TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.5, May 2014, pp
. 3491 ~ 35
0
0
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i5.3240
3491
Re
cei
v
ed Ma
y 23, 201
3; Revi
sed
De
ce
m
ber 9, 2013
; Accepte
d
Decem
b
e
r
30, 2013
Lane Detection Based on Object Segmentation and
Piecewise Fitting
Chuny
ang M
u
, Xing Ma*
Institute of Information a
nd C
o
mmunic
a
tio
n
T
e
chnol
og
y, B
e
ifan
g Univ
ersi
t
y
of Natio
nal
iti
e
s,
Yinch
uan 750
0
21,
Chi
n
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: maxi
ngsk
y
@
126.com
A
b
st
r
a
ct
A lane d
e
tectio
n alg
o
rith
m for compl
e
x envir
o
n
ment w
a
s propose
d
. It
w
a
s concer
ned o
n
selecti
n
g
cand
idat
e lan
e
regio
n
by ob
je
ct s
egmentatio
n. T
hen red
u
n
dancy e
d
g
e
s w
e
re extracted
b
y
Sobel o
per
ator.
F
u
rthermore, c
and
idat
e la
ne
mark
ers w
e
re obtai
ne
d by
th
resho
l
d se
lecti
on fro
m
the e
dges. F
i
n
a
lly l
a
n
e
mark
ers w
e
re
detecte
d by
piec
ew
ise fit
t
ing. T
he
pro
pose
d
a
l
g
o
rith
m w
a
s s
i
mul
a
ted
in MAT
L
AB.
Experi
m
ents show
ed that l
a
ne
markers
c
o
uld b
e
det
ecte
d correctly. Pi
ecew
ise li
ne
ar
transformatio
n
in
prepr
ocessi
ng
has enh
anc
ed perfor
m
a
n
ce of
detecti
on
w
h
il
e the
env
iro
n
m
ent w
a
s d
i
m. A
nd
li
mite
d re
gi
on
of interest he
lp
s to identific
ati
on la
ne i
n
an
a
ppro
p
ri
ate r
egi
on, w
h
ich hav
e
the e
ffect of enha
nce
m
e
n
t in
the
spee
d of oper
a
t
ion. F
eature-b
a
sed
meth
od is
usual
ly affected by intens
ity
of imag
e. Sever
a
l char
acteristi
c
s
of roads ne
ed t
o
be cons
id
ere
d
in fu
rther for detectio
n
more
precise
l
y.
Ke
y
w
ords
:
lan
e
detecti
on
, piecew
ise l
i
near transf
o
r
m
ati
on, OT
SU object se
g
m
entatio
n, thres
hol
d
selecti
on, pi
ecew
ise fitting
Copy
right
©
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
Majority of traffic accid
ent
s a
r
e
cau
s
ed
by
error
ope
ration
s o
r
di
straction
s
a
ccordin
g to
the traffic de
partme
n
ts’ st
atistics. Acci
dents
cau
s
e
d
by these
ca
se
s co
uld
be avoided
by
contin
ually monitor the po
sition of a car
within a lan
e
. The ch
alleng
ing pro
b
lem i
s
to detect la
ne
boun
dari
e
s
o
f
road
s. No
w,
Europe, Am
erica an
d Ja
pan have th
rown the
m
sel
v
es into pa
rt lane
detectio
n
sy
stem. Some systems,
su
ch
as th
e
RA
LP
H sy
st
em,
A
u
t
o
V
ue sy
st
em
,
S
t
art
sy
st
e
m
,
AUTO
RA system and ALVINN
system, a
r
e re
pre
s
e
n
ta
tive systems [
1
, 2].
Many resea
r
che
s
have
b
een
develo
p
ed in
this a
r
ea, ho
weve
r,
seve
ral
co
mplicated
nature
con
d
itions
can d
e
ci
sively degrad
e
the
performa
n
ce of lan
e
d
e
tection te
ch
nique
s:
a)
Shado
ws: tre
e
s, buildin
gs
and othe
r vehicl
e
s
proje
c
t shad
ows on
the road, cre
a
ting
false ed
ge
s.
b)
Solar p
o
sitio
n
:
dire
ct su
nlig
ht may satu
rate the a
c
qui
red im
age
s,
or
cau
s
e
spe
c
ula
r
reflexes.
c)
Climate: natu
r
al phe
nome
nal
(such as
fog, rain or s
now) may de
grad
e sig
n
ificantly
the quality of the image
s [3].
Nume
ro
us m
e
thod
s of vision-b
a
sed lan
e
detectio
n
h
a
ve been p
r
o
posed in an a
ttempt to
robu
stly dete
c
t lane
s. The
y
could be
ca
tegori
z
ed in f
eature
-
b
a
sed
and model
-b
ase
d
method.
Feature-b
a
se
d method p
o
s
ition
s
the la
nes’
im
age
s
by detecting
the obviou
s
feature
s
,
su
ch
as lan
e
edg
es.
Ho
u
gh transf
o
rm
is
a m
o
st
p
opula
r
m
e
tho
d
for dete
c
ti
ng a
nd l
o
cating
straig
ht line
s
in digital ima
ges [4], whi
c
h is not
sen
s
i
t
ive to noise.
However, th
e com
putatio
nal
compl
e
xity and stora
ge re
q
u
irem
ents a
r
e
the ma
in bottlene
cks of the
standa
rd Hough tra
n
sfo
r
m
scheme a
ppli
ed in real
-tim
e detectio
n
.
Edge-ba
sed
method
s ha
ve been p
r
opo
sed to u
s
e
straig
ht lines to m
o
d
e
l lan
e
boun
dari
e
s.
For
cu
rved
road
s, mo
re
compl
e
x mo
d
e
ls
su
ch
a
s
B-Spline
s
, B
e
zie
r
Spli
ne
s [5],
para
bola a
n
d
hyperb
o
la [6
, 7] fitting are
often
use
d
to provid
e su
pport. The
accuracy of the
s
e
model
s i
s
de
pend
ed
on th
eir
com
p
lexity. Simpler
m
odel
s d
o
n
o
t
fit lane b
oun
darie
s a
ccu
ra
tely
though they a
r
e more ro
bu
st to noise th
an com
p
lex model
s [8].
For im
provin
g the effici
e
n
t of pe
rformanc
e, lo
we
r a
r
ea
of a
lane im
age i
s
u
s
ually
con
s
id
ere
d
a
s
regio
n
of in
terest
(ROI) [
4
]. And this
region i
s
fu
rth
e
r divid
ed int
o
left and
rig
h
t
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3491 – 35
00
3492
sub
-
regio
n
s
by su
ppo
sin
g
the
width
of
lane
i
s
fixe
d [1, 9]. Se
g
m
enting
ROI will
re
du
ce
the
compl
e
xity of
lane dete
c
tio
n
.
This re
se
arch is con
c
e
r
n
ed o
n
sele
cti
ng
a
n
can
d
i
date
la
ne re
gion by
OTS
U
o
b
je
ct
segm
entation
.
Then
re
dun
dan
cy pote
n
tial ed
ge
pixels a
r
e
extra
c
te
d by Sob
e
l o
perato
r
. Pixel
s
i
n
these
e
dge
s are sele
cted by
statistic
value
s
.
Finally, lane marke
r
s are obtaine
d by piece
w
i
s
e
fitting.
The
re
st of t
h
is
pap
er i
s
orga
nized
as follows. Se
ct
ion 2
de
scrib
e
s
perfo
rma
n
ce
of the
lane dete
c
tio
n
alg
o
rithm. Several expe
rimental
re
sul
t
s a
r
e p
r
e
s
e
n
t
ed to
sup
port the validity
of
this method i
n
se
ction 3, a
nd se
ction 4
con
c
lu
de
s the pape
r.
2. Rese
arch
Metho
d
This metho
d
con
s
i
s
ts
of th
ree
pa
rts.
a)
Image pre
p
ro
ce
ssi
ng.
It co
ntains
colo
r spa
c
e
transfo
rmatio
n, and pie
c
e
w
ise linea
r transfo
rmatio
n. b) Ca
ndidat
e lane region
is obtaine
d by
OTSU se
gm
entation and ROI setting. c) Lan
e
m
a
rkers
dete
c
tion,
whi
c
h
is det
ected
by
Sob
e
l
edge o
perato
r
firstly. Then
candi
date
s
are
sele
cted
by threshold
value, and la
ne marke
r
s a
r
e
piecewi
s
e fitted finally. The flow ch
art o
f
algorit
hm is
sho
w
n in Fig
u
re 1, whi
c
h
will be de
scri
bed
in detail belo
w
.
Figure 1. Flow Ch
art of the Algorithm
2.1. Image Preproc
essin
g
2.1.1. Color Space Tra
n
s
f
ormatio
n
The col
o
r of
pixels are ori
g
inally rep
r
e
s
ented
in RGB spa
c
e that is highly correl
ated [4].
RGB value
s
can b
e
tran
sf
orme
d to YCbCr
col
o
r sp
a
c
e. The m
o
st
visually sig
n
ificant info
rmati
o
n
in the colo
r image is rese
rved in Y co
mpone
nt
of image. So Y comp
one
nt of image is used to
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Lane
Dete
ction Base
d on
Obje
ct Segm
entation an
d Piece
w
ise Fitting (Ch
u
n
y
a
ng Mu)
3493
detect
edg
es of de
si
red
la
ne. And
it al
so ha
s
advant
age
of savin
g
data
sto
r
ag
e an
d
redu
ci
ng
comp
uting time.
2.1.2.
Piecew
i
s
e Linear Trans
f
orma
tion
It’s necessa
ry to transfo
rm an ima
ge
whe
n
the e
n
virome
nt is di
m or the im
a
ge ha
s lo
w
contrast. Pie
c
e
w
ise lin
ear tran
sformati
on (P
LT)
is
adopte
d
to
a
d
justme
nt th
e imag
e
cont
rast.
PL
T
is ch
ar
ac
te
r
i
z
e
d b
y
2
n
p
a
ra
mete
rs
for
n
-1 lin
e segme
n
ts.
If the in
put
lumina
nce
wa
s
, 0
,
1
,
,
n
x
nm
and
the
out
put lumin
a
n
c
e will
be
, 0
,
1
,
,
n
yn
m
, the (
m
-1
)th tr
a
n
sform
function
s
T
m
-1
(
x
) will be:
1
11
1
1
()
()
(
)
()
mm
mm
m
mm
yy
Tx
x
x
y
xx
(1)
Segment poi
nts are
comp
uted ba
sed o
n
the hist
og
ra
m data. In this method, we
choo
se
the minim
u
m
and
maximu
m inp
u
t gray
value a
s
s
e
g
e
ment
point
s. And th
ey def
ine the
segm
ents
of the input data. Then ea
ch inp
u
t seg
m
ent is
map
p
ed to an outp
u
t segm
ent b
a
se
d on a lin
ear
transfo
rmatio
n for the co
rresp
ondi
ng se
gment.
E.g. three
se
gement
s tran
sform
a
tion fu
nction
T
0
(
x
),
T
1
(
x
) a
r
e
T
2
(
x
)
ar
e
s
h
ow
n in
F
i
g
u
r
e
2 (a). Given p
a
ram
e
ters are set as:
00
0,
y
0
x
(
2
)
11
,0
xM
I
N
y
(
3
)
22
,2
5
5
xM
A
X
y
(
4
)
33
255
,
255
xy
(
5
)
Then the PLT
function
T
0
(
x
),
T
1
(
x
) are
T
2
(
x
) are
sho
w
n
in Figure 2
(
b
)
.
(a) Pie
c
e
w
ise
linear tra
n
sfo
r
mation
(b) Pie
c
e
w
ise
linear tra
n
sfo
r
matio in this
method
Figure 2. Piece
w
ise Linea
r Tran
sform
a
tion Fun
c
tion
Figure 3 (a)
and (b)
are
a
n
origi
nal ima
ge an
d the P
T
L-b
a
sed im
age, re
sp
ecti
vely. The
origin
al ima
g
e
ha
s a
lo
w
contra
st, wh
ose detail
ca
nn
ot se
e
clea
rly. The PTL
-
b
a
s
ed
pe
rform
e
nt
result has b
e
en enh
an
ced.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3491 – 35
00
3494
(a) Original image
(b) PLT
-
ba
se
d image
Figure 3. Re
sult of Piecewi
s
e Lin
ear T
r
a
n
sformation
2.2. Candida
te Lan
e
Regi
on
2.2.1. OTSU Segmenta
tio
n
OTSU
se
gme
n
tation alg
o
rit
h
m was
prop
ose
d
by Jap
a
nese
researcher
No
buyu
k
i Otsu i
n
1979. Th
e cri
t
erion i
s
that
maximize th
e
sep
a
ra
bilitiy
of the re
sulta
n
t cla
s
ses i
n
gray level
s
[1
0].
For extended
,
m
-1 discrete thre
sh
old
s
can
divide im
age into
m
cl
asse
s. In this app
roa
c
h,
we
use two thre
shold
s
for se
p
a
rating th
ree
cla
s
ses of lan
e
image.
Let the pixels of a given
pi
cture
be represe
n
ted in
L
gray level
s
[1, 2, …,
L
]. Th
e numb
e
r
of pixels level
I
is denote
d
by
n
i
and the
total numbe
r
of pixels by
N
=
n
1
+
n
2
+…
+
n
L
. In order t
o
simplify the
discu
ssi
on, t
he g
r
ay-l
evel histo
g
ra
m i
s
normali
zed
and
reg
a
rded
as a
pro
bab
ility
distrib
u
tion:
1
,0
,
1
L
i
ii
i
i
n
pp
p
N
(6)
We u
s
e t
w
o thre
shol
ds:
12
1
kk
L
for se
pa
rating t
h
ree
cla
s
se
s,
C
o
for [1, … ,
k
1
],
C
1
for [
k
1
+1, …,
k
2
], and
C
2
for [
k
2
+1, …,
L
]. Then the prob
abilitie
s of class o
c
cu
rre
nce, and the
cla
ss me
an le
vels are:
1
00
1
1
()
(
)
k
ri
i
PC
p
k
(
7
)
2
1
11
2
1
()
(
)
k
ri
ik
PC
p
k
(
8
)
2
21
2
1
2
()
1
(
)
(
)
L
ri
ik
PC
p
k
k
(
9
)
11
00
0
1
1
11
()
(
()
)
|
kk
ri
ii
iP
C
i
p
ik
k
(
1
0
)
22
11
11
2
2
11
1
()
(|
(
)
)
kk
ri
ik
ik
iP
C
i
p
k
k
i
(
1
1
)
22
12
22
11
12
2
(
()
(
)
1
|)
()
(
)
LL
T
ri
ik
ik
kk
iP
C
i
p
kk
i
(
1
2
)
Whe
r
e:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Lane
Dete
ction Base
d on
Obje
ct Segm
entation an
d Piece
w
ise Fitting (Ch
u
n
y
a
ng Mu)
3495
1
1
1
()
k
i
i
ki
p
(
1
3
)
2
1
2
()
k
i
ik
ki
p
(
1
4
)
And the total mean level of
the original p
i
cture i
s
ߤ
்
,
1
()
L
Ti
i
Li
p
(
1
5
)
And there i
s
a relation fo
r choi
ce of
k
1
a
nd
k
2
,
00
1
1
2
2
0
1
2
,1
T
(16
)
The cla
s
s variances a
r
e gi
ven by:
11
22
2
00
0
0
0
11
((
|
)
(
))
kk
ri
ii
iP
i
C
i
p
(
1
7
)
22
11
2
11
22
1
1
1
11
((
|
)
(
))
kk
ri
ik
i
k
iP
i
C
i
p
(
1
8
)
22
22
2
2
22
2
2
11
((
|
)
))
(
LL
ri
ik
i
k
iP
i
C
i
p
(
1
9
)
The crite
r
ion
mea
s
ure
2
B
is then a fu
nct
i
on of two
variabl
es
and
an optim
al set of
threshold
s
1
k
and
2
k
. In ord
e
r to
evaluat
e the th
re
sh
old
k
1
,
k
2
, th
e follo
wing
d
i
scrimin
ant
c
r
iter
io
n
meas
u
r
e
s
ar
e
us
ed
:
22
2
2
2
2
/,
/,
/
BW
T
W
B
T
(20
)
Whe
r
e:
22
2
2
01
1
2
2
Wo
(
2
1
)
22
2
2
02
1
1
02
))
()
((
BT
T
T
(
2
2
)
And
22
1
()
L
Ti
i
T
ip
(23
)
An optimal se
t of threshold
s
1
k
and
2
k
is sele
cted by maximizing
2
B
:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3491 – 35
00
3496
12
12
22
12
1
(,
)
m
a
x
(
,
)
BB
kk
L
kk
k
k
(24
)
Re
sults of obj
ect
segm
enta
t
ion by OTS
U
are
sh
o
w
n in
Figu
re 4
-
Fig
u
re
6. They
a
r
e two,
t
h
ree an
d f
o
u
r
cla
ssif
i
cat
i
o
n
s re
sp
ect
i
v
e
l
y
.
In Figure 4, lane imag
e is represented
in bi
nary. In Figure 4(a
)
, the majo
rity road and
lane ma
rkers
are
con
s
id
ere
d
as o
ne unit,
which is
rep
r
ese
n
ted in
white. While th
e backg
ro
und
is
in bla
ck. In
th
is result, lane
marke
r
s are
mixed wi
th
ro
ad. In Fig
u
re
4(b
)
an
d (c), i
n
tensity of th
e
lane ma
rkers are high
er th
an roa
d
. The
r
efore mo
st
of the lane ma
rke
r
s co
uld b
e
detecte
d. But
the result of lane dete
c
tion
still have so
me bre
a
kpoin
t
.
In
Figu
re 5,
these
te
st scene
s are
divided
i
n
to
three
cla
s
sification
s. And
they a
r
e
rep
r
e
s
ente
d
by three co
nstant
s, e.g. 0,
0.498 an
d 1. The cla
s
sified
model is Z.
01
11
2
22
0,
,
[
1
,
]
0.498
,
,
[
1
,
]
1,
,
[
1,
]
Ci
k
Z
Ci
k
k
Ci
k
L
(25
)
The
se
gment
ation
re
sults i
n
Fig
u
re
5
are mo
re
cle
a
rl
y than th
at in
Figu
re
4. Th
e mo
st
promi
nent of
three
cla
ssifi
cations i
s
that the
lane ma
rkers,
road
and
backg
rou
nd
can b
e
divide
d
hiera
r
chi
c
ally.
Four
classifications of lan
e
image
s are sh
o
w
n in Figure 6. The road is div
i
ded into
several regio
n
s. Th
e mo
re
so
phisti
c
ate
d
cl
assifi
catio
n
re
sult
s a
r
e
helple
s
s for l
ane d
e
tectio
n
in
the next step
s, but wa
stin
g more
co
m
p
utation. Above all, we sele
ct
three
cla
s
sification
s of lane
image in this
method.
(a) S
c
ene 1
(b) S
c
ene 2
(c
) Sce
ne 3
Figure 4. Cla
ssifi
cation
Re
sults of T
w
o
Cla
s
ses
(a) S
c
ene 1
(b) S
c
ene 2
(c
) Sce
ne 3
Figure 5. Cla
ssifi
cation
Re
sults of Th
ree
Classe
s
There is
so
me sh
ado
w on rig
h
t si
de in Figu
re
5(a
)
. Morp
hologi
cal o
p
e
rate
can
enchan
ce th
e pe
rform
a
n
c
e of cla
s
sification resu
lt disturbed by
noisy
o
r
sh
ado
w. Re
sult
s of
dilation an
d e
r
osi
on a
r
e
sh
own i
n
Figu
re
7(a
)
and
(b
). The effectio
n
of sha
d
o
w
o
n
the rig
h
t la
ne
marker i
s
alm
o
st disape
ar
by dilation an
d ero
s
ion.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Lane
Dete
ction Base
d on
Obje
ct Segm
entation an
d Piece
w
ise Fitting (Ch
u
n
y
a
ng Mu)
3497
(a) S
c
ene 1
(b) S
c
ene 2
(c
) Sce
ne 3
Figure 6.
Cl
a
ssif
i
cat
i
on
Re
sult
s of
Fo
ur
Cla
s
s
e
s
(a) Dilation
result
(b) Er
osi
on re
sult
Figure 7. Dila
tion and Ero
s
i
on Re
sult
2.2.2. Appro
p
riate Limite
d Region o
f
Intere
st
Lane
imag
es are
captu
r
e
d
by a
came
ra
whi
c
h ca
n
be situate
d
i
n
sid
e
a
car n
ear re
ar
view mirro
r
. In upp
er
part
of these
lane
image
s a
r
e
s
k
y
, buildings, flyovers
, trees
,
s
t
reet lamps
,
hills, et
c. They have
solid li
ner structure
whi
c
h
could
be detected out by
edge
operator.
And they
might di
sturb
the d
e
tectio
n of
c
andi
dat
es
of lane
m
a
rkers i
n
lo
wer p
a
rt
of im
age
s. Thu
s
, i
t
is
necessa
ry to
set l
o
wer area of l
ane
i
m
age
as
RO
I. An app
rop
r
iate limited
region
of inte
rest
(ALROI) i
s
o
b
tained
by
combine
d
th
e
PLT-ba
se
d i
m
age
b
(
x
,
y
) with ROI and
the
m
odel
Z
(
x
,
y
),
whi
c
h
x
a
nd
y
are co
ordinat
e axes in ima
ge.
(,
)
(
,
)
(,
)
(
(,
)
)
a
A
L
R
OI
x
y
b
x
y
R
OI
x
y
Z
x
y
T
h
(26
)
Whe
r
e,
Th
a
i
s
a th
re
sh
old
for
sele
cting
ca
ndidate
la
ne regio
n
fro
m
the PLT
-
b
a
se
d ima
ge.
The
limited ROI h
e
lps to
give l
ane id
entifica
t
ion in a
n
ap
prop
riate
re
gi
on. Thi
s
p
r
o
c
ess will
have
the
effect of enha
ncem
ent in the spe
ed of op
eration.
Can
d
idate
la
ne
regi
on
s a
r
e
sh
own in
Figu
re
8, A
L
ROI
cove
re
d the
majo
rit
y
lane
markers, and
most of their surroun
ding
environ
m
ent wa
s maske
d
. But barriers in right side
of
Figure 8 (c)
are left. Because thei
r intensity is simi
l
a
r
to lane markers, and
they
are divided into
one unit by O
T
SU se
gment
aion.
2.3. Lane ma
rkers de
te
cti
on
2.3.1. Sobel Edge Oper
ator
Edges a
r
e im
portant fe
atures i
n
a
n
ima
g
e
si
nc
e they
repre
s
e
n
t si
gn
ificant lo
cal
in
tensity
cha
nge
s a
n
d
offer vital cl
ues to
sepa
rate re
gion
s
within a
n
obj
ect o
r
to ide
n
tify chang
es in
illumination [4].
(a) S
c
ene 1
(b) S
c
ene 2
(c
) Sce
ne 3
Figure 8. Can
d
idate La
ne
Regi
on
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3491 – 35
00
3498
There a
r
e m
any ways to
perfo
rm e
dge
dete
c
tion, Robert
s
, Cann
y, Prewitt an
d so o
n
.
We u
s
ed Sob
e
l edge d
e
tection in the limited regio
n
. Result
s are
sh
own in Fig
u
re
9.
Figure 9. Re
sult of Sobel Edge Detectio
n
There are stil
l some fal
s
e
edge
s after
Sobel
edg
e detectio
n
in an app
rop
r
iat
e
limited
regio
n
. The
s
e point
s may
disturb lan
e
markers fi
tting finally. Thre
shol
d sele
ction are
ado
pte
d
to
resolve it.
2.3.2. Thresh
old Selectio
n b
y
Data Statistic
Can
d
idate
of lane ma
rkers a
r
e obt
aine
d by
thre
shol
d sel
e
ctio
n of
Sobel ed
ge i
m
age
s.
By observin
g
lane image, i
n
tensity of lane marke
r
s
a
r
e high
er tha
n
road in the
same
scan li
ne.
And pixel nu
mbers of lan
e
marke
r
s a
r
e less. M
edi
um intensity is refe
r
e
nc
ed fo
r
s
t
a
t
is
tic
a
lly
gene
rated th
resh
old of eve
r
y scanline.
((
,
)
)
x
Th
M
e
dium
R
e
gion
x
n
(27
)
Whe
r
e,
Th
x
is the sele
ction
thresh
old of
x-
th line.
Region
(
x
,
n
) i
s
an image region
from
x
-
n
line to
x+
n
.
Medium
(
R
) is the
me
dium value of
regio
n
R
. An
d then
Th
x
i
s
use
d
to distin
guish po
ssi
bl
e
lane ma
rkers from false e
d
ges.
(
,
)
(
,)
(
(
,)
)
x
CandiLane
x
y
E
dge
x
y
A
LR
O
I
x
y
Th
(28
)
Whe
r
e,
Edge
is
a
bina
ry a
rray
after So
bel e
dge
det
ection,
and
(
x
,
y
)
pr
es
e
n
t
s th
e
po
s
i
tion
o
f
a
pixel. The
di
stinguish p
r
o
c
ess i
s
co
nce
r
ned
on th
e
e
dge
pixels.
T
he inte
nsity o
f
(
x
,
y
) in ALROI
image is
com
pare
d
with th
e statistic dat
a
Th
x
. If the intensity in ed
ge positio
n was larger tha
n
Th
x
,
thi
s
edge pixel is trut
h. Otherwise, it
will be considered
as a
false one, and
C
a
nd
iL
an
e
(
x
,
y
)
will be set 0 accordi
ngly.
Figure 1
0
sh
ows result of
can
d
idate
lan
e
ed
ge
by thresh
old
sele
cti
on. Some
of t
he fal
s
e
edge
pixels
are
remove
d
,
whe
r
e it is in botto
m o
f
the right
si
de. But there are
some
still
remai
ned
du
e to their int
ensity is sim
ilar to la
ne
markers. En
han
ceme
nt o
f
perfo
rman
ce is
need
ed in furt
her.
Figure 10. Edge Re
sult by Thre
sh
old Se
lection
2.3.3. Piecew
i
s
e
Lane M
a
rker Fitting
Piece
w
ise lan
e
marke
r
fitting is wi
dely u
s
ed in a
road
way unde
r a va
riety of compl
i
cated
con
d
ition
s
[3]. In this meth
od, sm
ooth p
i
ece
w
i
s
e p
o
lynomial fun
c
ti
ons
are
used
in re
pre
s
e
n
ting
lane ma
rkers.
2
,
(,
)
,
m
m
ax
bx
c
i
f
x
x
Lane
x
y
dx
e
i
f
x
x
(29
)
Whe
r
e
x
m
re
pre
s
ent
s the bord
e
r bet
we
en nea
r and
far fields. Th
e linear pa
rt of the model
is
use
d
to fit the ne
ar vi
si
on field, whil
e the pa
ra
bo
lic mo
del fits the far field
.
This p
r
op
o
s
ed
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Lane
Dete
ction Base
d on
Obje
ct Segm
entation an
d Piece
w
ise Fitting (Ch
u
n
y
a
ng Mu)
3499
techni
que i
s
robu
st in
the
pre
s
e
n
ce of
noi
se,
shad
ows, la
ck
of
lane p
a
inting
and
ch
ang
e
of
illumination condition
s.
3. Results a
nd Analy
s
is
The la
ne d
e
tection
algo
rithm is
simul
a
ted
in MATL
AB. Result
s
of lane d
e
tection are
pointed o
u
t in the gray ima
ge with yello
w line
s
.
As sho
w
n i
n
Figu
re 1
1
,
we
ca
n see
that
lane
m
a
rkers
ca
n b
e
dete
cted
correctly
althoug
h
the environ
ment are co
mp
licated. The
inten
s
ity of imag
e
in Fig
u
re
11
(a) and
(b)
a
r
e
dim, pe
rform
ance of
dete
c
tion
ha
s b
e
en e
nha
nced
by PLT
-
ba
sed. Mo
reove
r
there a
r
e
so
me
sha
d
o
w
s cover la
ne
markers in
Figu
re
11(c) a
nd
(d). Pie
c
e
w
ise lan
e
ma
rker fitting
cou
l
d
cou
n
tera
ct these di
sturbs.
And
one of the lane ma
rke
r
s is a d
o
tted
line in Figure 11(b
) an
d (e),
detectio
n
re
sults are al
so
corre
c
t.
(a) S
c
ene 1
(b) S
c
ene 2
(c) Sce
ne 3
(d) S
c
ene 4
(e) S
c
ene 5
Figure 11. Re
sults of La
ne
Markers Dete
ction
4. Conclusio
n
In this pape
r, a lane marker dete
c
ti
on al
go
rithm
for sophi
sti
c
ated envi
r
o
n
ment is
pre
s
ente
d
. This stu
d
y is con
c
e
r
ned
on sel
e
cti
n
g
candi
date l
ane re
gion
by OTSU o
b
ject
segm
entation
.
This is
help
f
ul for d
e
tecti
on a
nd
i
denti
f
ication
of lan
e
ma
rkers in
an a
pprop
riate
regio
n
. Experiment
s sho
w
n that lan
e
marker
s can be dete
c
ted corre
c
tly. For enha
nced
perfo
rman
ce
of detection,
piece
w
ise linear tr
a
n
sfo
r
mation sh
oul
d
be done i
n
prep
ro
ce
ssing
esp
e
ci
ally in a dim light en
vironme
n
t. And pie
c
e
w
is
e lane ma
rker fi
tting is rob
u
st
in the pre
s
en
ce
of shado
w and lack of lane paintin
g. There a
r
e little false lane detection
result
s be
cau
s
e
feature
-
ba
se
d method i
s
usu
a
lly affect
ed by inte
n
s
ity of image. Several
cha
r
a
c
teristics of
ro
ad
need to be
co
nsid
ere
d
for i
m
provin
g the perfo
rman
ce
of detecting.
Ackn
o
w
l
e
dg
ements
This
wo
rk was fin
a
n
c
ially su
ppo
rted
b
y
the Natio
n
a
l Natural
Scien
c
e
Fou
n
dation of
Chin
a (61
1
6
2005 a
nd 61
1630
02) a
n
d
the Indepen
dent
Scie
ntific Re
se
arch
Fund of Beif
ang
University of Nation
alities (2011Z
QY02
2
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3491 – 35
00
3500
Referen
ces
[1]
Ron
ghu
i Z
h
,
Hai
w
e
i
W
,
Xi
Z
h
, Lei
W
,
T
onghai J.
La
ne
detectio
n
alg
o
rithm
at n
i
ght
base
d
-o
n
distrib
u
tion
fea
t
ure of
bo
un
da
r
y
dots f
o
r ve
hi
cle
active
safet
y
.
Infor
m
ation
T
e
chno
logy
Jo
urna
l
. 2
012
;
11(5): 64
2-6
4
6
.
[2]
Borkar A,
Hayes M, Smith M
T
.
A nove
l
l
a
n
e
d
e
tectio
n s
ystem
w
i
th
effici
ent gr
ou
nd trut
h g
e
n
e
ratio
n
.
Intelli
gent T
r
an
sportatio
n
Systems
,
IEEE Transactions on
. 2
012; 13(
1): 365
-374.
[3]
Jung C
R
, Kel
ber CR.
A ro
bust lin
ear-
par
abo
lic
mod
e
l
for lane fo
llo
w
i
ng.
IEEE 17th Brazilian
S
y
mp
osi
u
m on
Computer Gra
phics a
nd Imag
e Processi
ng. Curitib
a
. 200
4: 72-7
9
.
[4]
Che
n
CY, C
h
e
n
CH,
Dai
Z
X
.
An ev
ol
ution
a
r
y
c
o
mputati
o
n
appr
oac
h for
l
ane
detecti
on
and tr
ackin
g
.
Advanc
ed Sci
e
nce Letters
. 20
12; 9(1): 34
2-3
47.
[5]
Aly
M.
R
e
a
l
ti
me
detecti
on
of lan
e
marker
s in ur
ban stre
ets
. IEEE Intellige
n
t Veh
i
cl
es
S
y
mp
osi
u
m.
Eind
hove
n
. 20
08: 7-12.
[6]
Choi
HC, P
a
rk
JM, Cho
i
W
S
, Oh SY. Visi
on
-base
d
fusi
on
of rob
u
st la
ne t
r
ackin
g
a
nd for
w
a
r
d v
e
h
i
c
l
e
detectio
n
in a r
eal dr
ivin
g env
ironme
n
t.
Internatio
nal J
ourn
a
l of Auto
motiv
e
T
e
chno
lo
gy
. 201
2; 13(4
)
:
653-
669.
[7]
Khalifa OO, Assidiq
AA M, Hashim AHA.
Vision-
bas
ed
lan
e
detecti
o
n
for auto
n
o
m
ous artifici
a
l
intell
ig
ent veh
i
cles
. IEEE Internatio
nal C
onfe
r
ence o
n
Sema
ntic Comp
uting
.
Berkele
y
. 20
0
9
: 636-6
41.
[8]
Javad
i
MS, Hann
an MA, Samad SA, Hussain A.
A robu
st vision-bas
e
d
lan
e
bou
nd
a
r
ies detecti
on
appr
oach for i
n
tellig
ent ve
hicl
es.
Informati
o
n
T
e
chnol
ogy Jo
urna
l
. 201
2; 11
(9): 1184-
11
92
.
[9]
Sharma S, Shah DJ. A Much Advance
d
and
Effi
cient Lane
Detectio
n Algo
rithm for Intelli
gent Hi
gh
w
a
y
Safet
y
.
Co
mpu
t
er Science & Informatio
n
T
e
chno
logy.
2
013;
9(1): 51-59.
[10]
Otsu N. A
threshol
d selecti
o
n
method from g
r
a
y
-
l
eve
l
histo
g
r
ams. Automati
ca. 1975; 9(
1): 62-6
6
.
Evaluation Warning : The document was created with Spire.PDF for Python.