TELKOM
NIKA
, Vol.13, No
.1, March 2
0
1
5
, pp. 93~1
0
2
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v13i1.1275
93
Re
cei
v
ed O
c
t
ober 1
3
, 201
4; Revi
se
d Decem
b
e
r
21, 2014; Accept
ed Ja
nua
ry 8,
2015
Hierarchical-map Updating Approach for Simultaneous
Localization and Mapping of Mobile Robots
Yingmin Yi*,
Zhimin Wan
g
F
a
cult
y
of Automation a
nd Inf
o
rmatio
n
Engi
n
eeri
ng, Xi’a
n U
n
iversit
y
of T
e
chno
log
y
,
Xi
’a
n71
00
48, Shaa
n
x
i, Chi
na
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: yi
ym@
x
a
u
t.edu.cn
A
b
st
r
a
ct
For the treme
n
dous
ly incre
a
si
ng of system s
t
ate in
w
ild fiel
d, the comput
ation
a
l co
mp
le
xities of
m
o
bile robot system
should be tak
en into account. This paper propos
es a hierarchic
al-
m
ap updating
appr
oach for
simulta
neo
us l
o
cali
z
a
tio
n
an
d ma
ppi
ng of
robots. T
he basic ide
a
of hierarch
ica
l
-ma
p is
defin
ing tw
o ki
nds of
ma
ps d
u
rin
g
the rec
u
r
s
ive u
pdati
ng process,
n
a
m
e
l
y
loca
l map
(
upp
er map) a
n
d
global
m
a
p (l
ow
er m
a
p). The system states
w
ill
be
updat
ed
by the pr
eset maps
. The hi
erarchical-
m
ap
upd
atin
g proc
e
ss is just for th
e up
per
map
a
nd the
lo
w
e
r
map is
up
date
d
after a certa
i
n
runn
ing
ter
m
. In
the ca
lcul
atio
n, the st
ate
d
a
ta of
the up
per map is
far less
t
han
t
hat
of the
low
e
r
ma
p. It is
vali
date
d
by th
e
exper
iments th
at, the appro
a
c
h is more o
p
ti
ma
l than
other
s in co
mp
utati
ona
l co
mpl
e
xiti
es w
h
ile e
n
suri
ng
the consiste
nc
y estimat
e
.
Ke
y
w
ords
: h
i
erarchic
al-
m
ap
up
datin
g, co
mp
utatio
nal
c
o
mpl
e
xity, si
multan
eo
us l
o
ca
li
z
a
tio
n
and
ma
p
buil
d
i
ng, mob
i
l
e
robots
1. Introduc
tion
The m
o
ving a
nd ma
p b
u
ildi
ng in
wild
env
ironm
ent i
s
of
gre
a
t impo
rt
ance in th
e
re
sea
r
ch
of robot filed.
Due to the l
a
rge
are
a
, the syst
em
stat
es in
crea
se e
x
ponentially. The re
se
arch
es
for ro
bot SL
AM follow th
e variation
s
of the
appli
c
ation environ
ments, fro
m
2-dim
ention
a
l
environ
ment
[1] to 3-dime
ntional envi
r
onment
[2], from
stru
cture
d
enviro
n
me
nt [3] to none-
stru
ctured e
n
v
ironme
n
t [4], from ind
oor [5] to
outdo
or [6] an
d ot
her
natural n
one-structu
r
e
d
environ
ment
[7]. For the
SLAM of ve
hi
cle
rob
o
ts, it
is n
e
cessa
r
y
to upd
ate the
po
se
and
m
ap
after every ne
w ob
servatio
n [8].
With the in
creasi
ng of the
feature
s
on t
he m
ap, the
comp
utationa
l compl
e
xity increa
se
s
as well. For the compu
t
ational pro
b
l
em, res
earchers put forward seve
ra
l map-divi
sio
n
approa
che
s
[
9
]. The e
s
sen
c
e of th
e ma
p-divisi
on a
p
p
roa
c
h
e
s is t
o
re
du
ce the
times of u
pda
ting
the glob
al m
ap while e
n
suring th
e opti
m
al estim
a
te. The ma
p-di
vision a
pproa
che
s
h
a
ve two
categ
o
rie
s
:
o
ne i
s
o
p
e
r
ati
ng o
n
the
lo
cal
pa
rts
of
the ma
p, re
maining
the
global
refe
re
nce
coo
r
din
a
tes. The Comp
re
ssed
Extend
ed
Kalma
n
Filtering
(CEKF) p
r
opo
se
d in p
ape
r [
10],
addresse
s th
e local a
r
ea
data combi
n
g
app
roa
c
h, lo
weri
ng th
e co
mputational
complexity of the
algorith
m
; the othe
r is the
sub
-
ma
p te
chnolo
g
y und
e
r
the lo
cal
co
ordin
a
te fra
m
es. Th
e eve
r
y-
increa
sing p
r
oblem of matrix
es is di
scussed in [11].
The EKF a
p
p
r
oa
ch fo
r SL
AM is b
a
sed
on the
minim
u
m me
an
sq
uare
erro
r, co
mpleting
the optimal
re
cursio
n for
ro
bot po
se
s in t
i
me dom
ain.
But for ea
ch
updatin
g, the
whol
e matrix
is
pro
c
e
s
sed, i
n
crea
sing
th
e comp
utatio
nal
com
p
le
xity, w
h
ic
h gre
a
t
ly r
e
s
t
r
i
c
t
s
th
e ap
p
r
oa
c
h
applying
in th
e la
rge
-
scal
e
environ
ment.
For t
he
ever-i
ncrea
s
ing
p
r
o
b
lem
of the
EKF-SLAM
wit
h
the in
cre
m
en
t of the stat
es, thi
s
pa
p
e
r p
r
o
pos
es
a hie
r
archi
c
a
l
-map
upd
ating a
pproa
ch
for
SLAM (HMU-SLAM). Du
rin
g
the re
cursio
n process of t
he robot, t
he
state
spa
c
e i
s
divided to t
w
o
layers,
nam
el
y the lo
we
r-map
(glo
bal)
state
sp
a
c
e
and upp
er-m
ap (lo
c
al)
sta
t
e
sp
ace.
In the
SLAM alg
o
rit
h
m, at eve
r
y
recursio
n
ste
p
, the
upd
ating
step
is ju
st fo
r updatin
g
the upp
er
st
ate
spa
c
e.
When
the robot n
a
v
igates out th
e uppe
r-m
ap
area, the
n
the lowe
r-map i
s
upd
ated. T
h
is
algorith
m
ca
n lowe
r the computation d
i
mensi
o
n
s
effectively so a
s
to lowe
r the comp
utatio
nal
compl
e
xity, makin
g
it possi
ble of its implementation in
large
-
scal
e e
n
vironm
ent.
The next section p
r
e
s
e
n
ts the p
r
o
c
e
ss
and o
b
se
rvation
model
s u
s
e
d
in ou
r
experim
ents,
whi
c
h
sets t
he
context fo
r the
s
e
re
sult
s in
term
s
of a 2
-
D vehi
cl
e with
a
ra
n
ge
beari
ng
sensor. The l
andmark model
is di
scus
sed in the
section.
Section III describes
hiera
r
chi
c
al-map upd
atin
g approa
ch
for simulta
n
e
ous lo
cali
zati
on and ma
p
p
ing of rob
o
t
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 1, March 2
015 : 93 – 10
2
94
Section IV presents the
results of th
e simula
tio
n
experime
n
ts in terms of
computatio
n
a
l
compl
e
xities
and e
s
timate con
s
i
s
ten
c
y. The final se
ct
ion provid
e di
scussio
n
and
con
c
lu
sio
n
s.
2. Sy
stem model
2.1. SLAM s
y
stem model
The de
scribe
d SLAM system is com
p
o
s
ed by
rob
o
t’s po
sition an
d dire
ction a
nd the
observed coo
r
dinate
s
on static
enviro
n
m
ent landma
r
k. The
united
state vector u
nder
k
state is
s
h
ow
n
as
1
1
x
x
n
vk
vk
rk
vk
k
N
N
x
y
x
y
x
y
(1
)
In the form
ula,
vk
vk
r
k
x
,y
,
r
e
pr
es
en
t r
e
sp
ec
tive
ly th
e
r
o
bo
t’s
co
o
r
d
i
na
te
s
in
tw
o-
dimen
s
ion
a
l spa
c
e a
nd direction a
ngle.
The map i
s
static, param
eter
11
n,
,
,
,
NN
x
yx
y
T
has n
o
time index. The rob
o
t’s mo
vement mode
l is rolling mo
tion con
s
trai
n
t
s (i.e., assu
ming ze
ro wh
eel
s
lip) [12].
11
11
1
1
co
s
(
)
x(
x
,
u
)
s
i
n
(
)
sin(
)
k
vk
k
r
k
k
kv
v
k
k
v
k
k
r
k
k
vT
rk
k
B
xv
T
G
fy
v
T
G
G
(2)
The time
int
e
rval from
1
k
to
k
mom
ent i
s
T
, spe
ed
k
v
and
driving
an
gl
e
k
G
are
con
s
tant, whi
c
h con
s
titutes controll
ed qu
antity
u,
kk
k
vG
T
, robot’
s
wh
eel ba
se
is
B
.
2.2. SLAM observ
a
tion model
The ob
se
rved
model is
wh
e
r
e
ik
z
is the ob
servation ve
ctor at time k a
nd
i
h
is the mo
del
of the ob
se
rvation of the
i
th land
mark.
The ve
ctor
ik
z
is a ob
se
rvatio
n of the la
nd
mark lo
cation
i
p
relative to the robot’s
location
vk
x
. This type of observ
ation will be
referred to as a vehicle-
landma
r
k ob
servation o
r
a VLM observa
tion.
22
()
(
)
z(
x
)
a
r
ct
an
iv
k
iv
k
iv
k
i
v
k
ik
i
k
yy
rk
xx
xx
y
y
h
(3)
The mod
e
l is not assume
d to be perfe
ct and unm
o
delled
sen
s
o
r
characte
ri
stics a
nd
noise
corrupti
on a
r
e l
u
mpe
d
into a
ob
se
rvation e
r
ror
vector
ik
. The
o
b
se
rvation
error vecto
r
i
s
again ta
ken t
o
be a tempo
r
ally uncorrel
a
ted and
zero
mean ra
ndo
m sequ
en
ce.
2.3. Landma
r
k model
Landm
arks a
r
e fixed and
con
s
pi
cu
ou
s featur
e
s
withi
n
the enviro
n
m
ent. Landm
arks
can
have many
p
h
ysical form
s; corner
s,
pla
nes, rou
gh surfaces, pole
s
,
natu
r
al
o
r
artificial
te
rrai
n
feature
s
can
all be
con
s
i
dere
d
lan
d
m
a
rks if
they
are
rep
eatedl
y and reli
abl
y obse
r
ved b
y
a
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Hierarchi
c
al
-m
ap Updating Approach for Si
m
u
ltaneous Localization and .... (Yingm
in Yi)
95
sen
s
o
r
. Exact
l
y what
co
nst
i
tutes a
lan
d
m
ark i
s
dr
ive
n
by th
e p
h
ysics
of the
ob
servin
g
sen
s
or
-
landma
r
ks a
r
e con
s
pi
cu
o
u
s through t
he eyes of
the ob
servin
g
sen
s
or. Thi
s
se
nsor-cen
tric
definition of
a
landm
ark m
e
ans that
it i
s
not al
ways
po
ssi
ble to
re
ad
ily asso
ciate
a lan
d
mark
with
visually perce
ived feature
s
.
Mathemati
c
al
ly, landma
r
ks are rep
r
e
s
e
n
ted a
s
a ve
ctor of p
a
ra
meters that
d
e
fine the
locatio
n
and
other prope
rti
e
s of t
he landmark. This t
hesi
s
gen
erally
employs the simple
st of all
landma
r
k mo
dels: a
lan
d
m
ark is a
st
ationary
poi
nt like
entity in two
dime
nsio
ns. A p
o
i
nt
landma
r
k is d
e
fined by two
param
eters
spe
c
ifying it
s
positio
n in ca
rtesia
n spa
c
e
with re
spe
c
t to
some
glob
al coo
r
din
a
te frame. A point
landma
r
k i
s
visible from
a
ll viewing a
n
g
les. In ge
ne
ral,
more
com
p
le
x landmarks
can b
e
in
corporate
d
into the map
s
an
d
model
s empl
oyed thro
ugh
out
this
thes
is
.
The
i
th point l
andma
r
k in t
he environm
e
n
t will be
den
oted a
s
i
p
and will
be define
d
as
follows
T
i
i
i
x
y
p
(4)
The rel
a
tion
ship between t
he point lan
d
m
ark
state at times k
+ 1 a
nd k is trivial.
The land
mark is stationa
ry and so
(1
)
(
)
ii
i
kk
pp
p
(5)
Importantly, and in cont
ra
st to the vehicle m
odel, there is no additi
ve unce
r
taint
y
term in
the landma
r
k model.
The vehicl
e motion mod
e
l
, the obse
r
vation model,
and the me
asu
r
ed valu
e
s
of the
control pa
ra
meters
u,
kk
k
vG
T
, are not exa
c
t, but are su
bject to n
o
ise, whi
c
h le
a
d
to
uncertainty in
the state est
i
mate. For thi
s
re
ason,
we
requi
re a p
r
obabili
stic filter to re
cu
rsiv
ely
estimate a di
stributio
n ove
r
the
state giv
en noi
sy information.
3. Hierarchic
al-map upda
ting algorith
m
for SLAM
3.1. Basic id
eas of hier
ar
chical-map
Whe
n
the
ro
b
o
t co
ndu
cts
SLAM in the
wild
enviro
n
m
ent, the m
a
p data
an
d th
e sy
stem
state
spa
c
e
will increa
se
ex
pone
ntially. Whe
n
the
ro
b
o
t is wo
rking
i
n
the
wild,
wh
at co
ncerns a
r
e
the intere
sted
area
s a
r
ou
n
d
and the
ge
neral
obje
c
ts.
Hen
c
e, for th
e map buil
d
in
g of the rob
o
t, it
is available
to divide the intereste
d
area
s
an
d the gene
ral
obje
c
ts to hiera
r
chical maps.
Acco
rdi
ng to
the re
qui
rem
ents, the
ma
p can
be divi
ded i
n
to
N la
yers. T
he to
p
layer i
s
den
oted
by 1. And
th
e bottom
lay
e
r i
s
den
oted
by
N. Th
e
schem
atic gra
ph of
the
hie
r
archi
c
al
ma
p
s
i
s
s
h
ow
n
in
F
i
gu
r
e
1
.
Figure 1. Sch
e
matic g
r
ap
h of the hiera
r
chical ma
ps
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 1, March 2
015 : 93 – 10
2
96
3.2. Upda
ting for the hier
archical map
s
For the p
r
obl
em of com
p
u
t
ational co
mp
lexity
increa
se re
sulted by
the increme
n
t of the
state sp
ace, the system a
r
ea is divided i
n
to tw
o part
s
durin
g the EKF recursion p
r
ocess, nam
e
l
y
the
up
per-m
ap state spa
c
e and
the
l
o
we
r-m
ap st
at
e spa
c
e.
During
ea
ch
st
ep of th
e SL
AM
recursio
n ste
p
, the updati
ng step j
u
st
update
s
the uppe
r-m
ap state
spa
c
e. Whe
n
the
ro
bot
navigate
s
out
of the up
per-map a
r
ea, th
en the lo
we
r-
map i
s
up
dat
ed. As it i
s
sh
own i
n
Fig
u
re
2,
the recta
ngul
ar Area A is t
he initial upp
er-m
ap stat
e
spa
c
e; Are
a
B is the initial lower-ma
p st
ate
spa
c
e. When
the sen
s
o
r
of the m
obile ro
bot observe
s the feature
s
o
f
the lower-m
ap state space
while navigating in the upper-map
state space,
the lower m
ap
will be updated. Besides, the
uppe
r-m
ap st
ate spa
c
e a
n
d
the lowe
r-m
ap state spa
c
e are re-divid
ed, as sho
w
n
in Figure 2.
The HM
U-S
L
AM algorithm
is described
as follo
ws:
Initialisation
and p
r
edi
ctio
n: the uppe
r-map
a
r
ea i
s
indepe
nde
nt from the lo
wer-ma
p
area. Th
e sta
t
e vectors are
divided to two parts
.
Figure 2. Upd
a
ting for the state spa
c
e
s
To
p
Lo
w
To
p
To
p
L
o
w
Lo
w
n
n
X
,X
R
,
X
R
X
X
(6)
whe
r
e, n presents the num
ber of
the virtual feature
s
. The cova
ria
n
c
e matrixe
s
a
r
e divided a
s
follows
:
TT
TL
LT
LL
n
PP
E
(
X
X
)(X
X)
R
PP
P
(7)
The auxiliary matrix:
Top
L
o
w
To
p
T
o
p
L
o
w
,,
*
LL
,,
,
nn
n
n
n
RR
Q
R
The initial time
1
k
:
*
11
1
L
L
1
(
)
,
(
)0
,
(
)0
,
(
)0
kI
k
k
Q
k
(8)
Different fro
m
the EKF-SLAM algo
rithm,
the HM
U-SLAM
alg
o
rithm
con
d
u
cts th
e
predi
ction
in t
he u
ppe
r-m
a
p
area. At thi
s
time,
only
Top
T
T
,
X
P
are i
n
volved i
n
the
pre
d
icti
on. Th
e
auxiliary matrix has a recursiv
e equation as the followi
ng.
TT
()
(
1
)
kJ
k
(9)
()
(
1
)
kk
(10)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Hierarchi
c
al
-m
ap Updating Approach for Si
m
u
ltaneous Localization and .... (Yingm
in Yi)
97
()
(
1
)
kk
(11)
TT
To
p
T
o
p
To
p
(/
)
|
(
)
J
fX
X
k
(12)
**
LL
L
L
LL
()
(
1
)
(
)
Qk
Qk
Qk
(13)
Upd
a
ting: the
upd
ating i
s
divided to
up
per-map
up
d
a
ting a
nd l
o
wer-m
ap
upd
ating. Th
e
uppe
r-m
ap
u
pdating
is ju
st for the
lo
cal
area
s a
nd th
e lo
we
r-ma
p
updatin
g i
s
condu
cted
via
the
recursion of the auxiliary
matrix.
()
(
(
)(
1
)
)
kI
k
k
(14)
T1
To
p
T
o
p
()
()
()
()
kH
k
S
k
H
k
(15)
T
(
)
(1
)
(
1
)
(
)
(1
)
kk
k
k
k
(16)
To
p
()
()
(
(
)
,
)
Z
ky
k
h
X
k
k
(17)
T
To
p
T
T
T
o
p
()
()
(
)
(
)
Sk
H
k
P
k
H
k
R
(18)
To
p
T
o
p
To
p
()
(
/
)
|
()
Hk
h
X
X
k
(19)
TT
()
()
k
Pk
k
(20)
TT
1
To
p
(
)
(1
)
(
2
)
(1
)
(
1
)
(1
)
kk
k
H
k
S
k
Z
k
(21)
Upd
a
ting of
Low
T
L
L
L
,,
XP
P
: Once the
are
a
of the
lower state
sp
ace i
s
o
b
served
(a
t time K2
here
)
, the lower-m
ap up
dat
ing is cond
uct
ed via the recursi
on form
ul
a.
TL
2
2
TL
1
()
()
(
)
Pk
k
P
K
(22)
*
L
L
2
L
L
1
TL
1
2
TL
1
L
L
2
()
(
)
(
)
()
(
)
()
Pk
Pk
Pk
k
P
k
Q
k
(23)
Low
2
L
o
w
1
2
()
(
)
(
)
Xk
X
k
k
(24)
3.3. Steps fo
r HMU-SLAM
HMU-SLAM
pro
c
e
ss i
s
pe
rforme
d as fol
l
ows.
Step1 Initialisation
:
Initiall
y define
and
assign
value
s
for th
e p
r
edi
ction l
o
cation,
syste
m
covari
an
ce matrixes,
hi
e
r
archi
c
al are
a
s,
t
he
auxi
liary matrix,
the sy
stem
pro
c
e
s
s no
ise
covari
an
ce
matrix, the system observation noise
covaria
n
ce matrix, control variable
s
,
the
maximum ob
servatio
n dist
ance and the
time interval.
Step2 Location pre
d
ictio
n
for the robo
t: predict the curre
n
t location and co
varian
ce
matrix of the robot ac
cording to the pos
i
tion
es
timat
e
, c
o
varianc
e
matr
ix, auxili
ary matrix
and
moving mod
e
l
of the former moment in t
he upp
er-ma
p
area.
Step3 Pra
c
tical ob
se
rvatio
n: Obtain th
e
pra
c
ti
cal ob
servatio
ns by
the ob
se
rvations
of
the environ
m
ent feature
s
u
s
ing the
sen
s
ors of the rob
o
t.
Step4 Predi
ct
ion ob
servati
on: Cal
c
ulate
the
predi
ctio
n observation
s of the environment
feature
s
usi
n
g the
mea
s
u
r
ement m
odel
acco
rdin
g
to
the
coo
r
di
na
tes of
the l
o
cation p
r
edi
cti
on
and ob
se
rvation enviro
n
me
nt f
eature
s
on
the lower m
a
p.
Step5 Data
asso
ciation (relating
)
: con
duct
the dat
a asso
ciation
for the observation
feature
s
of the uppe
r map.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 1, March 2
015 : 93 – 10
2
98
|1
1
ˆˆ
()
(
,
)
kk
k
k
In
n
z
k
h
x
m
(25)
1
kk
k
k
DI
n
n
S
I
n
n
T
(26)
Step6 State updatin
g: the stat
e updati
ng is divide
d
to upper-ma
p
updatin
g a
nd lower-
map up
dating
.
The uppe
r-map up
dating
is just for
the
associ
ated f
eature
s
. The
data is g
ene
rally
far less than t
hat of the lower map.
The lower-m
ap updatin
g is gene
rally condu
cted a
fte
r a certai
n distance of mov
e
ment of
the robot, to modify the data of the global map.
Step7 State increme
n
ting: Add
the ne
w observed fe
ature
s
to the m
ap. The HMU-SLAM
algorith
m
i
s
recu
rsive
by
steps i
n
seq
u
e
n
ce
of
p
r
edi
ct
ion, ob
se
rvation,
data
a
s
so
ciation,
upp
er-
map upd
ating
or lowe
r-map
updating a
n
d
state increm
enting.
1
(,
)
new
kn
k
v
k
h
xz
x
(27)
T
k
k
new
k
x
x
x
(28)
4. Results a
nd Analy
s
is
The initial state and co
varian
ce
ma
trix of the
experim
ents are
0/
0
ˆ
(3
,
1
)
xz
e
r
o
s
,
0/
0
(3
)
Pz
e
r
o
s
, res
p
ec
tively. In the initia
lisation pro
c
e
ss, the noise co
varian
ce an
d the observati
on
cov
a
ri
an
ce a
r
e
22
[0
.
3
,
(
3
/
180)
]
Qd
i
a
g
,
22
[0.
1
,
(
/
180)
]
Rd
i
a
g
,respe
ctively.The expe
riment
s also
perfo
rmed th
e cla
ssi
c EKF
-SLAM, Fast
SLAM, UKFS
LAM , for co
mpari
ng with
and an
alyzin
g the
prop
osed alg
o
rithm. Figure 3 sho
w
s cl
assical
EKF-SLAM algorit
hm. Figure 4
shows cl
assica
l
Fast-S
LAM a
l
gorithm. Fig
u
re 5
sho
w
s
the hiera
r
chi
c
al-map SLA
M
of the pro
posed alg
o
rit
h
m.
Figure 6 sh
o
w
s the two-la
yer map SLA
M
. The m
ap is built by usin
g the prop
ose
d
algorith
m
.
Figure 3. The
map of the robot EKF-SL
A
M
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Hierarchi
c
al
-m
ap Updating Approach for Si
m
u
ltaneous Localization and .... (Yingm
in Yi)
99
Figure 4. The
map of the robot Fa
st-SL
A
M
Figure 5. State spa
c
e u
pda
ting gram
The data p
r
oce
s
sing q
u
antity of a single-step
cal
c
ulatio
n is a
n
importa
nt index for
deci
d
ing th
e
efficien
cy of
an alg
o
rithm
[10]. For
the
large
-
scale m
ap, to verify the comp
utation
compl
e
xity of the algorith
m
, the experi
m
ents al
so p
e
rf
orm
ed t
he
cla
ssi
c E
K
F
-
S
LA
M,
Fast
S
L
A
M
and
UKFSL
A
M re
spe
c
ti
vely in the
same
envi
r
o
n
ment, for compa
r
ing
wit
h
the
pro
p
o
s
ed
algorith
m
in this pap
er. Fi
gure 7
sho
w
s the ti
me co
nsumi
ng average
s of the above algo
rith
ms
unde
r 50 si
ng
le-ste
p cal
c
ul
ations. i pre
s
ents ste
p
len
g
th. T presen
ts time (unit: s).
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 1, March 2
015 : 93 – 10
2
100
Figure 6. Hierarchical map
p
ing of rob
o
t SLAM
Figure 7. Co
mpari
s
io
n for singl
e-st
ep consusi
ng of the algo
rithm
s
As it can be
seen from
Figure 7, under t
he sam
e
con
d
itions,
the single
-
step time
con
s
umi
ng
curve of the
FastSLAM
algorithm
goe
s top; UKFSL
A
M is infe
rio
r
; EKFSLAM is
bene
ath UKF
S
LAM; HM
USLAM is
at the bottom.
T
h
is in
dicates the p
r
op
ose
d
algo
rithm
has
minimum
sin
g
le-step time
con
s
umin
g,
optimal
tha
n
the other
algorith
m
s in
comp
utation
a
l
c
o
mplexity.
For the prop
ose
d
algo
rith
m, the estimation issue of
con
s
isten
c
y sho
u
ld be co
nsid
ere
d
.
For line
a
r
Gau
ssi
an filter, the filter perfo
rm
an
ce can b
e
chara
c
te
rized
through
NEES
(no
r
mali
zed e
s
timation e
rro
r squ
a
red) [1
1].
1
xx
P
x
x
||
|
ˆ
ˆ
()
()
kk
k
k
k
k
k
k
k
T
(29)
Und
e
r the hy
pothe
sis that
the filter is co
n
s
i
s
tent an
d approximately linear-Ga
u
ssian,
NEES obeys
2
distrib
u
tion.
Con
s
i
s
ten
c
y of the algo
rithm is eval
ua
ted by perfo
rming N time
s
Monte Carlo runs
.
The
algorith
m performance
indicato
rs are evaluated
by
the average NEES.
Whe
n
N
,
k
approache
s to the state vector
dimen
s
ion.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Hierarchi
c
al
-m
ap Updating Approach for Si
m
u
ltaneous Localization and .... (Yingm
in Yi)
101
1
1
N
ki
k
N
i
(30)
Given the hy
pothe
sis of a
con
s
i
s
tent linear-G
au
ssi
a
n
filter,
k
N
has
a
2
dens
i
ty with
N
dim (
x
k
). Thu
s
, for the 3
-
dime
nsi
onal
robot p
o
se
, with N=50
, the 95% prob
ability
c
o
nc
en
tr
a
t
io
n r
e
g
i
on
fo
r
k
is
boun
ded by t
he interval [2.
36, 3.72]. If
k
rise
s si
gnifica
ntly higher
than the
upp
er b
oun
d, the filter is
opt
imistic, if it tend
s bel
ow
t
he lo
we
r bo
u
nd, the filter
is
c
o
ns
er
va
tive
.
To verify the
con
s
i
s
ten
c
y estimate, a
s
to EKF-SLAM, FastSLA
M
, UKFSLA
M and the
prop
osed alg
o
rithm, we co
ndu
cted 50
MonteCarl
o
runs on the ro
bot for each. Figure 8 sho
w
s
the NEES consi
s
tency esti
mate cu
rves.
As it can be seen from
Figure 8, the
NEES curves o
f
the algorithm
s are alm
o
st within [2.36
,
3.72]. Hence, they are all
con
s
i
s
ten
c
y estimate
s. From
the ab
ove, th
e p
r
opo
se
d
algorith
m
i
s
optimal in
co
mputational
complexity wh
ile en
su
ring
the
c
o
ns
is
tenc
y es
timate.
5. Conclusio
n
For the
probl
em of the
ever-i
ncrea
s
in
g
com
putation
a
l co
mplexity re
sulted
by the eve
r
-
increa
sing
st
ate spa
c
e i
n
the large
-
scal
e environm
e
n
t, this p
ape
r pro
p
o
s
e
s
a
hiera
r
chi
c
al-map
updatin
g alg
o
rithm for si
multaneo
us l
o
cali
zatio
n
and mappi
ng.
Acco
rding t
he algo
rithm, the
robot
will
buil
d
multiple
ma
p laye
rs at m
ap b
u
ild
in
g.
Duri
ng
state
updatin
g, onl
y a fe
w
state
s
of
the uppe
r ma
p are up
date
d
, thus lowe
ri
ng the com
p
utational co
m
p
lexity of the
algorithm. T
h
e
experim
ents
validated that the pr
opo
se
d algorithm h
a
s the minim
u
m com
putational co
mplex
i
ty
while e
n
surin
g
the co
nsi
s
t
ency e
s
timat
e
. The idea
s
of this pap
er
provide
a me
aningful
clue
for
the map build
ing of the larg
e-scal
e un
kn
own e
n
viron
m
ent for ro
bo
t.
Figure 8. NEES c
ons
is
tenc
y es
timate
curves
of the algorithms
Ackn
o
w
l
e
dg
ments
This
work
was
supp
orte
d
by National
Natural Sci
ence Fou
n
d
a
tion of Chi
na (No.
5127
5405
) a
nd the scie
n
c
e research
prog
ram
s
of
edu
cation d
e
partme
n
t of Shaanxi Prov
ince
(201
3JK
107
8
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 1, March 2
015 : 93 – 10
2
102
Referen
ces
[1]
Yang P. Effici
ent partic
l
e fi
lter al
gorithm f
o
r
ultraso
n
ic se
nsor-b
ased
2D
rang
e-on
l
y
si
multan
eo
u
s
local
i
sati
on a
n
d
mapp
ing
app
licatio
n.
IET Wi
reless Sensor
System
s.
20
12
; 2(4): 394 -40
1
.
[2]
Bosse M, Zlot
R, Flick PZ. Desi
gn
of a
Spr
i
n
g
-Mou
n
ted
3-D
Ra
nge
Se
nsor
w
i
th
App
licati
o
n to M
obi
le
Mapp
ing.
IEEE
T
r
ansactions
on Ro
botics.
2
012; 28(
5): 110
4-11
19.
[3]
Mitch B, Sal
a
h
S. Architectur
e
s for C
oop
er
ativ
e
Air
bor
ne Simulta
neo
us Loca
lisati
on a
nd
Ma
pp
ing.
Journ
a
l of Intell
ige
n
t and R
obo
tic Systems.
20
09;
55(4): 2
67-
297.
[4]
W
e
izhe
n Z
,
Miro JV, Diss
ana
y
a
ke G. Information-Efficient 3-D
Vis
ual S
L
AM for
Unstructure
d
Domains.
IEEE Transactions
on Robotics.
2
008; 24(
5): 107
8-10
87.
[5]
S T
h
run, D Fox, W
Burg
ard.
A pro
bab
ilisti
c
appr
oac
h to
concurr
ent m
app
ing
an
d l
o
calizati
on f
o
r
mobil
e
ro
bots.
Machi
ne Lear
n
i
ng.
19
98
;
(31)
: 29-53.
also
a
ppe
ared
in
Aut
ono
mous Rob
o
ts
5. 19
98;
253-
271 (j
oi
nt issue).
[6]
Jose G, E
dua
rdo
N, Step
ha
n B. A
u
tonom
ous
Navi
gati
o
n a
n
d
Map
b
u
ild
in
g Us
in
g
Laser
Ra
ng
e
Sensors i
n
Outdoor Ap
plic
atio
ns.
Journa
l of Rob
o
tic Systems.
20
00;
17(
1
0
): 565-5
83.
[7]
Jose G, Eduar
do N, Juan N, F
a
vio M. Navig
a
ti
on a
nd Map
p
in
g in Lar
ge u
n
structured En
vironme
n
ts
.
T
he Internati
o
n
a
l Jour
nal of R
obotics R
e
sear
ch.
2004; 2
3
(4)
:
449-47
2.
[8]
Si-Yao F
,
Xi
n-
Kai K, Rui Z
,
Guo-She
ng Y,
Z
eng-Gua
ng
H.
Compress
iv
e sensi
ng a
p
p
r
oach b
a
se
d
ma
pp
ing
an
d l
o
cali
z
a
tio
n
for
mo
bil
e
ro
bot in
an in
do
or w
i
reless se
nsor
n
e
tw
ork
. 2010 I
n
ternati
o
n
a
l
Confer
ence
on
Net
w
o
r
ki
ng, Sensi
ng a
nd Co
ntrol. Chic
ago,
USA,
IEEE. 2010: 122-
12
7.
[9]
Bai-F
an
C, Z
i
-Xi
ng
C, Z
h
i-R
o
ng Z
.
A hybr
id
data ass
o
ci
atio
n ap
pro
a
ch for
mo
bi
le ro
bot
SLAM
. 2010
Internatio
na
l C
onfere
n
ce o
n
Contro
l Autom
a
tion
a
nd S
y
st
ems. Gy
eo
ng
g
i
-do, Kore
a, IEEE. 2010:
190
0-19
03.
[10]
Jose G, Edu
a
rdo N.
I
m
pr
ovin
g Co
mput
ation
a
l a
nd
Memory
Re
qu
ire
m
ents of
Simulta
neo
us
Loca
l
i
z
a
t
i
on
an
d Map B
u
il
din
g
Algorit
h
m
s
. Proceedings of the 2002 IEEE
international Conference on
Robotics 8 Automation
.
Washingto
n
, DC, IEEE. 2002: 27
3
1
-27
36.
[11]
Y Bar-Sh
a
l
o
m,
XR
Li, T
Kiru
bara
j
an.
Estimation
w
i
th a
p
p
l
icatio
ns to
trac
king
an
d
nav
ig
ation
. John
W
ile
y
an
d Son
s
. 2001: 23
4-2
35.
[12]
R Smith, M Self, P Chees
em
an. A stochasti
c map for unce
r
tain spati
a
l rel
a
tions
hips.
In International
Sym
p
osium
of Robotics Research.
198
7: 46
7-47
4.
Evaluation Warning : The document was created with Spire.PDF for Python.