter
which
is
kno
wn
as
Bilateral
Filter
(BF)
[24].
This
filter
is
capable
t
o
remo
v
e
noise
and
preserv
ed
object
edges.
The
third
stage
is
optimization
based
on
WT
A
strate
gy
.
Last
stage,
the
BF
is
applied
once
ag
ain
to
remo
v
e
unw
anted
or
remaining
in
v
alid
pix
els.
The
BF
is
also
capable
to
increase
the
accurac
y
at
object
boundaries.
2.
RESEARCH
METHOD
The
diagram
of
the
proposed
w
ork
is
dispalyed
by
Figure
1.
The
stereo
matching
algorithm
starts
with
STEP
1
to
get
the
preliminary
disparity
map.
The
impro
v
ed
SAD
has
been
proposed
which
the
weighted
technique
is
used
on
the
block
matching
process.
The
combination
of
impro
v
ed
SAD
with
gradient
match-
ing
in
this
article
should
be
able
to
increase
the
ef
fecti
v
eness
of
corresponding
process
and
accurac
y
.
Then
at
STEP
2,
the
BP
is
utilized
to
reduce
the
noise
and
preserv
ed
the
object
edges.
The
BP
is
capable
to
ef
ficiently
remo
v
e
noise
on
the
lo
w
te
xture
re
gions
and
sharping
the
object
boundaries.
The
optimization
uses
WT
A
strate
gy
which
this
method
normalizes
the
floating
point
s
numbers
and
selects
minimum
disparity
v
alues
on
the
disparity
map.
Final
stage
at
STEP
4
is
also
using
the
BP
b
ut
with
the
disparity
v
alues.
This
filter
is
a
type
of
nonlinear
filter
and
capable
to
impro
v
e
final
disparity
map.
Figure
1.
A
flo
wchart
of
the
proposed
algorithm.
Int
J
Elec
&
Comp
Eng,
V
ol.
10,
No.
3,
June
2020
:
2375
–
2382
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Elec
&
Comp
Eng
ISSN:
2088-8708
r
2377
2.1.
Matching
cost
computation
The
first
stage
of
the
proposed
frame
w
ork
is
using
the
weighted
SAD.
The
preliminary
disparity
map
is
produced
at
this
stage.
Hence,
rob
ust
function
must
be
used
to
increase
the
ef
fecti
v
eness
on
the
dispar
-
ity
map.
The
problem
on
matching
process
at
this
stage
on
the
lo
w
te
xture
re
gions
must
be
at
minimum.
The
weight
is
proposed
at
SAD
to
impro
v
e
the
v
alues
on
the
lo
w
te
xture
re
gions.
Thus,
the
consistenc
y
of
the
weight
at
t
he
lo
w
te
xture
re
gion
is
capable
to
mak
e
the
matching
process
accurate
and
reduces
the
mismatch
or
in
v
alid
pix
els.
The
RGB
v
alues
are
used
with
the
weight
of
sum
of
intensity
dif
ferences
on
right
image
I
r
and
left
image
I
l
which
is
gi
v
en
by
(1):
SAD
(
x;
y
;
d
)
=
1
W
X
(
x;y
)
2
w
j
I
i
l
(
x;
y
)
i
I
i
r
(
x
d;
y
)
j
(1)
where
(
x;
y
)
are
the
coordinates
pix
el
of
interest
with
d
represents
the
disparity
v
alue,
W
is
the
proposed
weight,
RGB
channels
numbers
are
i
and
w
represents
k
ernel
of
SAD
algorithm.
The
second
part
is
gradient
matching
components.
It
contains
the
magnitude
dif
ferences
from
each
image.
There
will
be
tw
o
directions
that
need
to
be
calculated
on
this
gradient
dif
ferences.
V
ertical
direction
G
y
and
horizontal
direction
of
G
x
are
the
directions
with
the
equations
are
gi
v
en
by
(3)
and
(2):
G
y
=
2
4
1
0
1
3
5
I
m
(2)
G
x
=
1
0
1
I
m
(3)
where
I
m
is
input
image
and
represents
con
v
olution
operation
on
the
gradient
matching.
The
G
x
and
G
y
are
the
gradient
magnitude
for
m
which
is
gi
v
en
by
(4):
m
=
q
G
2
x
+
G
2
y
(4)
(5)
is
the
gradient
matching
k
ernel
G
(
x;
y
;
d
)
.
G
(
x;
y
;
d
)
=
j
m
l
(
x;
y
)
m
r
(
x
d;
y
)
j
(5)
The
matching
cost
function
at
this
stage
is
gi
v
en
by
(6)
where
the
input
v
ol
ume
of
S
AD
(
x;
y
,
d
)
and
G
(
x;
y
,
d
)
are
combined
together
.
M
C
(
x;
y
;
d
)
=
S
AD
(
x;
y
;
d
)
+
G
(
x;
y
;
d
)
(6)
2.2.
Cost
aggr
egation
This
second
stage
more
lik
ely
to
filter
the
preliminary
disparity
map
from
stage
one.
Normally
the
preliminary
disparity
map
contains
high
noise
and
it
must
be
remo
v
ed.
Some
of
in
v
alid
and
uncertainties
pix
els
are
constructed
during
the
matching
process.
Hence,
at
this
stage
the
filter
must
be
rob
ust
and
is
capable
to
remo
v
e
high
noise
of
in
v
alid
pix
els
and
preserv
ed
the
object
boundaries.
The
BP
is
used
due
to
strong
preserving
object
edges
a
nd
at
the
same
time
ef
ficient
to
remo
v
e
high
noise
especially
on
the
plain
color
and
lo
w
te
xture
re
gions.
(7)
is
the
BF
function
used
in
this
article.
W
B
F
(
p;
q
)
=
X
q
2
w
B
exp
j
p
q
j
2
2
s
exp
j
I
p
I
q
j
2
2
c
(7)
where
p
is
the
location
pix
el
of
interest
at
(
x
,
y
),
w
B
and
q
are
windo
w
size
of
BF
and
neighboring
pix
els
respecti
v
ely
.
The
s
denotes
a
f
actor
of
spatial
adjustment
and
c
equals
to
similarity
f
actor
for
the
color
detection.
The
p
q
is
spatial
Euclidean
interv
al
and
j
I
p
I
q
j
denotes
the
Euclidean
distance
in
color
space.
Hence,
(8)
is
the
cost
aggre
g
ation
function
of
BF
with
the
matching
cost
computation
input.
C
(
p;
d
)
=
W
B
F
(
p;
q
)
M
C
(
p;
d
)
(8)
De
velopment
of
ster
eo
matc
hing
algorithm
based
on...
(Rostam
Af
fendi
Hamzah)
Evaluation Warning : The document was created with Spire.PDF for Python.
2378
r
ISSN:
2088-8708
2.3.
Disparity
optimization
This
stage
optimizes
the
disparity
v
alues
on
disparity
map.
The
normalization
is
based
on
the
minimum
disparity
v
alues
with
the
floating-point
number
which
the
WT
A
is
selected
in
this
article.
The
WT
A
is
normally
being
used
in
the
local
based
methods
due
to
f
ast
im
plementation.
The
WT
A
function
is
gi
v
en
by
(9).
d
x;y
=
ar
g
min
d
2
D
C
(
p;
d
)
(9)
where
D
represents
a
set
of
v
alid
disparity
v
alues
for
an
image
and
C
(
p;
d
)
denotes
the
second
stage
of
aggre
g
ation
step.
Fundamentally
,
after
this
stage
the
disparity
map
still
contains
noise
or
in
v
alid
pix
els.
Thus,
this
map
needs
to
be
impro
v
ed
and
the
last
stage
is
will
remo
v
e
remaining
noise.
2.4.
Disparity
r
efinement
The
last
st
age
of
the
algorithm
frame
w
ork
is
kno
wn
as
refinement
or
post
processing
stage.
It
has
se
v
eral
continuous
processes
which
starts
with
handling
the
occlusion
re
gions,
filling
the
in
v
alid
pix
els
and
filtering
final
disparity
map.
The
left-right
consistenc
y
checking
process
is
conducted
to
identify
occlusion
areas
and
some
in
v
alid
pix
els.
Then,
these
in
v
alid
pix
els
are
restored
with
v
alid
pix
el
v
alues
through
the
filling
process.
Some
of
artif
acts
and
unw
anted
pix
els
will
be
remo
v
ed
using
the
BF
and
at
the
same
time
preserv
ed
the
object
boundaries.
The
BF
smoothes
the
final
disparity
map
as
indicates
by
(7).
3.
RESUL
T
AND
AN
AL
YSIS
This
section
e
xplains
about
the
disparity
map
results
that
will
be
represented
by
color
-scale
intensity
.
The
dif
ferent
color
tones
sho
w
that
the
respected
objects
are
mapped
based
on
the
dispari
ty
v
alues
and
the
distance
sensor
(i.e.,
stereo
camera).
Most
probably
the
lighter
intensity
v
olume
indicates
that
the
object
is
closer
to
the
sensor
.
The
e
xperimental
analysis
has
been
e
x
ecuted
on
a
personal
computer
with
W
indo
ws
10,
3.2GHz
and
8G
RAM.
The
input
images
are
from
the
Middleb
ury
stereo
e
v
aluation
dataset
[24]
which
contains
15
standard
images
and
must
be
submitted
online.
These
images
are
v
ery
comple
x,
and
each
image
consists
of
dif
ferent
characteristics
and
properties
such
as
light
settings
objects
depth,
incoherence
re
gions,
dif
ferent
resolutions
and
lo
w
te
xture
areas.
The
v
alues
of
f
w
;
s
;
c
;
w
B
g
are
f
9
x
9
;
17
;
0
:
4
;
11
x
11
g
.
Figure
2
sho
ws
a
sample
Jadeplant
image
(i.e.,
left
and
right)
from
the
Middleb
ury
training
dataset
with
dif
ferent
brightness
and
hi
gh
contrast.
Generally
,
due
to
the
brightness
dif
ference,
these
input
images
are
v
ery
challenging
to
be
matched.
It
contains
dif
ferent
pix
el
v
alues
at
the
same
corresponding
point.
Ho
we
v
er
,
the
proposed
algorithm
is
correctly
disco
v
ered
the
disparity
locations.
The
le
v
el
of
disparity
contour
are
precisely
assigned
and
object
distance
are
well-recognized.
Figure
3
sho
ws
the
final
disparity
map
results
of
15
training
images
from
the
Middleb
ury
dataset.
The
accurac
y
attrib
utes
for
error
e
v
aluation
are
nonocc
(non-occluded)
and
all
error
.
The
nonocc
error
is
the
error
e
v
aluation
based
on
t
he
non-occluded
re
gions
on
dis
p
a
rity
map
while
all
error
represents
the
all
pix
els’
e
v
aluation
on
an
image
of
disparity
map.
W
ithin
these
15
images,
Pipes
and
Jadeplant
images
are
the
most
dif
ficult
images
to
be
matched.
These
images
comprise
se
v
eral
piping
lines
and
lea
v
es
with
dif
ferent
sizes
respecti
v
ely
.
Y
et,
the
propose
algorithm
can
reconstruct
almost
accurate
disparity
map
with
clear
discontinuities
re
gions.
Fundamentally
,
real
images
from
the
Middleb
ury
are
dif
ficult
and
v
ery
challenging
to
get
an
accurate
corresponding
point.
It
w
as
de
v
eloped
to
test
the
rob
ustness
of
an
algorithm
where
same
corresponding
point
maybe
contains
dif
ferent
pix
el
v
alues.
Additionally
,
each
image
contains
dif
ference
characteristics
such
as
plain
color
objects,
shado
w
,
discontinuity
re
gions
and
occluded
areas.
W
ith
referring
to
Figure
3,
the
disparity
maps
of
lo
w
te
xture
surf
aces
such
as
Motorc
ycle,
Motorc
y-
cleP
,
Playtable
and
PlaytableP
are
well
recreated
with
dif
ferent
depth
and
disparity
contour
.
Other
re
gions
dif
ficult
to
be
matched
are
plain
colour
objects
and
shado
w
such
as
images
of
ArtL,
Rec
ycle,
Piano
and
PianoL.
These
re
gions
consist
of
similar
pix
el
v
alues
and
possibility
to
get
wrong
matching
are
v
ery
high.
The
dispa
rity
maps
from
the
proposed
w
ork
display
almost
accurate
matching
for
these
images.
It
sho
ws
that
the
proposed
w
ork
is
able
to
get
correct
matching
pix
els
o
n
these
re
gions
and
rob
ust
ag
ainst
the
plain
colour
areas.
The
quantitati
v
e
measurement
from
the
Middleb
ury
online
results
are
gi
v
en
in
T
ables
1
and
2.
These
results
are
produced
by
the
Middleb
ury
online
benchmarking
e
v
aluation
system
with
tw
o
error
attrib
utes
as
e
xplained
abo
v
e.
Some
established
methods
are
also
included
in
these
T
ables
to
sho
w
the
competiti
v
eness
of
the
proposed
w
ork.
Ov
erall,
an
a
v
erage
error
measurement
is
assessed
to
rank
the
best
results.
F
or
T
able
1,
the
proposed
method
is
rank
ed
at
top
of
the
table
with
6.11%,
and
T
able
2
with
9.15%.
It
sho
ws
that
the
proposed
w
ork
is
competiti
v
e
with
other
recently
published
methods
and
can
be
used
as
a
complete
algorithm.
The
proposed
Int
J
Elec
&
Comp
Eng,
V
ol.
10,
No.
3,
June
2020
:
2375
–
2382
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Elec
&
Comp
Eng
ISSN:
2088-8708
r
2379
w
ork
is
rank
at
top
compared
to
[15-17,19,25,26]
for
nonocc
error
.
The
weight
a
v
erage
error
is
6.11%
where
Jadepl,
Playrm
and
V
intge
images
are
the
lo
west
error
produced.
F
or
the
al
l
error
attrib
ute
in
T
able
2,
the
proposed
w
ork
is
produced
at
9.15%
which
is
the
lo
west
a
v
erage
error
.
It
sho
ws
that
the
proposed
w
ork
in
this
article
is
competiti
v
e
with
some
established
methods.
Figure
2.
Dif
ferent
brightness
and
contrast
of
the
input
Jadeplant
stereo
images.
Accurate
disparity
map
result
is
produced
by
the
proposed
w
ork
compared
to
the
image
without
the
proposed
w
ork.
The
plant
structures
are
clearly
displayed
and
smooth
disparity
map
can
be
notified
Figure
3.
These
final
disparity
images
are
from
the
Middleb
ury
dataset
which
the
results
tab
ulated
in
T
able
1
and
T
able
2
T
able
1.
The
nonocc
error
results
using
the
Middleb
ury
dataset.
The
comparison
results
with
other
published
methods
Algorithms
Adiron
ArtL
Jadepl
Motor
MotorE
Piano
PianoL
Pipes
Playrm
Playt
PlayP
Rec
yc
Shelvs
T
eddy
V
intge
W
eight
A
v
e
Proposed
Algorithm
3.66
4.23
19.54
3.11
3.87
5.01
11.02
6.32
5.11
24.85
5.10
3.33
7.52
2.21
7.90
6.11
SNCC
[19]
2.89
4.05
18.10
2.68
2.52
3.52
7.08
6.14
5.64
45.40
3.13
2.90
7.59
1.58
13.50
6.97
ELAS
[26]
3.09
4.72
29.70
3.28
3.29
4.30
8.31
5.61
6.00
21.80
2.84
3.09
9.00
2.36
10.90
7.22
MPSV
[15]
3.83
6.00
19.70
5.85
5.53
5.68
34.30
9.59
5.86
15.30
4.20
4.59
13.00
3.70
14.30
8.81
ADSM
[16]
13.30
6.10
15.00
3.67
5.67
7.08
20.60
6.57
13.20
23.10
3.55
5.76
17.20
3.05
10.10
8.95
DoGGuided
[17]
15.20
9.57
27.10
5.64
8.31
8.09
32.40
9.67
14.00
24.50
5.32
5.56
16.20
4.15
15.00
12.00
BSM
[25]
7.27
11.40
30.50
6.67
6.52
10.80
32.10
10.50
12.50
24.40
12.80
7.42
16.40
4.88
32.80
13.40
De
velopment
of
ster
eo
matc
hing
algorithm
based
on...
(Rostam
Af
fendi
Hamzah)
Evaluation Warning : The document was created with Spire.PDF for Python.
2380
r
ISSN:
2088-8708
T
able
2.
The
al
l
error
results
using
the
Middleb
ury
dataset.
These
comparisons
sho
w
the
competiti
v
eness
of
the
proposed
method
Algorithms
Adiron
ArtL
Jadepl
Motor
MotorE
Piano
PianoL
Pipes
Playrm
Playt
PlayP
Rec
yc
Shelvs
T
eddy
V
intge
W
eight
A
v
e
Proposed
Algorithm
4.74
7.45
27.21
5.57
5.54
6.01
11.64
11.77
7.31
27.05
8.54
3.86
9.21
3.33
8.87
9.15
SNCC
[19]
3.63
6.78
39.80
5.12
5.11
4.65
8.23
11.80
8.05
45.60
4.36
3.29
8.10
2.55
14.80
10.40
ELAS
[26]
4.08
7.18
52.80
5.39
5.45
4.96
9.00
10.70
7.94
23.20
3.83
3.78
9.46
3.34
11.60
10.60
ADSM
[16]
14.30
10.60
34.10
6.00
8.00
7.37
20.40
12.10
16.90
25.50
5.84
5.83
17.20
4.11
11.10
12.30
MPSV
[15]
5.87
9.43
40.20
9.11
8.80
7.03
34.20
15.80
8.58
16.90
5.89
6.78
13.70
4.82
16.80
12.70
DoGGuided
[17]
20.10
28.00
56.50
13.80
16.80
13.40
37.30
23.80
30.30
30.80
13.00
9.13
19.00
13.40
23.60
22.30
BSM
[25]
12.70
2
8.70
58.70
14.80
14.70
16.00
35.80
24.50
29.40
31.00
20.20
12.10
19.20
14.30
39.30
23.50
T
o
v
erify
the
potentiality
of
the
proposed
algorithm,
the
images
from
the
KITTI
[27]
are
also
tested.
These
images
are
more
dif
ficult
and
challenging
to
be
matched.
It
contains
comple
x
edges
and
structures
such
as
shado
w
,
plain
color
surf
aces,
high
dif
ferent
contrast
and
brightness
areas
with
lar
ge
unte
xtured
re
gions.
The
e
xperimental
results
are
sho
wn
in
Figure
4.
The
disparity
map
results
sho
w
accurate
disparity
v
alues
estimation
in
grayscale.
As
for
reference,
the
signage,
a
c
yclist,
trees
and
cars,
are
well-reconstructed
with
correct
disparity
le
v
el.
It
sho
ws
the
proposed
w
ork
in
this
article
capable
to
w
ork
with
dif
ficult
stereo
images
from
real
en
vironment.
Figure
4.
The
disparity
map
results
from
the
KITTI
training
dataset.
This
article
utilizes
images
from
the
number
of
#000004
10-#000007
10
4.
CONCLUSION
In
this
w
ork,
the
combination
of
SAD
algorithm
based
RGB
color
and
gradient
matching
are
producing
accurate
results.
The
second
stage
where
edge
preserving
filter
is
utilized.
The
BF
at
the
aggre
g
ation
stage
is
capable
to
filter
high
noise
and
conserv
e
the
object
bound
a
ries
of
the
preliminary
disparity
map.
The
WT
A
strate
gy
w
as
implemented
at
the
optimization
stage
to
normalize
the
floating
points
numbers
to
the
disparity
v
alues.
The
second
edge
preserving
filter
w
as
used
at
the
last
stage
of
the
proposed
w
ork
using
the
same
BF
.
This
nonlinear
filter
remo
v
ed
remaining
noise
and
increase
the
ef
ficienc
y
of
final
disparity
map.
Ov
erall,
these
edge
preserving
filters
used
in
the
proposed
frame
w
ork
were
able
to
rem
o
v
e
noise
especially
on
the
lo
w
te
xture
re
gions
and
able
to
preserv
e
the
object
edges
as
sho
wn
by
Figure
2.
The
quantitati
v
e
measurement
from
the
standard
benchmarking
Middleb
ury
system
also
demonstrated
lo
w
a
v
erage
errors
were
produced
by
the
proposed
frame
w
ork
at
6.11%
and
9.15%
of
non-occluded
and
all
pix
el
errors
respecti
v
ely
.
The
training
images
are
sho
wn
by
Figure
3.
From
real
images
of
the
KITTI,
the
proposed
w
ork
w
as
also
demonstrated
accurate
results.
Int
J
Elec
&
Comp
Eng,
V
ol.
10,
No.
3,
June
2020
:
2375
–
2382
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Elec
&
Comp
Eng
ISSN:
2088-8708
r
2381
A
CKNO
WLEDGEMENT
This
project
w
as
sponsored
by
a
grant
from
the
Uni
v
ersiti
T
eknikal
Malaysia
Melaka
with
the
Nu
m
ber:
JURN
AL/2018/FTK/Q00008.
REFERENCES
[1]
S.
F
.
Gani,
et
al.,
”De
v
elopment
of
portable
automatic
number
plate
recognition
(ANPR)
system
on
Rasp-
berry
Pi,
”
International
Journal
of
Electrical
and
Computer
Engineering
,
v
ol.
9,
pp.
1805–1813,
2019.
[2]
R.
A.
Hamzah,
et
al.,
”Stereo
Matching
Algorithm
for
3D
Surf
ace
Reconstruction
Based
on
T
riangulation
Principle,
”
in
International
Conference
on
Information
T
echnology
,
Information
Systems
and
Electrical
Engineering
(ICITISEE),
,
2016,
pp.
119–124,
2016.
[3]
S.F
.
Gani,
et
al.,
”De
v
elopment
of
a
portable
community
video
surv
eillance
system,
International
Journal
of
Electrical
&
Computer
Engineering
,
v
ol.
9,
pp.
1814–1821,
2019.
[4]
D
Scharstein,
et
al.,
”A
T
axonomy
and
Ev
aluation
of
Dense
T
w
o-frame
Stereo
Correspondence
Algo-
rithms,
”
in
Stereo
and
Multi-Baseline
V
ision,
2001.(SMBV
2001).
Proceedings.
IEEE
W
orkshop
on
,
2001,
pp.
131–140.
[5]
R.
Christian,
et
al.,
”Dens
e
W
ide-Baseline
Scene
Flo
w
From
T
w
o
Handheld
V
ideo
Camera
s,
”
IEEE
T
rans-
action
on
Circuit
and
Systems
for
V
ideo
T
echnology
,
v
ol.
26,
pp.
1–18,
2016.
[6]
Q.
Y
ang,
”A
Non-local
cost
aggre
g
ation
Method
for
Stereo
Matching,
”
in
IEEE
Conference
on
Computer
Vision
and
Pattern
Recognition
(CVPR)
,
2012,
pp.
1402–1409.
[7]
H.
Asmaa,
et
al.,
”F
ast
Cost-v
olume
Filtering
for
V
isual
Correspondence
and
Be
yond,
”
IEEE
T
ransactions
on
P
attern
Analysis
and
Machine
Intelligence
,
v
ol.
35,
pp.
504–511,
2015.
[8]
H.
Ibrahim,
et
al.,
”Stereo
matching
algorithm
based
on
illumination
control
to
impro
v
e
the
accurac
y
,
”
Image
Analysis
&
Stereology
,
v
ol.
35,
pp.
39–52,
2016.
[9]
L.
Qian,
et
al.,
”Stereo
Matching
Algorithm
Based
on
Ground
Control
Points
Using
Graph
Cut,
”
in
Image
and
Signal
Processing
(CISP),
2014
7th
International
Congress
on
,
2014,
pp.
503–508.
[10]
W
.
Sih-Sian,
et
al.,
”Ef
ficient
Hardw
are
Architecture
for
Lar
ge
Disparity
Range
Stereo
Matching
Based
on
Belief
Propag
ation,
”
in
Signal
Processing
Systems
(SiPS),
2016
IEEE
International
W
orkshop
on
,
2016,
pp.
236–241.
[11]
H.
Hirschm
¨
uller
,
et
al.,
”Real-time
Correlation-based
Stereo
V
ision
with
Reduced
Border
Errors,
Interna-
tional
Journal
of
Computer
V
ision
,
v
ol.
47,
pp.
1-3,
2002.
[12]
K.
Jedrzej,
et
al.,
”Real-time
Stereo
Matching
on
CUD
A
Us
ing
an
Iterati
v
e
Refinement
Method
for
Adap-
ti
v
e
Support-weight
Correspondences,
”
IEEE
T
ransactions
on
Circuits
and
Systems
for
V
ideo
T
echnology
,
v
ol.
23,
pp.
94–104,
2013.
[13]
Q.
Y
ang,
et
al.,
”F
ast
Stereo
Matching
Using
Adapti
v
e
Guided
Filtering,
”
Image
and
V
ision
Computing
,
v
ol.
32,
pp.
202–211,
2014.
[14]
J.
Zbontar
and
Y
.
LeCun,
”Computing
The
Stereo
Matching
Cost
with
a
Con
v
olution
Neural
Netw
ork,
”
in
IEEE
Conference
on
Computer
V
ision
and
P
attern
Recognition
,
2015,
pp.
1592–1599.
[15]
B.
Jean-Charles
et
al.,
”Morphological
Processing
of
Stereoscopic
Image
Superimpositions
for
Disparity
Map
Estimation,
”
Hal-01330139
,
pp.
1–17,
2016.
[16]
M.
Bing
et
al.,
”Accurate
Dense
Stereo
Matching
Based
on
Image
Se
gmentation
Using
an
Adapti
v
e
Multi-Cost
Approach,
Symmetry
,
v
ol.
8,
p.
159,2016.
[17]
K.
Masamichi,
et
al.,
”High
Accurac
y
Local
Stereo
Matching
Using
DoG
Scal
e
Map,
”in
Fifteenth
IAPR
International
Conference
on
Machine
V
ision
Applications
,2017,
pp.
258–261.
[18]
W
.
Hu,
et
al.,
”V
irtual
Support
W
indo
ws
for
Adapti
v
e-weight
Stereo
Matching,
”
in
V
isual
Communica-
tions
and
Image
Processing
(VCIP),
2011
IEEE
,
2011,
pp.
1–4.
[19]
E.
Nils
and
E.
Julian,
”Anisotropic
Median
Filtering
for
Stereo
Disparity
Map
Refinement.,
”
in
VISAPP
(2)
,
2013,
pp.
189–198.
[20]
K.A.A.
Aziz,
et
al.,
”A
pix
el
to
pix
el
correspondence
and
re
gion
of
interest
in
stereo
vision
application”,
in
2012
IEEE
Symposium
on
Computers
&
Informatics
(ISCI)
,
pp.
193–197,
2012.
[21]
M.G.Y
.
W
ei,
et
al.,
”Stereo
matching
based
on
absolute
dif
ferences
for
multiple
objects
detection,
”
T
elk
omnika
,
v
ol.
17,
pp.
261-267,
2019.
[22]
A.
H.
Hasan,
et
al.,
”Disparity
Mapping
for
Na
vig
ation
of
Stereo
V
ision
Autonomous
Guided
V
ehicle,
”
in
Soft
Computing
and
P
attern
Recognition,
2009.
SOCP
AR’09.
International
Conference
of
,
2009,
pp.
De
velopment
of
ster
eo
matc
hing
algorithm
based
on...
(Rostam
Af
fendi
Hamzah)
Evaluation Warning : The document was created with Spire.PDF for Python.
2382
r
ISSN:
2088-8708
575–579.
[23]
R.
A.
Hamzah
et
al.,
”V
isualization
of
image
distortion
on
camera
calibration
for
stereo
vision
applica-
tion,
”
in
Control
System,
Computing
and
Engineering
(ICCSCE),
2012
IEEE
Inter
national
Conference
on
,
2012,
pp.
28–33.
[24]
S.
Daniel
and
S.
Richard,
”
Middleb
ury
Stereo
Ev
aluation
-
V
ersion
3
(Accessed
date
:
No
v
ember
2019,
http://vision.middleb
ury
.edu/stereo/e
v
al/references.,
”
2019.
[25]
K.
Zhang,
et
al.,
”Binary
Stereo
Matching,
”
in
P
attern
Recognition
(ICPR),
2012
2
1s
t
International
Con-
ference
on
,
2012,
pp.
356–359.
[26]
A.
Geiger
et
al.,
”Ef
ficient
Lar
ge-scale
Stereo
Matching,
”
in
Asian
Conference
on
Computer
V
ision
,
2010,
pp.
25–38.
[27]
M.
Menze
and
G.
Andreas,
”Object
Sflo
w
for
Autonomous
V
ehicles,
”
in
IEEE
Conference
on
Computer
Vision
and
Pattern
Recognition
(CVPR)
,
2015,
pp.
3061–3070.
BIOGRAPHIES
OF
A
UTHORS
Rostam
Affendi
Hamzah
w
as
graduated
from
Uni
v
ersiti
T
eknologi
Malaysia
where
he
recei
v
ed
his
B.Eng
majoring
in
Electronic
Engineering.
Then
he
recei
v
ed
his
M.
Sc.
majoring
in
Electronic
System
Design
engineering
from
the
Uni
v
ersiti
Sains
Malaysia
in
2010.
In
2017,
he
recei
v
ed
PhD
majoring
in
Electronic
Imaging
from
Uni
v
ersiti
Sains
Malaysia.
Currently
he
is
a
lecturer
in
the
Uni
v
ersiti
T
eknikal
Malaysia
Melaka
teaching
digital
electronics,
digital
image
processing
and
embedded
system.
Melvin
Gan
Y
eou
W
ei
w
as
born
in
1992,
Melvin
Gan
Y
eou
W
ei
graduated
from
Uni
v
ersiti
T
eknikal
Malaysia
Melaka
where
he
recei
v
ed
his
B.Eng
majoring
in
Electrical
in
2017.
Curre
ntly
,
he
is
pursuing
a
Master
De
gree
in
the
Uni
v
ersiti
T
eknikal
Malaysia
Melaka.
Nik
Syahrim
Nik
Anwar
w
as
born
in
1981,
Nik
Syahrim
graduated
from
Uni
v
ersity
of
Applied
Science
Heilbronn,
German
y
where
he
recei
v
ed
his
Diplom
in
Mechatronik
und
Mikrosystemtechnik
majoring
in
Mechatronics
in
2006.
In
2010
he
recei
v
ed
his
M.
Sc.
majoring
in
Mechatronics
from
the
Uni
v
ersity
of
Applied
Science
Aachen,
German
y
.
In
2018,
he
recei
v
ed
PhD
majoring
in
Electrical
from
the
Uni
v
ersiti
Sains
Malaysia.
Currently
he
is
a
lecturer
in
Uni
v
ersiti
T
eknikal
Malaysia
Melaka
teaching
Electrical
and
Mechatronics
subjects.
Int
J
Elec
&
Comp
Eng,
V
ol.
10,
No.
3,
June
2020
:
2375
–
2382
Evaluation Warning : The document was created with Spire.PDF for Python.