Intern
ati
o
n
a
l Jo
urn
a
l
o
f
R
o
botics
a
nd Au
tom
a
tion
(I
JR
A)
V
o
l.
4, N
o
. 1
,
Mar
c
h
20
15
,
pp
. 73
~81
I
S
SN
: 208
9-4
8
5
6
73
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJRA
Mobile Robot Localization: A
Review of Probabilistic Map-
Based Techniques
Sal
v
ad
or M.
Mal
a
g
o
n
-
S
o
l
d
ara
,
Ma
nuel
T
o
ledano-Ayal
a
, Gen
a
ro
S
o
to-Z
araz
ua,
Roberto
V.
Ca
rrillo
-Serra
n
o
,
Edga
r
A. Ri
v
a
s
-Ara
iza
* Division d
e
Estudios de Posgrado, F
acu
ltad de I
ngenier
ia, Universidad Aut
onoma de Queretaro
,
Queretaro
,
Mex
i
co
Article Info
A
B
STRAC
T
Article histo
r
y:
Received Aug 8, 2014
Rev
i
sed
No
v
12
, 20
14
Accepted Nov 28, 2014
This work presents a com
p
rehensiv
e rev
i
e
w
of current probabilist
i
c
developments u
s
ed to calculate
position b
y
mobile robots in indoor
environm
ents. I
n
this calcul
a
t
i
o
n
, best
known as locali
zat
ion, it
is necessar
y
to develop most of the task
s de
l
e
gat
e
d to
the m
obile robo
t.
It is
then
cruci
a
l
that
the methods used for positio
n calcul
ations b
e
as precise as po
ssible,
and
accur
a
t
e
l
y
r
e
pre
s
ent the
loc
a
ti
o
n
of the
robot
within a
given
environm
ent.
The research
community
h
a
s devoted
a considerable amount of time to
provide solution
s
for the localiz
ation
problem.
Several method
ologies have
been proposed the most commo
n
of which are b
a
sed in the Bay
e
s rule. Other
m
e
thodologies i
n
clude
the K
a
l
m
a
n filte
r and
t
h
e Monte C
a
rlo
locali
zat
ion
filte
r wich
will
be addr
essed in
next
se
ctions.
T
h
e m
a
jor
contrib
u
tion of
this
review res
t
s
in
offering a wide arra
y of tech
niques
that res
e
archers
ca
n
choose.
Ther
efo
r
e, method-s
e
nsor comb
inations and th
eir
main
advan
t
ages
are d
i
s
p
la
ye
d.
Keyword:
Bayesian
in
feren
c
e
Kalm
an
filter
Lo
calizatio
n
M
obi
l
e
r
o
bot
Particle filter
Copyright ©
201
5 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
Edgar A.
Rivas
-
Araiza,
Di
vi
si
o
n
de
E
s
t
udi
o
s
de
P
o
s
g
rad
o
,
Fac
u
l
t
a
d de In
ge
ni
eri
a
,
Uni
v
ersi
dad Autonom
a de
Queretaro,
Cerro
de las ca
m
p
anas S/
N, Queretaro, Mexi
co.
Em
a
il: eriv
as@u
aq
.m
x
1.
INTRODUCTION
Mo
b
ile ro
bo
t h
a
v
e
seen
an
in
cr
ease
pre
s
ence in the industrial fiel
d per
f
o
r
m
i
ng t
a
sks as vari
e
d
a
s
clean
in
g fl
o
o
rs, lo
ad
ing
an
d un
lo
ad
ing
in indu
strial p
l
an
ts, t
r
an
spo
r
ting
sam
p
les fro
m
o
n
e lab
o
raty to
an
o
t
h
e
r,
am
ong m
a
ny
o
t
hers.
Al
l
t
h
at
wi
t
h
o
u
t
i
g
n
o
ri
ng t
h
e
gr
owi
n
g i
n
t
r
o
duct
i
o
n
of t
h
i
s
t
y
pe o
f
ro
b
o
t
as a d
o
m
est
i
c
worker
b
ecau
s
e o
f
its
flex
ib
il
ity, s
m
all size and low cost
.
Nearly all of
th
ese ap
p
lication
s
requ
ire
kn
owledg
e
o
f
th
e po
sitio
n
o
f
th
e ro
bo
t; th
erefo
r
e it is n
ecessary to
p
e
rfo
rm
a
lo
calizat
io
n
calcu
latio
n [1
]. In
lo
calizatio
n
,
th
e po
sitio
n
of th
e ro
bo
t relat
i
v
e
to
a m
a
p
of an
env
i
ro
n
m
en
t is esti
m
a
ted
an
d
t
h
is calculatio
n
rep
r
esents o
n
e
of t
h
e m
o
st
rel
e
vant
p
r
obl
e
m
s
i
n
m
obi
l
e
ro
b
o
t
i
c
s [2]
.
Furt
herm
ore, these calculations are used i
n
othe
r
m
odul
es of
t
h
e ro
b
o
t
co
nt
r
o
l
soft
ware t
h
at
are i
n
c
h
ar
ge
of
deci
di
ng
h
o
w
t
h
e
ro
b
o
t
sh
oul
d act
i
n
t
h
e
nex
t
m
ovem
e
nt
.
[3
]
estab
lish th
at t
h
e
robo
t m
u
st n
a
v
i
g
a
te safely
with
i
n
its en
v
i
ron
m
en
t as a
key p
r
ereq
u
i
site fo
r a
t
r
ul
y
aut
o
nom
ous
ro
b
o
t
.
R
e
l
i
a
bl
e navi
g
a
t
i
o
n i
n
m
obi
l
e
robot
i
c
s re
qui
res
t
h
e com
put
at
ion
of r
o
bust
m
o
ti
on
ap
pro
x
i
m
a
tio
n
s
.
So
lu
tion
s
based
on
in
ertial
m
easu
r
em
ent
u
n
i
t
s
or gl
obal
p
o
si
t
i
oni
ng
sy
st
em
(GPS) ca
n
pr
o
v
i
d
e
posi
t
i
on a
p
pr
oxi
m
a
t
i
ons a
nd t
h
ei
r co
rres
p
on
di
ng
u
n
cert
a
i
n
t
i
es [4]
.
Ho
we
ver
,
t
h
i
s
sol
u
t
i
on i
s
i
m
p
r
actical in
in
do
or app
licatio
n
s
wh
ere GPS sign
als are n
o
t
reliab
l
e.
Wh
ile ou
tdoo
r lo
calizatio
n
in
op
en
areas has
been largely solv
ed with the adva
nces in satellite-base
d
GPS s
y
stem
s, indoor lo
calization present
s
on
g
o
i
n
g chal
l
e
nge
s d
u
e t
o
t
h
e l
a
rge ra
ng
e of
vari
a
b
l
e
s t
h
at
requi
re di
f
f
e
rent
t
ech
ni
q
u
e
s [5]
.
A
s
i
t
is not
pos
si
bl
e t
o
hav
e
a cal
cul
a
t
i
on
usi
n
g G
PS, t
h
e
use o
f
ot
her t
y
pes o
f
se
ns
ors
i
s
necessary
t
o
col
l
ect
i
n
fo
rm
at
i
o
n
fr
om
t
h
e envi
r
onm
ent
.
Tw
o di
ffe
re
nt
sou
r
c
e
s of i
n
fo
rm
ation m
a
y
be used t
o
m
a
p navi
gat
i
on:
p
r
op
ri
o
cept
i
v
e
(g
yro
s
cop
e
, inclin
o
m
eter) and
ex
tero
cep
tive (co
m
p
a
ss).
So
m
e
au
th
ors
call th
ese sen
s
o
r
s as id
i
o
th
et
ic and
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
089
-48
56
IJR
A
V
o
l
.
4,
No
. 1,
M
a
rc
h 20
1
5
:
7
3
– 81
74
allo
th
etic sen
s
o
r
s, resp
ectiv
ely [6
]. Th
e rob
o
t
g
a
t
h
er
s
d
a
ta th
roug
h
ex
tero
cep
tiv
e sens
ors
whic
h survey the
world and propri
oceptive
se
nsors
which c
o
ntinuously m
onitor the
m
o
tion
of the
robot
in s
p
ace
via c
o
m
p
ass
readi
ngs
,
wh
ee
l
enc
ode
rs,
an
d
ot
he
rs
[
7
]
.
T
h
ese sen
s
o
r
s a
r
e
use
d
t
o
det
e
r
m
i
n
e t
h
e
ori
e
nt
at
i
on a
n
d i
n
cl
i
n
at
i
o
n
o
f
t
h
e rob
o
t; the p
r
o
cess t
o
calcu
l
ate th
e orien
t
atio
n
also
it is called
attitu
de esti
m
a
t
i
o
n
.
At th
e sam
e
t
i
m
e, th
e
tech
n
i
qu
e
o
f
esti
m
a
t
i
n
g
th
e
po
sitio
n
t
h
ro
ugh
th
e in
itial
positio
n
,
co
urse an
d
sp
eed
,
is called
d
e
ad
reck
on
ing
[8]
.
I
n
dea
d
re
cko
n
i
n
g (
h
ea
di
ng se
ns
ors
)
an
d o
dom
et
ry
(w
heel
sens
ors
o
n
l
y
) t
h
e p
o
si
t
i
on
up
dat
e
i
s
b
a
sed o
n
pr
o
p
ri
oce
p
t
i
v
e
sens
ors
.
T
h
e m
ovem
e
nt
o
f
t
h
e
ro
b
o
t
i
s
se
nse
d
wi
t
h
wheel
e
n
co
de
rs a
nd/
or
hea
d
i
n
g se
ns
o
r
s t
h
a
t
later it is integrate
d
to com
pute
po
sition.
Because t
h
e sensor m
easure
m
en
t errors al
so a
r
e inte
grat
ed, t
h
e
position error
is accum
u
lated ove
r ti
m
e
. Thus the
positi
on has to be
updated from
ti
me to ti
me by other
lo
calizatio
n
mech
an
ism
s
; o
t
herwise the ro
bo
t is un
ab
le t
o
m
a
in
tain
a mean
ing
f
u
l
po
sit
i
o
n
estim
ate in
lo
ng
ru
ns. I
n
sh
o
r
t
,
m
obi
l
e
rob
o
t
effect
o
r
s i
n
t
r
o
duce
unce
r
t
a
i
n
t
y
about
t
h
e n
e
xt
st
at
e. Then
ce, i
t
i
s
im
port
a
nt
t
o
un
de
rst
a
n
d
t
h
e
p
r
eci
se nat
u
re
of
t
h
e e
ffect
or
noi
se
t
h
at
a
ffe
ct
s m
obi
l
e
ro
b
o
t
s
. F
r
om
t
h
e r
o
b
o
t
pers
pect
i
v
e, t
h
i
s
erro
r i
n
m
o
tio
n
is
v
i
ewed
as an
erro
r i
n
o
d
o
m
etry, o
r
th
e in
ab
ility o
f
t
h
e robo
t to
esti
mate its o
w
n
po
sitio
n
ove
r t
i
m
e
usi
ng k
n
o
w
l
e
d
g
e o
f
i
t
s
ki
nem
a
ti
cs and
dy
nam
i
cs. In t
h
e
fi
el
d o
f
m
obi
l
e
robot
i
c
s, i
t
i
s
co
m
m
o
n
t
h
at
od
om
et
ry
error
sou
r
ces be di
vi
de
d i
n
t
o
t
w
o
di
ffere
nt
g
r
oup
s. Th
e first so
urce is th
e syste
m
atic
erro
r
th
at is
d
e
term
in
istic.
Syste
m
atic
erro
r sources inclu
d
e
un
e
qual
wheel
di
am
et
ers, m
i
sal
i
gnm
ent
of whee
l
s
, or
ki
nem
a
t
i
c
m
o
del
i
ng e
r
r
o
rs.
There
f
ore, i
t
i
s
po
ssi
bl
e t
o
decrease t
h
e
err
o
r i
f
ki
nem
a
t
i
c
param
e
t
e
r
s
are
calib
rated
.
Th
e secon
d
sou
r
ce is th
e non
syste
m
atic erro
r,
whi
c
h i
s
st
oc
h
a
st
i
c
. Possi
bl
e sou
r
ces
of
t
h
es
e ki
n
d
s
o
f
errors are
en
v
i
ron
m
en
tal
con
d
ition
s
such
as
un
ev
en
g
r
ou
nd
o
r
wheel
slip
p
a
g
e
. Non
s
ystem
a
t
i
c
erro
rs
cannot be directly
com
p
ensated, bu
t the errors a
r
e
just m
odeled as the
stoc
hastic unce
rtainty.
If
n
o
n
s
ystem
a
t
i
c erro
rs are too
l
a
rg
e, th
en
it is
d
i
fficu
lt to
u
s
e
p
u
re
od
o
m
etry for
p
o
s
ition
est
i
m
a
t
i
o
n
[9
].
Th
e tru
e
sou
r
ce o
f
error g
e
n
e
rally lies
in
an
in
co
m
p
lete
m
o
d
e
l o
f
th
e en
v
i
ro
n
m
en
t, wh
ich rep
r
esen
t a
no
nsy
s
t
e
m
a
ti
c err
o
r
.
F
o
r
i
n
st
a
n
ce, t
h
e
ro
b
o
t
doe
s
not
m
ode
l the fact t
h
at the
floo
r m
a
y be sloped, the
wheels
may slip
, o
r
that a hu
m
a
n
m
a
y p
u
s
h
t
h
e
r
obo
t.
A
ll of
thes
e unm
odeled s
o
urces
of
erro
r resu
lt in
in
accu
racy
b
e
tween
th
e ph
ysical
m
o
tio
n o
f
th
e
robo
t,
th
e in
tend
ed
m
o
ti
on of
t
h
e ro
b
o
t
,
an
d
t
h
e pr
o
p
ri
oce
p
t
i
v
e
sens
or
est
i
m
a
t
e
s of m
o
t
i
o
n
[
1
0]
. F
u
r
t
herm
ore, i
n
t
e
r
act
i
on
be
tween
th
e rob
o
t
and
th
e en
v
i
ronmen
t, alon
g with
the
prese
n
ce
of
n
o
i
sy
senso
r
read
i
ngs m
a
ke t
h
e
pr
o
b
l
e
m
m
o
re di
ffi
c
u
l
t
t
o
s
o
l
v
e.
A
not
her t
y
pe
of
pr
o
b
l
e
m
occu
rs
whe
n
t
h
e m
easurem
ent
s
of se
nso
r
s ar
ri
ve
de
l
a
y
e
d t
o
t
h
e localizatio
n
m
o
du
le du
e to
m
u
l
tip
le facto
r
s such
as
t
h
e phy
si
cal
di
st
ri
but
i
o
n o
f
t
h
e senso
r
s, t
h
e c
o
m
m
uni
cat
i
on net
w
or
k, an
d t
h
e t
i
m
e
used t
o
pre-
pr
ocess t
h
e raw
m
easurem
ent
s
t
o
ext
r
act
t
h
e
i
n
f
o
rm
at
i
on t
h
a
t
i
s
sent
t
o
t
h
e
l
o
cal
i
zat
i
on
m
odul
e. A
n
ot
h
e
r
di
ffi
c
u
l
t
sce
n
ar
i
o
appea
r
s
whe
n
t
h
e del
a
y
s
an
d
t
h
e seq
u
ence
of t
h
e a
rri
val
of i
n
f
o
rm
at
i
on t
o
t
h
e l
o
cal
i
zat
i
on m
odul
e
are n
o
t
fix
e
d
,
con
s
titutin
g
th
e ou
t-of-sequ
e
n
ce prob
lem (OOSP)
. In
o
r
d
e
r to
deal with
th
e measu
r
em
en
t
arriv
a
l
d
e
lays, th
e lo
calizatio
n
m
o
d
u
le can
b
a
sically i
m
p
l
e
m
en
t fou
r
d
i
fferen
t so
l
u
tio
ns, as su
ggested
in [1
1
]
.
Briefly, th
e au
t
o
no
m
o
u
s
m
o
b
i
le rob
o
t
starts
fro
m
an
in
itial p
o
s
ition
withou
t prior
k
nowled
g
e
of th
e
en
v
i
ron
m
en
t a
n
d
tries to
gain
in
fo
rm
atio
n
ab
ou
t its su
rrou
nd
ing
s
, th
ro
ug
h
its on
bo
ard sen
s
or m
easu
r
e
m
en
ts.
Th
e ro
bo
t n
e
ed
s to
con
s
id
er all o
f
th
e
m
eas
urem
ents from
the sensors to create a belief of its next state. In
th
is o
r
d
e
r t
o
ach
iev
e
th
is it
is n
ecessary
e
m
p
l
o
y
a p
r
ob
ab
ilistic
m
e
t
h
od
.
Here, the classical Ba
yesian
form
u
l
atio
n
is
ad
op
ted to upd
ate a
h
ypo
th
esis. Hen
ce,
se
nsor m
easure
m
ents are c
o
mb
in
ed
to calcu
l
a
te th
e
l
o
cat
i
on
o
f
sal
i
ent
feat
ures
o
f
t
h
e e
nvi
ro
nm
ent
(m
appi
ng
p
r
o
cess) an
d simu
ltan
e
ou
sly the ro
bo
t estim
at
es its
o
w
n
po
sitio
n
i
n
th
is co
n
tinuou
sly en
r
i
ch
ed
map
(
l
o
calizati
o
n
p
r
o
cess)
. In g
e
n
e
r
a
l, th
e
maj
o
r
ity o
f
wor
k
s in
th
e literatu
re relies o
n
p
r
ob
ab
ilistic fra
m
e
works to
so
lve th
e lo
calizati
o
n
p
r
o
b
l
em
. Th
e id
ea un
d
e
rp
inn
i
ng
su
ch
app
r
o
a
ches is to
recu
rsi
v
ely
m
a
in
tain
a p
r
o
b
a
b
ility
d
i
strib
u
tion
,
cal
led
b
e
lief, ov
er all p
o
s
itio
n
s
(state
space
points) in the
e
nvironm
ent. Probabil
istic localizat
i
o
n algorithm
s
are
varia
n
ts
of the Bayes
filter. The
straig
h
t
forward
ap
p
lication
o
f
Bayes filters to
t
h
e l
o
calizatio
n
p
r
ob
lem
is called
Markov
lo
calizat
io
n
.
Th
e
Marko
v
l
o
calizatio
n
m
o
d
e
l can
rep
r
esen
t an
y
p
r
o
b
a
b
ility d
e
n
s
ity fun
c
tio
n
reg
a
rd
ing
rob
o
t
p
o
s
ition
.
Howev
e
r,
t
h
i
s
ap
pr
oach i
s
ext
r
em
el
y
general
a
nd s
o
m
e
aut
h
ors
de
scrib
e
it as in
effi
cien
t. Con
s
id
erin
g
th
e
fu
nd
amen
tal
d
e
m
a
n
d
s
on
a ro
bo
t lo
calizatio
n
syste
m
, o
n
e
can
argu
e th
at th
is filter is n
o
t
th
e co
rrect so
l
u
tio
n
to
th
e
lo
calizatio
n
pro
b
l
em
b
u
t
sen
s
o
r
fu
si
n
g
is a key ele
m
en
t
to
ro
bu
st lo
calizatio
n. Th
e
Kalm
a
n
filter is p
r
esen
ted
in
th
e n
e
x
t
sectio
n
.
Th
is meth
od
is co
mm
o
n
l
y ap
p
lied
to
satisfacto
r
ily co
m
b
in
e sen
s
o
r
m
easu
r
emen
ts
fo
llowed
b
y
an an
alysis of
o
t
h
e
r al
g
o
rith
m
s
d
e
ri
v
e
d fro
m
th
e Bayes ru
le. In th
is
work,
fram
e
wo
rk
s
wi
th
th
e
sam
e
sensor t
y
pe u
s
ed a
r
e
di
vi
de
d i
n
di
f
f
er
ent
sect
i
o
n
s
, a
nd acc
o
r
di
ngl
y
i
t
i
s
possi
bl
e
com
p
are t
h
e fe
at
ure
s
containe
d i
n
si
m
ilar syste
m
s.
2.
KAL
M
AN FILTER
To c
ontrol a
m
obile robot,
as explained a
b
ove,
fre
quent
l
y it
is necess
a
ry
to com
b
ine inform
ation
fro
m
m
u
ltip
le so
urces.
Howev
e
r, d
i
fferen
t
typ
e
s o
f
sen
s
ors
h
a
v
e
d
i
ff
eren
t reso
lu
tion
s
an
d d
e
grees
o
f
error.
C
onse
q
uent
l
y
, t
h
e i
n
f
o
rm
at
i
on t
h
at
com
e
s from
t
r
ust
w
ort
h
y
source
s sh
ou
l
d
be m
o
re im
po
rt
ant
o
r
carr
y
m
o
re
weigh
t
th
an
less reliab
l
e
on
es. A g
e
n
e
ral way to
co
m
p
u
t
e
th
e inform
at
io
n
fro
m
so
u
r
ces th
at are m
o
re
o
r
less
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISS
N
:
2088-8708
Mob
ile rob
o
t
l
o
ca
liza
tio
n:
a
review o
f
proba
b
ilistic ma
p-
ba
sed
techn
i
qu
es (S
M Ma
lagon
-So
l
da
ra
)
75
t
r
ust
w
o
r
t
h
y
an
d w
h
at
wei
g
ht
s
m
u
st
be gi
ve
n t
o
t
h
e dat
a
of eac
h so
urce
;
i
s
by
m
a
ki
ng
a wei
ghe
d p
o
u
n
d
e
r
ad
d
ition
of th
e
m
easu
r
em
en
ts. Th
is
p
r
o
cess is b
e
tter
kn
own
as
Kalm
an
filter an
d
it is o
n
e
o
f
t
h
e meth
od
s
m
o
re wid
e
ly
used
fo
r sen
s
orial fusion
in mo
b
ile
robo
tic
s appl
i
cat
i
o
ns [1
2]
. In Fi
g
u
r
e 1, a Kalm
an filter is
illu
strated
wh
ere th
e
b
l
o
c
k
s
rep
r
esen
t th
e m
easu
r
em
en
ts
, dev
i
ces, an
d
t
h
e en
v
i
ron
m
en
t. Th
is filter is
u
s
ed
wh
en
th
e system to
b
e
m
o
d
e
led
fails for h
a
v
i
ng
a no
n
lin
ear Gau
ssian
n
o
ise d
i
strib
u
tion.
W
h
ile th
e erro
rs are
ap
pro
x
i
m
a
tel
y
Gau
ssian
, th
e Kalm
an
filter
can
b
e
u
s
ed
nev
e
rt
h
e
less bu
t will p
r
ob
ab
ly n
o
t
b
e
op
ti
m
a
l. For
n
o
n
lin
ear system
s,
th
e ex
ten
d
ed
Kalm
an
filt
er (EKF) is u
s
ed
. Th
is inv
o
l
ves th
e lin
earizatio
n
o
f
th
e p
l
an
t, and
i
f
nece
ssary
,
t
h
e l
i
n
ea
ri
zat
i
o
n
of
t
h
e m
easurem
ent
.
T
hus
, hi
gh
or
der
t
e
rm
s of t
h
e
T
a
y
l
or e
xpa
nsi
o
n a
r
e
can
celled
.
Th
e ex
istin
g lin
earized
error
p
r
opag
a
tio
n in
t
h
e
fam
i
ly o
f
Kal
m
an
filters can resu
lt in
larg
e erro
rs
an
d in
co
nsisten
c
y in
t
h
e simu
ltan
e
ou
s lo
cal
izatio
n
and
m
a
ppi
ng
(S
LAM
)
pr
o
b
l
e
m
.
On
e
approach to all
e
viate
th
is situ
ation
is th
e
u
s
e
o
f
iteratio
n
in th
e EK
F and
t
h
e si
g
m
a po
in
t
Kalm
a
n
filter (SPKF) [1
3
]
.
Fig
u
re
1
.
Typ
i
cal Kalm
an
filt
er ap
p
licatio
n
To
rev
i
ew t
h
e work
do
n
e
b
y
th
e scien
tific commu
n
ity
, th
is
sectio
n
is d
i
v
i
ded
in
to
t
w
o
d
i
fferen
t typ
e
s
o
f
fram
e
wo
rk
s. First, work
s
th
at u
s
e land
mark
s are d
i
sp
layed
.
The m
a
j
o
rity o
f
th
ese
wo
rk
s in
cl
u
d
e
visio
n
sens
ors a
nd t
r
i
a
ng
ul
at
i
on m
e
t
h
o
d
s. A
n
d s
econ
d
, m
e
t
hod
s based
on l
a
ser sens
or
s ar
e sho
w
n
.
The
s
e t
w
o
fram
e
wo
rk
s rep
r
esen
t th
e effo
rt to
i
m
p
r
ov
e th
e so
lu
tio
n
t
o
th
e lo
calizati
o
n
p
r
o
b
l
em
, a
n
d
th
eir im
p
o
r
t
a
n
ce is
hi
g
h
l
i
ght
e
d
i
n
d
i
vi
dual
l
y
.
2.
1.
Landm
ar
k and
trian
g
ul
ati
o
n me
th
ods
As na
vi
gat
i
o
n
st
rat
e
gy
, m
e
t
h
ods
wi
t
h
l
a
n
d
m
arks an
d t
r
ian
g
u
l
atio
n
o
f
si
g
n
a
ls rely on
id
en
tification
of
feature
s
or objects of the e
nvi
ronm
ent. The features a
n
d objects m
u
s
t
be kn
o
w
n a
pri
o
ri
o
r
ext
r
act
ed
dynam
i
cally. The e
nvi
ronm
ent feature
s
are
divided into four types:
1) a
c
tive
beacons
tha
t
are fi
xed at known
p
o
s
ition
s
and
activ
ely tran
sm
i
t
u
ltrason
ic, IR
(infrared)
o
r
RF (rad
i
o
frequ
e
n
c
y
)
sign
al
fo
r th
e calcu
latio
n of
the a
b
sol
u
te robot
position
through
t
h
e direction of
r
ecei
ving i
n
cide
nce
;
2) artificial landm
ark
which a
r
e
specially designed objects
or m
a
rker
s
placed at
known locations i
n
the
envi
ronm
ent; 3) nat
u
ral la
ndmarks
whi
c
h are
di
st
i
n
ct
i
v
e en
vi
r
o
n
m
ent
a
l
feat
ures
and ca
n b
e
ext
r
act
ed
by
sens
ors;
a
nd
4) e
n
v
i
ro
nm
ent
m
odel
s
t
h
at
are
b
u
ilt fro
m
p
r
i
o
r
kn
owledg
e ab
ou
t t
h
e env
i
ro
n
m
en
t an
d
can
b
e
u
s
ed
for
m
a
tch
i
n
g
n
e
w
sen
s
or o
b
s
ervatio
n
s
.
Am
ong t
h
e e
nvi
ro
nm
ent
a
l
feat
ure
s
di
sc
u
ssed,
nat
u
ral
l
a
ndm
ark-
base
d
navi
gat
i
on i
s
fl
exi
b
l
e
beca
use n
o
ex
p
licit artificial lan
d
m
ark
s
are n
e
ed
ed, b
u
t
m
a
y n
o
t
work
well when la
ndm
arks are spa
r
se
or the
envi
ro
nm
ent
i
s
not
k
n
o
w
n a p
r
i
o
ri
.
Al
t
h
o
u
g
h
t
h
e art
i
f
i
c
i
a
l
landm
arks are e
nha
nce
d
an
d m
a
p b
u
i
l
d
i
n
g p
r
ocess
i
s
sim
p
l
i
f
i
e
d.
No
wa
day
s
, t
h
e
em
ergence o
f
vi
sual
se
ns
or
s
has
res
u
l
t
e
d i
n
a t
r
e
n
d t
o
wa
rds
t
h
e
use
of
di
gi
t
a
l
cam
e
ras as t
h
e
m
a
i
n
sens
or t
o
gat
h
er i
n
f
o
r
m
at
i
on. The
si
m
u
lt
aneou
s
p
r
ocess
of l
o
cal
i
zat
i
on a
nd m
a
ppi
ng
through cam
er
as is co
mm
only called visual SLAM and
solve
d
with E
K
F. T
h
e basic
strengt
h of EKF i
n
so
lv
i
n
g
th
e SLAM p
r
ob
lem li
es in
its iterati
v
e
app
r
o
a
ch
of d
e
term
in
in
g
th
e esti
m
a
tio
n
.
Hen
c
efo
r
th
b
u
i
ld
in
g
of
an
au
gm
ent
e
d m
a
p o
f
i
t
s
sur
r
o
u
ndi
n
g
e
nvi
ro
nm
ent
w
h
ere
t
h
e
r
o
b
o
t
s
na
vi
gat
e
t
h
r
o
ug
h
som
e
way
poi
nt
s.
[1
4]
gra
d
ual
l
y
bui
l
d
t
h
e m
a
p
by
con
s
i
d
eri
n
g
i
t
as an
augmentation
of estimated st
ates,
whic
h are nothing but
a col
l
ect
i
on
of
po
si
t
i
ons
of t
h
e feat
ures
(
o
r
l
a
ndm
arks) i
n
t
h
e en
vi
r
onm
ent
,
al
o
n
g
wi
t
h
t
h
e
po
si
t
i
on
of t
h
e
robo
t. Thu
s
, to
so
lv
e th
e l
o
calizatio
n
p
r
oble
m
, th
e rob
o
t
p
o
sitio
n
an
d
t
h
e lo
cation
s
of ob
serv
ed
stat
io
n
a
ry
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
089
-48
56
IJR
A
V
o
l
.
4,
No
. 1,
M
a
rc
h 20
1
5
:
7
3
– 81
76
landm
arks
(for exam
ple line
segm
ents) m
u
st be estim
ated
. Th
e ob
ser
v
ati
o
n-u
p
d
a
te st
ep req
u
i
res th
at
all th
e
lan
d
m
ark
s
an
d
th
e jo
in
t co
v
a
ri
an
ce m
a
trix
b
e
u
p
d
a
ted ev
ery ti
m
e
an
o
b
s
erv
a
tio
n
is m
a
d
e
. Th
is m
ean
s th
at th
e
ext
e
nt
of t
h
e c
o
m
put
at
i
on e
x
pan
d
s
q
u
ad
rat
i
cal
l
y
wi
t
h
t
h
e
num
ber
of l
a
n
d
m
a
rks i
n
a m
a
p. B
e
si
des,
vi
si
on
-
base
d ap
pr
oa
ches p
r
ese
n
t
several
p
r
obl
e
m
s wi
t
h
occl
usi
o
ns, re
al
-t
i
m
e operat
i
o
n,
and en
vi
r
o
n
m
ent
m
o
d
i
ficatio
n
s
.
Co
n
s
equ
e
n
tly,
th
e robo
t can
o
n
l
y d
e
tect th
e p
r
esen
ce
o
f
the tag
s
wh
en
it
is trav
elin
g
i
n
th
eir
pr
o
x
i
m
i
t
y
. As
a resul
t
,
t
h
e i
m
port
a
nce of
com
b
i
n
i
ng t
h
i
s
i
n
fo
rm
ati
on w
i
t
h
dat
a
obt
ai
n
e
d fr
om
ot
her sens
or
s
(e.g
.
o
d
o
m
etry
) is ob
serv
ed
[1
5
]
[16
]
. Below a laser
ran
g
e
find
er is sho
w
n
.
Th
is typ
e
o
f
sen
s
o
r
facilitates th
e
d
a
ta p
r
o
cessi
ng
b
y
th
e lo
calizatio
n
alg
o
rithm
.
Besid
e
s, it
s recog
n
ition
d
o
e
s n
o
t
d
e
p
e
n
d
o
n
ch
ang
e
s in
th
e
envi
ro
nm
ent
.
2.
2.
Laser
range
finder
The l
o
cal
i
zat
i
on sy
st
em
based
on t
h
e l
a
ser
scan
ner
an
d
r
e
t
r
o-
refl
ect
i
v
e
l
a
ndm
arks i
s
a
pr
om
i
s
i
n
g
abs
o
l
u
t
e
p
o
si
t
i
oni
ng t
e
c
hni
qu
e i
n
t
e
rm
s of p
e
rf
orm
a
nce
and cost.
The las
e
r actively
e
m
its a signal and
records
i
t
s
echo, t
h
e o
u
t
p
ut
si
gnal
be
i
ng a l
i
ght
bea
m
. Lasers pr
ov
i
d
e
m
u
ch m
o
re focu
sed
beam
s t
h
an ot
he
r se
nso
r
s
lik
e son
a
rs. Th
is is cru
c
ial wh
en
h
ittin
g
a sm
o
o
t
h
surface
at an
an
g
l
e. [17
]
u
s
e senso
r
fusion
b
e
t
w
een
an
om
ni
di
rect
i
ona
l
cam
e
ra and
a 3D l
a
se
r ra
n
g
e fi
nde
r (
L
R
F
). T
h
i
s
a
p
p
r
o
ach t
a
kes a
d
v
a
nt
age
of t
h
e
m
e
t
r
i
c
i
n
f
o
rm
at
i
on p
r
ovi
ded
by
t
h
e
LR
F an
d c
o
m
b
i
n
es i
t
wi
t
h
t
h
e
om
nidirectional vision. T
h
e
n
cam
era extracts the
vert
i
cal
l
i
n
es i
n
t
h
e e
n
vi
ro
nm
ent
an
d
usi
n
g a
scan m
a
t
c
hi
ng
t
echni
que
, sol
v
es t
h
e
SL
AM
pr
o
b
l
e
m
.
How
e
ver
,
th
e au
t
h
ors
do
n
o
t
con
s
id
er
o
c
clu
s
ion
s
an
d illu
m
i
n
a
tio
n
ch
an
g
e
s.
In [
1
8]
, t
h
e E
K
F i
s
use
d
t
o
l
o
cal
i
ze t
h
e
m
obi
l
e
robot
wi
t
h
a LR
F sens
or i
n
a
n
envi
ro
nm
ent
d
e
m
a
rcated
with
lin
e seg
m
en
ts. Si
m
u
latin
g
th
e k
i
n
e
m
a
tic
m
o
d
e
l o
f
th
e ro
bo
t p
e
rfo
rm
s
a p
r
ed
iction
step
.
A
m
e
t
hod
fo
r es
t
i
m
a
t
i
ng t
h
e c
ova
ri
ances
o
f
t
h
e l
i
n
e pa
ra
m
e
t
e
rs based
on cl
assi
c l
e
a
s
t
squa
res
(L
SQ)
i
s
propose
d
.
Thi
s
m
e
thod is c
o
m
p
ared
with the m
e
t
hod resulting
from
the orthogonal LSQ in te
rm
s of
com
putational
com
p
lexity.
The res
u
lts of a co
m
p
aris
on show that the use of
classic LSQ instead of
ort
h
o
g
onal
LS
Q re
duce
s
t
h
e
num
ber o
f
c
o
m
put
at
i
ons i
n
a l
o
cal
i
zat
i
on al
go
ri
t
h
m
t
h
at
is a part
o
f
a S
L
AM
. I
n
the input noise
cova
riance m
a
trix of
the E
K
F the standa
rd
deviation of each
angula
r
spe
e
d of
robot wheels is
cal
cul
a
t
e
d as
bei
n
g
pr
o
p
o
r
t
i
onal
t
o
t
h
e a
n
gul
a
r
s
p
eed
of the
robot wheels. A corr
ect
i
on
st
ep i
s
pe
r
f
o
r
m
e
d
minimizing the differe
n
ce
between t
h
e m
a
tched line
se
gm
ent
s
fr
om
t
h
e l
o
cal
an
d
gl
obal
m
a
ps [
1
9
]
. If t
h
e
o
v
e
rlapp
i
ng
rate b
e
tween
t
h
e
m
o
st si
milar
lo
cal an
d
g
l
obal lin
e seg
m
en
ts is b
e
lo
w the th
resh
o
l
d
,
the lin
e
segm
ents are
paire
d
. T
h
e line covaria
n
ces
of
pa
ram
e
ters, which a
r
ise
from
the LRF distance-m
easure
m
ent
erro
r, co
m
p
rise th
e
ou
tpu
t
no
ise co
v
a
rian
ce m
a
trix
o
f
the EKF.
Line
s
e
gm
ents were
chosen beca
us
e they
req
u
i
r
e a
sm
al
ler am
ount
of
c
o
m
put
er m
e
m
o
ry
i
n
c
o
m
p
ari
s
on
wi
t
h
t
h
e
occ
upa
ncy
gri
d
s
m
e
t
hod
[
20]
.
Traditionally, m
a
ny nonlinea
r Bayesian estim
a
tion
proble
m
s
are solved
using the EKF
.
But whe
n
the dy
nam
i
c
m
odels and m
easurem
ents are highly n
onli
n
ear a
n
d the
measurem
ent noise is
not Gaussian,
linearized m
e
thods m
a
y not
always be
a good approac
h
[2
1
]
. Po
pu
lar altern
ativ
es to
Gau
ssian techn
i
qu
es are
n
onp
aram
etric
filters. No
np
ara
m
etric filters
d
o
n
o
t
rely
on a fix
e
d
fun
c
tio
n
a
l form
o
f
t
h
e po
steri
o
r, su
ch
as
G
a
u
s
sian
s.
I
n
st
ead
, th
ey app
r
o
x
i
m
a
te p
o
s
terio
r
s b
y
a f
i
n
ite n
u
m
b
e
r
o
f
v
a
lu
es, each
r
ough
ly co
r
r
e
spon
din
g
to
a re
gion in state space
.
3.
PARTICLE FILTER
Particle filters (PF
)
are se
que
n
tial Monte Carlo
m
e
thods
unde
r the Bayes
i
an estim
a
tion fram
e
work
and
have
bee
n
wi
del
y
use
d
i
n
m
a
ny
fiel
ds suc
h
as si
gnal
p
r
oce
s
s
i
ng, t
a
r
g
et
t
r
a
c
ki
n
g
, m
obi
l
e
rob
o
t
localization, i
m
age proce
ssing, a
n
d
various econom
i
cs a
p
p
lication
s
.
Th
e
k
e
y id
ea is to
represen
t t
h
e
n
e
x
t
p
r
ob
ab
ility d
e
n
s
ity fun
c
tion (PDF)
of th
e state v
a
riab
les b
y
a set of rando
m
sa
m
p
les o
r
p
a
rticles with
associated
wei
ghts
,
and c
o
mpute estim
ates
base
d on the
s
e
sam
p
les and
weights. PF ca
n estim
a
t
e the syste
m
states su
fficiently wh
en th
e nu
m
b
er of
p
a
rti
c
les (estim
a
t
i
ons of
t
h
e
st
at
e vect
o
r
s whi
c
h evol
ve
i
n
paral
l
el
)
i
s
large. T
h
e PF
can be a
p
plied to an
y state tran
sitio
n
o
r
m
easu
r
em
en
t
m
o
d
e
l, and
it do
es n
o
t
m
a
tter if so
me
err
o
rs i
n
i
n
ert
i
al
senso
r
s e
xhi
bi
t
com
p
l
e
x st
ocha
st
i
c
ch
ara
c
teristics. Thes
e errors a
r
e ha
rd to m
odel us
ing a
linear estim
a
tor s
u
ch as the Kalm
an
filter because of their high inhe
re
nt nonline
a
rity and random
n
ess.
Ho
we
ver
,
t
h
i
s
m
e
t
hod
has
n
o
t
y
e
t
bec
o
m
e
po
pul
a
r
i
n
t
h
e
i
n
d
u
st
ry
beca
use i
m
pl
em
ent
a
t
i
on det
a
i
l
s
are
m
i
ssi
ng
in
th
e av
ailab
l
e research
literatu
re, and
becau
s
e its co
m
p
utatio
n
a
l co
m
p
lex
ity h
a
s to
b
e
h
a
nd
led
in
real-ti
m
e
appl
i
cat
i
o
ns. T
h
e fi
rst
m
e
t
hod
di
scusse
d i
t
i
s
t
h
e t
r
i
a
ng
ul
at
i
o
n by
W
i
Fi
(I
E
EE 80
2.
1
1
W
L
AN
), w
h
i
c
h co
nsi
s
t
s
in
id
en
tifyin
g
access po
in
ts i
n
th
e env
i
ron
m
e
n
t.
On
e adv
a
n
t
ag
e in u
s
i
n
g
WiFi tech
no
log
y
is its freq
u
e
n
t
u
s
e i
n
in
do
or
env
i
r
onmen
ts.
3.
1.
Wi
Fi
Accord
ing
to
W
i
Fi-allian
ce,
o
v
e
r 700
milli
o
n
p
e
op
le u
s
e
W
i
Fi and
th
ere are ab
ou
t 800
mil
lio
n
n
e
w
W
i
Fi de
vices e
v
ery year. This
freely availabl
e wireless
i
n
fr
ast
r
uct
u
re
pr
o
m
pt
ed
m
a
ny
researche
r
s t
o
de
vel
o
p
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Mob
ile rob
o
t
l
o
ca
liza
tio
n:
a
review o
f
proba
b
ilistic ma
p-ba
sed
techn
i
qu
es (S
M Ma
lagon
-So
l
da
ra
)
77
W
i
Fi-b
ased
positio
n
i
ng
system
s
f
o
r
indo
or
en
v
i
r
o
n
m
en
ts
. Three m
a
in approac
h
es
fo
r
W
i
Fi-b
ased
positio
n
i
ng
exist: tim
e
-based, a
n
gle-base
d, a
n
d
si
gnal-strength-ba
s
ed a
p
proaches
.
Ofte
n tim
e
s, howe
v
e
r
, the
r
e
are no a
v
ailable
W
i
Fi
access
poi
nts and it is necessa
ry to
find a ne
w
m
a
nner
o
f
i
d
e
n
t
i
f
y
i
ng t
h
e e
nvi
ro
nm
ent
.
O
m
ni
di
rect
i
onal
cam
eras repr
esent
a c
h
eap
sol
u
t
i
o
n a
n
d
m
a
ny
feature
s
of t
h
e
envi
ronm
ent can
be e
x
tracted from
an im
age.
3.2.
Om
nidirectional c
a
me
ras and
laser
range finder
Acc
o
r
d
i
n
g t
o
[2
2]
t
w
o m
e
t
h
od
ol
o
g
i
e
s ha
ve
been p
r
eval
e
n
t
i
n
l
i
v
e
m
o
ti
on an
d st
ruct
ure
est
i
m
a
t
i
o
n
fro
m
a sin
g
l
e
m
o
v
i
ng
v
i
d
e
o
cam
era: i)
filterin
g
ap
proach
es th
at fu
se
m
easu
r
em
en
ts fro
m
all i
m
ag
es
sequ
en
tially by u
p
d
a
ting
pro
b
a
b
ility d
i
strib
u
tion
s
ov
er
featu
r
es and
ca
m
e
ra p
a
ram
e
ters; and
ii) bu
nd
le
adj
u
st
m
e
nt
(B
A) m
e
t
hods t
h
at
perf
orm
batch o
p
t
i
m
i
zati
on o
v
er sel
ect
ed
im
ages from
the l
i
v
e st
ream
, such as
slid
in
g wi
n
dow,
o
r
in
p
a
rticu
l
ar sp
atially d
i
strib
u
t
ed k
e
y
f
ram
e
s. In
m
e
n
tio
n
e
d
work pro
b
a
b
ility d
i
strib
u
tion
s
are studied, a
n
d acc
ordingly to loca
lizace t
h
e robot it is important
ha
ve
ob
j
ect rec
o
gnition. Aut
h
ors
like
[23]
b
u
ild
a
p
r
ob
abilistic
o
b
j
ect reco
gn
itio
n. In
p
r
op
o
s
ed
fr
amework
,
a static recog
n
ition
mo
du
le th
at prov
id
es
class prop
ab
ilit
ies for each
p
i
x
e
l of a set
o
f
lo
cal RGB
featu
r
es. For th
is
pu
rpo
s
e two
m
e
th
od
s are
p
r
esen
ted:
i
)
a B
a
y
e
si
an
m
e
t
hod ba
sed
on a m
a
xim
u
m
li
kel
i
h
o
od;
and i
i
)
a ne
ura
l
net
w
o
r
k t
h
at
by
aut
h
or re
su
l
t
s
, i
t
i
s
d
e
m
o
n
s
trate someti
mes wo
rk
b
e
tter th
an
th
e Bayesian
ap
proach
when t
h
e
y
are inte
grated withi
n
a trac
king
fram
e
wo
rk
.
[24
]
, [2
5
]
,
[26
]
stu
d
i
ed
th
e
feasib
ility o
f
th
e tech
n
i
q
u
e
s b
a
sed
o
n
t
h
e g
l
o
b
a
l ap
p
e
aran
ce of a
set of
om
nidirectiona
l im
ages capt
u
red by a
cam
e
r
a to s
o
lve t
h
e lo
calizatio
n
prob
lem
.
First,
th
ey stud
ied ho
w to
descri
be t
h
e
vi
sual
i
n
f
o
rm
at
i
on
gl
o
b
al
l
y
so t
h
at
it correctly represen
ts lo
cation
s
an
d th
e g
e
o
m
etrical
relatio
n
s
h
i
p
s
between
th
ese l
o
catio
ns. Th
en th
ey in
teg
r
ated
th
is in
fo
rm
at
io
n
using
an
ap
pro
ach
b
a
sed o
n
a
sp
ri
n
g
-m
ass-d
a
m
p
er
m
o
d
e
l, to
create a topo
log
i
cal
m
a
p
o
f
th
e env
i
ro
nmen
t. On
ce t
h
e
m
a
p
is b
u
ilt, th
ey
p
r
op
o
s
ed
the use o
f
a Mon
t
e
Carlo
lo
calizatio
n
app
r
o
ach to
estim
a
t
e th
e m
o
st p
r
o
b
ab
le
p
o
s
ition
o
f
t
h
e
v
i
sion
syste
m
an
d
its
traj
ectory wit
h
in
th
e m
a
p
.
[2
7]
prese
n
t
a
m
e
t
hodol
ogy
t
o
bui
l
d
i
n
c
r
em
ent
a
l
t
opol
ogi
cal
m
a
ps. They
use
d
o
m
ni
di
rect
i
onal
im
ages bot
h i
n
ro
bot
m
a
ppi
n
g
an
d l
o
cal
i
zat
i
on. T
h
ese s
o
l
u
t
i
o
n
s
can
be
cat
ego
r
i
zed i
n
t
o
t
w
o m
a
i
n
grou
ps:
feat
ure
-
based a
nd a
p
peara
n
ce-
base
d sol
u
t
i
o
n
s
. I
n
t
h
e fi
r
s
t ap
pro
ach, a num
b
e
r
o
f
sign
if
i
can
t po
in
ts or
reg
i
ons
are ext
r
acted
from
each om
nidirectional im
a
g
e a
nd ea
ch
point is desc
ribe
d
using a
n
invariant de
scri
ptor. All
t
h
e ex
pe
ri
m
e
nts ha
ve
bee
n
ca
rri
ed
o
u
t
wi
t
h
a set
o
f
om
nidirectional im
ages capt
u
re
d by a
catadioptric
s
y
ste
m
m
ounted on the m
obile platform
. Each s
cene is firs
t filtered to avoid lighting de
pende
nce a
nd t
h
en is
descri
bed t
h
r
o
ug
h a F
o
u
r
i
e
r
-
b
ase
d
si
g
n
at
ur
e t
h
at
prese
n
t
s
a go
o
d
pe
rf
or
m
a
nce i
n
t
e
rm
s of am
ount
of
m
e
m
o
ry
and
pr
ocessi
n
g
t
i
m
e
. In t
h
at
wor
k
, t
h
e
aut
h
ors
have
eval
uat
e
d t
h
e i
n
fl
ue
nce o
f
t
h
e desc
ri
pt
o
r
i
n
t
h
e
lo
calizatio
n
b
y
v
a
rying
the num
b
e
r o
f
p
o
s
sible asso
ciatio
n
s
. Th
e system
is
ab
le to
estim
at
e th
e po
sition
o
f
t
h
e
robo
t in
th
e case o
f
an
unk
nown
i
n
itial p
o
s
itio
n
an
d
it is
able to
track
th
e
p
o
s
ition
of th
e
robo
t wh
ile mo
v
i
n
g
.
In
th
e ev
alu
a
ted
m
e
th
o
d
s, as th
ey in
crease the num
b
er of
particles in th
e
syste
m
,
the average
of locali
zation
d
ecreases
rap
i
dly. Also, it is
po
ssib
l
e to
correct th
e
weig
h
ting
o
f
th
e p
a
rticles b
y
co
m
b
in
ing
a ph
ysical syste
m
of forces
with
a Ga
ussian wei
ght.
Ap
pr
oac
h
es
be
fo
re t
h
e
p
r
ese
n
t
,
do
n
o
t
re
p
r
esent
al
l
t
h
e t
echni
que
s
use
d
i
n
t
h
e
vi
sual
fram
e
wo
rk
.
There e
x
i
s
t
ot
h
e
r m
e
t
hods wi
t
h
m
o
re t
h
an a
cam
e
ra l
i
k
e st
ereo
vi
si
o
n
. [
2
8
]
sol
v
e t
h
e SL
AM
p
r
o
b
l
e
m
wi
t
h
an
obs
er
vat
i
on m
odel
t
h
at
c
o
ns
i
d
er
bot
h t
h
e
3D
p
o
si
t
i
ons
and t
h
e S
I
FT
descri
pt
o
r
o
f
t
h
e l
a
n
k
m
a
rks. O
n
e
ad
v
a
n
t
ag
e
o
f
stereo
v
i
sion
is th
e
m
easu
r
e of th
e d
e
p
t
h
an
d th
erefo
r
e th
e
p
o
s
sib
ility o
f
realice a p
r
o
b
a
b
ilistic
m
odel
for vi
s
u
al
od
om
et
ry
. In t
h
e ne
xt
sect
i
on a com
p
i
l
a
t
i
on
of
paral
l
e
l
t
echni
q
u
es i
s
prese
n
t
e
d
,
m
a
ny
of
w
h
ich
ar
e fo
cused
on
r
e
du
cing
th
e nu
m
b
er
of
th
e co
m
p
u
t
atio
n
s
.
4.
OTHER MET
HODS
[29
]
estab
lish
e
s th
at th
e ti
m
e
an
d
m
e
m
o
ry requ
irem
en
ts of th
e basic EKF–
SLAM ap
pro
ach
resu
lt
fro
m
th
e co
st o
f
m
a
in
tain
in
g th
e fu
ll cov
a
rian
ce m
a
trix
, wh
ich
is
O
(
n
2)
w
h
er
e
n
i
s
t
h
e
num
ber
of f
eat
ures i
n
the m
a
p. Many recent e
f
forts
have
co
nce
n
tra
t
ed on
reducing the c
o
m
puta
tional c
o
m
p
lexity of SL
AM in large
envi
ro
nm
ent
s
. Several
c
u
r
r
e
n
t
m
e
t
hod
s ad
dress t
h
e com
put
at
i
o
nal
co
m
p
l
e
xi
ty
pro
b
l
e
m
by
wor
k
i
n
g o
n
a
li
mited
reg
i
on o
f
th
e m
a
p
.
Po
stpon
em
en
t a
n
d
t
h
e co
m
p
ressed
filter significan
tly red
u
c
e th
e co
m
p
u
t
atio
n
a
l
co
st w
itho
u
t
sacr
if
icing
p
r
ecisio
n
,
altho
ugh th
ey r
e
q
u
i
r
e
an
O
(
n
2
)
step
on
th
e to
tal n
u
m
b
er o
f
lan
d
m
ark
s
to
obtain t
h
e full m
a
p. The s
p
lit cova
riance i
n
tersection m
e
thod lim
its the com
putational burde
n
but sacri
f
ice
s
p
r
ecision
: it o
b
t
ain
s
a con
s
erv
a
tiv
e estim
at
e. Th
e sp
arse
ex
tend
ed
i
n
fo
rmatio
n
filter i
s
ab
le to
ob
tain
an
app
r
oxi
m
a
t
e
m
a
p i
n
c
o
n
s
t
a
nt
t
i
m
e
per st
ep
,
except
d
u
ri
ng
l
o
o
p
cl
osi
n
g.
Al
l
ci
t
e
d m
e
t
hods
wo
rk
on
a
si
n
g
l
e
abs
o
l
u
t
e
m
a
p
rep
r
ese
n
t
a
t
i
on,
and co
nf
r
ont
di
ve
rge
n
ce
due to
n
o
n
lin
eari
ties as u
n
certain
ty in
creases when
m
a
ppi
n
g
l
a
r
g
e
areas.
I
n
co
nt
r
a
st
, l
o
cal
m
a
p joi
n
i
n
g a
n
d
t
h
e
co
nst
r
ai
ne
d l
o
cal
su
bm
ap fi
l
t
er,
pr
o
pose
t
o
bui
l
d
sto
c
h
a
stic m
a
p
s
relativ
e to a lo
cal re
fere
nce, gua
ra
nteed
t
o
be
statistically
in
d
e
p
e
nd
en
t. B
y
li
mitin
g
th
e size o
f
th
e lo
cal m
a
p
,
th
is o
p
e
ration is th
e co
n
s
tant ti
me p
e
r step. Lo
cal m
a
p
s
are jo
in
ed
p
e
riod
ically in
to
a g
l
ob
al
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
089
-48
56
IJR
A
V
o
l
.
4,
No
. 1,
M
a
rc
h 20
1
5
:
7
3
– 81
78
ab
so
lu
te m
a
p
,
i
n
a
O
(
n
2
)
st
ep.
Gi
ven t
h
at
m
o
st
of t
h
e u
pdat
e
s ar
e carried
out on a local map, these techniques
also reduce the harm
ful effe
cts
of l
i
n
eari
z
at
i
on. T
o
av
oi
d t
h
e
O
(
n
2) st
ep, t
h
e co
nst
r
a
i
ned rel
a
t
i
v
e s
ubm
ap
filter p
r
o
p
o
s
es
to
m
a
in
tain
th
e in
d
e
p
e
nd
en
t lo
cal m
a
p
stru
ctu
r
e. Each
m
a
p
co
n
t
ains lin
k
s
t
o
o
t
her neigh
b
o
r
i
n
g
m
a
ps, fo
rm
i
ng a t
r
ee st
ruct
ur
e (w
here l
o
op
s
cann
o
t
be
re
p
r
esente
d)
. In t
h
e Atlas fram
e
wo
rk
[3
0]
, net
w
o
r
k
coupled fe
ature
m
a
ps, and c
onsta
nt tim
e
SLAM the lin
ks between loc
a
l
m
a
ps form
an adjacency
gra
p
h.
Th
ese techn
i
ques d
o
no
t i
m
p
o
se lo
op
con
s
ist
e
n
c
y in
th
e
g
r
ap
h
t
h
u
s
sacrifi
c
in
g
th
e
op
ti
mality o
f
th
e resu
ltin
g
gl
o
b
al
m
a
p. Hi
erarchi
cal
SL
AM
pr
o
p
o
s
es a l
i
n
ear t
i
m
e
techni
que t
o
i
m
pose l
oop c
o
nsi
s
t
e
ncy
,
obt
a
i
ni
ng
a
cl
ose t
o
opt
i
m
al
gl
obal
m
a
p.
The Fast
S
L
A
M
t
echni
q
u
e
u
s
es p
a
rticle filt
ers to
es
tim
ate
the ve
hicle trajectory
and eac
h
one
has a
n
ass
o
ciated set of inde
pende
nt E
K
F
to esti
m
a
te
th
e lo
catio
n of
eac
h feat
ure i
n
the
m
a
p.
Th
is p
a
rtitio
n o
f
SLAM into
a lo
calizatio
n
and
a m
a
p
p
i
ng
prob
lem
,
allo
ws to
o
b
tain
a co
m
p
u
t
atio
n
a
l
co
m
p
lex
ity
O
(log
(
n
)) with
the n
u
m
b
e
r o
f
featu
r
es in
th
e map
.
Howev
e
r, its co
m
p
lex
i
t
y
is
lin
ear with
the
num
ber
of
pa
rt
i
c
l
e
s used
. T
h
e
scal
i
ng
o
f
t
h
e
num
ber
of
pa
rt
i
c
l
e
s neede
d
w
i
t
h
t
h
e si
ze a
n
d c
o
m
p
l
e
xi
t
y
of t
h
e
envi
ro
nm
ent
rem
a
i
n
s uncl
e
a
r
. I
n
pa
rt
i
c
ul
a
r
, cl
osi
ng l
oop
s causes dramatic p
a
rticle
ex
tin
ctio
ns that
m
a
p
resu
lts i
n
o
p
t
i
m
istic (i.e., in
co
n
s
isten
t
)
un
certain
t
y esti
m
a
ti
o
n
s
.
In
[3
1
]
,
research
ers i
n
v
e
sti
g
ate th
e
po
ten
tial to
im
p
r
ov
e th
e non
quan
tized
(NQ) m
e
th
o
d
,
b
y
ex
p
l
o
itin
g
th
e
en
tro
p
y
-d
iscri
m
in
ativ
ity rela
tio
n
.
In
th
is
work
th
ey inv
e
sti
g
ate th
e non
quan
tized
rep
r
esen
tatio
n
as a so
l
u
tio
n to
th
e
g
l
ob
al lo
calizatio
n
p
r
ob
lem
.
In
particu
l
ar th
ey
fo
cu
s
on
p
e
rfo
r
m
a
n
ce g
a
i
n
s th
is
rep
r
ese
n
tation
of
fers
o
v
e
r
the
Bo
W (Ba
g
o
f
Wor
d
s
)
m
o
d
e
l an
d
o
f
t
h
e
po
ten
tial to
im
p
r
ov
e efficien
cy and
me
m
o
ry size at a re
duc
ed acc
uracy l
o
ss. As
a first c
o
nt
ri
b
u
t
i
on, t
h
i
s
pa
per
p
r
esen
ts
a c
o
m
p
arative eval
uation
of
qua
nt
i
zed (
Q
) a
nd
no
n
qua
nt
i
zed re
prese
n
t
a
t
i
ons i
n
a ro
b
o
t
l
o
cal
i
zat
i
on t
a
sk. As a sec
o
nd c
ont
ri
b
u
t
i
o
n, t
h
ey
pr
o
pose
m
odul
at
i
ng t
h
e
i
m
port
a
nce
of
feat
u
r
es acc
or
di
n
g
t
o
t
h
e
ent
r
o
p
y
m
easure,
w
h
i
c
h i
s
e
xpe
ri
m
e
nt
al
ly
shown t
o
bene
fit localization accuracy.
As
a third c
ont
ri
bution, it
proposes two a
p
proa
ches t
o
s
p
ee
d
up the
NQ
m
e
t
hod at
r
u
n
t
i
m
e. In t
h
e
fi
rst
a
p
pr
oa
ch,
t
h
ey
pr
o
p
o
s
e a
hi
era
r
chi
c
al
l
o
cal
i
zat
i
on
schem
e
perf
o
r
m
e
d at
two
stag
es is
propo
sed. In
t
h
e secon
d
app
r
o
a
ch
,
ob
j
ectiv
e was to
cap
italize
on
th
e sp
ecifi
cities o
f
th
e t
r
ain
i
n
g
d
a
ta for lo
calizatio
n
.
Exp
e
rimen
t
al resu
lts ob
tain
ed
w
ith th
is m
e
th
o
d
su
ppo
rt its
superio
r
ity in
th
e
g
l
o
b
a
l
l
o
cal
i
zat
i
on t
a
s
k
a
n
d
s
u
g
g
est
t
h
at
pe
rf
o
r
m
a
nce gai
n
s ca
n
be
achi
e
ve
d i
n
t
h
e
l
o
o
p
cl
os
ure
p
r
o
b
l
e
m
.
Vallet
[32
]
p
r
esen
t a m
o
b
ile ro
bo
t used
t
o
sim
u
ltaneously locate the
no
de
s of a
wi
rel
e
ss
net
w
or
k a
n
d
calibrate the param
e
ters of received signal
strengt
h.
The
y
assum
e
that
the position of all the nodes is
unknown and
use a m
obile robot, ca
pable
of SLAM, a
s
a
m
obile beacon.
While
the robot m
oves around, i
t
bui
l
d
s a m
a
p o
f
t
h
e e
n
vi
r
onm
ent
usi
n
g
a l
a
s
e
r sca
nne
r a
n
d
o
dom
et
ry
i
n
fo
rm
ati
on. T
h
us,
i
t
s
po
si
t
i
on
w
i
t
h
i
n
the m
a
p is known at a
n
y mom
e
nt. As the
robot m
ove
s, it also collects RSS (r
eceived signal
strengt
h)
measu
r
em
en
ts
fro
m
th
e n
o
d
e
s o
f
th
e n
e
twork
.
All th
is in
form
at
io
n
is th
en ex
p
l
o
ited
to
esti
m
a
te
th
e p
o
s
itio
n
of the
node
s.
The e
fficacy
of the
m
odels c
a
n
be c
o
m
p
are
d
using t
h
e likelihood of the
data. Howeve
r, t
h
e
au
tho
r
s also
con
s
id
er th
at a
mo
re m
ean
in
gfu
l
co
m
p
aritiv
e
measu
r
e in
the co
n
t
ex
t o
f
th
is research
is to
u
s
e th
e
real
err
o
r
of
m
a
xim
u
m
li
kel
i
hoo
d (M
L
)
p
o
si
t
i
on est
i
m
ates. O
n
e ad
van
t
age of t
h
e M
L
fo
rm
ul
at
i
on of t
h
e
lo
calizatio
n
p
r
o
b
l
em
is th
at it do
es
no
t requ
ire calcu
la
t
i
n
g t
h
e i
nve
rse m
odel
of t
h
e R
S
S
-
di
st
ance
,
w
h
i
c
h c
a
n
be
di
f
f
i
c
ul
t
.
In
pa
rt
i
c
ul
ar,
i
f
t
h
e R
S
S
-
di
st
a
n
ce m
a
ppi
ng
is no
t
b
i
j
ectiv
e, th
e i
n
v
e
rse mo
d
e
l
(d
istan
ce-RSS)
m
i
ght
cont
ai
n
several
di
st
a
n
c
e
s fo
r t
h
e sam
e
R
SS val
u
es. T
h
i
s
can be a se
r
i
ous
dra
w
bac
k
fo
r som
e
al
gori
t
hm
s
that require a
direct estim
a
te
of th
e
di
st
anc
e
fr
om
R
SS, and
req
u
i
r
es a
d
di
t
i
onal
w
o
r
k
t
o
ch
o
o
se bet
w
een t
h
e
pos
si
bl
e al
t
e
rn
at
i
v
es. The M
L
fo
rm
ul
at
i
on of t
h
e
pr
obl
em
sim
p
l
y
does not
suf
f
er
fr
om
t
h
i
s
pr
obl
em
,
and i
t
can
wo
rk
with an
y fun
c
tion
o
f
th
e
d
i
st
ance
, as long as the
m
odel is a va
lid
PDF. To
l
earn
m
o
re abou
t
W
i
Fi
si
gnal
st
ren
g
t
h
sens
ors
,
re
ad
[
33]
,
[
34]
,
[
35]
,
[
36]
.
5.
DIS
C
USSI
ON
There
are
studi
e
s which com
p
ares
the effectivene
ss
of E
K
F and
PF s
u
c
h
a
s
[37], whe
r
e t
h
e E
K
F ha
s
been
em
pl
oy
ed f
o
r
t
h
e l
o
cal
i
zat
i
on
of a
n
a
u
t
o
n
o
m
ous ve
hi
cl
e by
f
u
si
n
g
d
a
t
a
com
i
ng fr
o
m
di
ffere
nt
se
nso
r
s.
In t
h
e E
K
F a
p
proach t
h
e st
ate vector is
approxim
a
ted
b
y
a Gau
ssian rando
m
v
a
riab
le,
wh
ich
is
th
en
p
r
op
ag
ated
an
alytical
ly th
ro
ugh
th
e
first
o
r
d
e
r lin
eariza
t
i
o
n of
t
h
e no
nl
i
n
ea
r
sy
st
em
. The s
e
ries approximation
i
n
t
h
e
EK
F al
go
ri
t
h
m
can,
ho
we
ver
,
l
ead
t
o
po
o
r
rep
r
e
s
ent
a
t
i
ons
o
f
t
h
e
no
nl
i
n
ea
r
f
unct
i
o
ns
an
d
of
t
h
e
asso
ciated prob
ab
ility d
i
stribu
tio
ns.
As a
resu
lt, so
m
e
ti
mes th
e filter
will b
e
d
i
v
e
rg
ent. Related
work
h
a
s
sh
own
th
at th
e p
a
rticle filter is sup
e
rior th
an th
e EKF
i
n
term
s
o
f
th
e accuracy o
f
t
h
e state v
ector esti
m
a
tio
n
,
as well as in
term
s o
f
rob
u
stness and
to
lerance to
m
easu
r
emen
t n
o
i
se. The p
e
rform
a
n
ce o
f
th
e
p
a
rticle filter
alg
o
rith
m
d
e
p
e
n
d
s
o
n
th
e num
b
e
r o
f
p
a
rticles and
t
h
eir initializa
tio
n
.
It
can
b
e
seen that th
e PF al
g
o
rith
m
s
g
e
n
e
rate
b
e
tter esti
m
a
tes o
f
the state v
ector
of th
e m
o
b
ile
robot as
the
number
of pa
rticles
increases
, but at
the
expe
nse
o
f
hi
g
h
er
com
put
at
i
o
nal
ef
fo
rt
.
Th
e
o
p
tim
al fi
lter for a lin
ear m
o
d
e
l with
Gau
s
sian
n
o
i
se is th
e Kalm
an
filter. State esti
m
a
tio
n
for
no
nl
i
n
ea
r sy
st
em
s wi
t
h
no
n-
G
a
ussi
an
noi
se i
s
a di
ffi
cul
t
p
r
obl
em
and i
n
gene
ral
t
h
e o
p
t
i
m
a
l
sol
u
t
i
on cann
o
t
be e
x
presse
d in cl
osed-form
.
In order t
o
i
n
crease the
accu
ra
cy of vis
u
al
SL
AM it is
us
ually m
o
re profitable to
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Mob
ile rob
o
t
l
o
ca
liza
tio
n:
a
review o
f
proba
b
ilistic ma
p-ba
sed
techn
i
qu
es (S
M Ma
lagon
-So
l
da
ra
)
79
increase the
num
ber of feat
ures than
t
h
e num
b
e
r o
f
fram
e
s
. Th
is is th
e
ke
y reason why
BA is m
o
re efficient
than filtering for vis
u
al SLAM. On
the other hand, the P
F
suffers from
the so-called s
a
m
p
le im
poverishm
e
nt
problem in which sam
p
les
tend to conve
r
ge to a co
nfi
n
ed re
gi
on in the solution
space, m
a
king state
est
i
m
a
ti
ons t
r
appe
d i
n
l
o
c
a
l
opt
i
m
a. In [3
8]
, sam
p
l
e
s of
part
i
c
l
e
s are up
dat
e
d
and
pr
opa
ga
t
e
d by
im
pl
em
ent
i
ng a seq
u
ent
i
a
l
i
m
port
a
nce sa
m
p
li
ng (
S
I
S
)
p
r
oces
s rec
u
rsively as new
m
easurem
ent inform
ation
becom
e
s available. As the
num
b
er of
samp
les b
e
co
m
e
s
v
e
ry large and ap
pro
a
ch
es infin
ity, th
e SIS
p
a
rticle
filter app
r
ox
imates th
e
o
p
tim
a
l
Bayesian
estimate.
ACKNOWLE
DGE
M
ENTS
Th
e au
tho
r
s
wan
t
to
t
h
ank fo
r th
eir
fin
a
n
tial su
ppo
rt t
o
:
-
C
onse
j
o
Naci
o
n
al
de
C
i
en
cia
y Tecnolgia, M
e
xico.
-
FOF
I
, U
n
i
v
e
r
si
dad
A
u
t
o
n
o
m
a
de Q
u
eret
ar
o,
M
e
xi
co.
REFERE
NC
ES
[1] C.
Gamallo
, et
al.
, "Omnivision-based kld-M
onte Carlo localization,"
Robotics and Autonomous Systems
, vo
l.
58(3), pp
. 295-3
05, 2010
.
[2] A.
Gasparri
, et al.
, "Monte
car
l
o
filter
in m
obil
e
roboti
c
s locali
zat
ion: a
cluster
e
d evolut
ionar
y
point of view
,"
Journal of Intelligent and
Robotic Systems
, vol. 4
7
(2), pp
. 155-17
4, 2006
.
[3] M.-H.
Li
, et a
l
.
, "Novel indoor
mobile robot n
a
vi
gation using
monocular visio
n
,"
Engineering
Applica
tions of
Artifi
cial
Int
e
ll
ig
ence
, vo
l. 21(3),
pp. 485-497
, 20
08.
[4] Z.
Sun
, et
al.
, "Inferring laser-scan match
i
ng
uncerta
inty
w
i
th conditional
random fields,"
Robotics and
Autonomous Sys
t
ems
, vol. 60(1
)
,
pp. 83-94
, 2012
.
[5] M.
D’Souza
, et
al.
, "Evaluat
io
n of realtim
e p
e
ople
tracking f
o
r
indoor envi
r
onm
ents using ubiquitous m
o
ti
on
sensors and limited wire
less ne
twork infrastruc
t
ure,"
Pervasive and Mobile Co
mputing
, vol. 9(
4), pp. 498-515,
2013.
[6] D.
Fill
iat
, et
al.
,
"Map-based nav
i
gation
in
mobile robots::
I
.
a r
e
view of
loca
lization stra
tegies,"
Cognitiv
e S
y
ste
m
s
Research
, vol. 4
(
4), pp
. 243-282
, 2003.
[7] Y.
Elor
, et al.
, "A “thermody
namic” approach to mu
lti-robot cooperative
localization,"
Theoretical Computer
Scien
c
e
, vol. 45
7, pp
. 59-75
, 20
12.
[8] M.
Ahmadi
, et
a
l
.
, "Attitud
e
estimation b
y
divid
e
d diffe
ren
ce
filter in quaternion s
p
ace,"
Acta Astr
onautica
, vol. 75
,
pp. 95-107
, 201
2.
[9] K.
Lee
, et al.
,
"Kinematic par
a
meter calibratio
n of a car-like mobile robot to
improve odometr
y
accur
a
cy
,
"
Mechatronics
, v
o
l. 20(5)
, pp
. 58
2-595, 2010
.
[10] R.
Siegwar
t
, et
al
.
, "Introduction
to au
tonomous mobile robots:
MIT press; 2011
.
[11] E.
Besada-Portas
, et al.
, "Lo
c
alization of Non-Line
arly
Modeled Autonomous Mobile Robots Using
Out-of-
S
e
quence
M
eas
u
r
em
ents
,"
Sensor
s
, vol. 12(3)
, pp
. 2487-2518, 201
2.
[12]
L. F. de Melo
, et al.
, "Tr
a
jector
y
pl
anning f
o
r nonholonom
ic m
obile
robot using extended Kalm
an filter
,
"
Mathematical Problems
in Engin
eering
, vol. 201
0, pp
., 2011.
[13] K.
Shojaei
, et a
l
.
, "Exp
erim
enta
l
Stud
y
of It
era
t
e
d
Kalm
an Filter
s
for Sim
u
ltaneous Locali
z
a
tion
and Mapping of
Autonomous Mo
bile Robots,"
Jo
urnal of In
tellig
ent
&
Robotic Sys
t
ems
, vol. 63(3
-
4
)
, pp
. 575-594
, 2
011.
[14] A.
Chatter
j
ee
, et al.
, "Developm
ent of a real-lif
e EKF based SL
AM
sy
stem for mobile robots emplo
y
ing vision
sensing,"
Expert Systems with
Ap
plications
, vo
l. 3
8
(7), pp
. 8266-8
274, 2011
.
[15] M.
Boccadoro
, et
al.
, "Constrained and quan
tized Kalman filteri
ng for an RFI
D
robot localization prob
lem,"
Autonomous Ro
bots
, vol. 29(3-4
)
, pp
. 235-251
, 2
010.
[16] K.
Yu
, et al.
, "
L
ocalization
algor
ithm using a vi
rtual label for
a mobile robo
t in
in
door and outdoo
r environments,"
Artifi
cial
Lif
e
an
d Robot
ics
, vol.
16(3), pp
. 361-3
65, 2011
.
[17] G.
Gallegos
, et
al.
, "Indoor SLAM based on
composite sensor
mixing
laser scans and omnidir
ection
a
l images,"
in
Robotics
and Au
tomation (
I
CRA)
, 2010
IE
EE In
ternational Conference on
; 2010:
I
EEE, Year, pp. 3
519-3524.
[18] L.
Tesli
ć
, et al.
, "Using a LRF sensor
in the Kalm
an-filter
i
n
g
-based localization of a m
obi
le robot,"
ISA
transactions
, vo
l. 49(1)
, pp
. 145-
153, 2010
.
[19] A.
Llar
ena
, et al.
, "Odom
etr
y
-Ba
s
ed Viterbi Lo
ca
liz
ation with Art
i
fi
ci
al Neur
al Ne
tworks
and Laser Range Finders
for Mobile Robo
ts,"
Journal of
In
telligent
&
Robo
tic Systems
, vol.
66(1-2), pp
. 75-
109, 2012
.
[20] L.
Tesli
ć
, et al.
, "EKF-based localization of
a wheeled mobile robot
in stru
ctured
environments,"
Journal o
f
Intell
igen
t
&
Robotic
Systems
, v
o
l. 62(2)
, pp
. 18
7-203, 2011
.
[21] P.
Axelsson
, et
al.
, "Bay
es
ian state
estimation o
f
a flexib
le indu
strial robot,"
Co
ntrol Engineerin
g Practice
, vo
l.
20(11), pp
. 1220
-1228, 2012
.
[22] H.
Strasdat
, et
al
.
,
"Visual SLAM: wh
y fi
lte
r
?
,"
I
m
age and Vision
Computing
, vol. 30(2), pp. 65-77
, 2012
.
[23] F.
Se
rra
tosa
, et al.
, "A probabi
li
stic int
e
gra
t
ed o
b
jec
t
recogn
itio
n and track
ing f
r
am
ework,"
Ex
pe
rt Sy
ste
m
s with
Applica
tions
, vo
l. 39(8)
, pp
. 730
2-7318, 2012
.
[24] Z.
Chen
, et
al.
, "Visual detection
of lintel-o
cclud
ed doors b
y
in
tegrating m
u
ltipl
e
cues using a dat
a
-driven Marko
v
chain
Monte Car
l
o process,"
Rob
o
tics and
Autono
mous Systems
, v
o
l. 59(11)
, pp
. 9
66-976, 2011
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
089
-48
56
IJR
A
V
o
l
.
4,
No
. 1,
M
a
rc
h 20
1
5
:
7
3
– 81
80
[25] M.
Hentschel
, et
al.
, "An Adaptive Memor
y
Model for
Long-Term
Navigation o
f
Autonomous
Mobile Robots,"
Journal of Robotics
, vo
l. 2011, p
p
., 2012.
[26] B.
Bacca
, et
a
l
.
, "Appearan
ce-based mapping
a
nd localizatio
n for mobile r
obot
s using a
featur
e stability
histogram,"
Rob
o
tics and
Autono
mous Systems
, v
o
l. 59(10)
, pp
. 8
40-857, 2011
.
[27] L.
Pay
á
, et al.
,
"Map building and monte carlo
localization usin
g global app
ear
an
ce of omnidir
ectional
images,"
Sensors (
B
asel,
Switzerland)
, vo
l. 10(12)
, pp
. 11
468, 2010
.
[28]
F. A. Moreno
, et al.
, "Stereo vi
sion specifi
c m
odels for part
icl
e
filter
-
based SL
AM,"
Robotics and
Autonomous
Systems
, vol. 57
(
9
), pp
. 955-970
,
2009.
[29]
S. S. Ge, "Auton
omous mobile
robots: sensing, co
ntrol, decision m
a
king
and
applications: CRC press; 2010.
[30] M.
Bosse
, et al.
, "An Atlas framework for scalable mapping,"
in
Roboti
c
s and Automation, 2
003 Proceedin
g
s
ICRA'03 I
EEE I
n
ternational Co
nference on
; 200
3: IEEE, Year, p
p
. 1899-1906
.
[31]
F. M. Campos
, et
al.
, "Global
localization with
non-qua
ntized
local
im
age f
eatures,"
Robotics and
Autonomou
s
Systems
, vol. 60
(
8
), pp
. 1011-102
0, 2012
.
[32] J.
Vallet
, et al.
,
"Simultaneous RSS-based localiz
ation and
model calib
ration in
wi
reless networks with a mobile
robot,"
Procedia
Computer Scien
c
e
, vol. 10
, pp
. 1
106-1113, 2012
.
[33] M.
Atia
, et al.
, "Particle-Filter-
Based WiFi-Aid
ed Reduced In
ertial Sensors Na
vigation
S
y
stem for Indoor an
d
GPS-Denied En
vironments,"
International Journ
a
l of Naviga
tion
and Observation
, vol. 2012
, pp
.,
2012.
[34]
J. M. Alonso
, et al.
, "Enhan
ced
WiFi localizatio
n sy
stem based on So
ft Computi
ng techniqu
es to deal with small-
s
cale
var
i
at
ions
i
n
wire
le
ss se
nsors,
"
Applied So
ft
Computing
, vo
l.
11(8), pp
. 4677-
4691, 2011
.
[35]
G. A. Hollinger
, e
t
al
.
, "Target tracking
without
line of
sight using range
from radio,"
Autonom
ous Robots
, vol.
32(1), pp
. 1-14
,
2012.
[36] J.
Zhou
, et al.
, "
R
FID localizatio
n algorithms
and applications—
a
rev
i
ew,"
Journal of In
telligen
t Manufacturing
,
vol. 20(6)
, pp
. 6
95-707, 2009
.
[37]
G. G. Rig
a
tos,
"Extended
Kalm
an
and Par
t
i
c
le
Filtering
for sen
s
or fusion in m
o
tion con
t
rol of
m
obile robots,"
Mathematics an
d computers in
s
i
mulation
, vol. 8
1
(3), pp
. 590-60
7, 2010
.
[38] B.
Cao
, et al.
, "I
mproved particle fi
lter
based on
fine
resampling
algorithm,"
The Journal of Chin
a Universities of
Posts and Telecommunications
,
vol. 19(2)
, pp
. 1
00-115, 2012
.
BIOGRAP
HI
ES
OF AUTH
ORS
Salvador M. Malagon-Soldar
a is stud
y
i
ng his Ph.D
. in the Univer
sidad Autonoma de Queretaro.
He has com
p
leted his M.Sc. in instrum
e
ntation
and autom
a
tic
control in the s
a
m
e
universit
y.
Terminal lin
es o
f
Salvador
ar
e ar
tificial
in
tellig
en
ce, robot control a
nd embedded
sy
stems.
Edgar A
.
Riv
a
s-Araiza receiv
e
d
his M. of
Sc.
degre
e
i
n
Aut
o
ma
tic
Cont
rol
from the
Uni
v
e
r
sity
of Querét
aro
and
his Ph.D. from
t
h
e sam
e
institu
ti
on in 2007
. His
r
e
search
in
terests
include
signal pro
cessin
g
, computer
vision and motion
control.
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISS
N
:
2088-8708
Mob
ile rob
o
t
l
o
ca
liza
tio
n:
a
review o
f
proba
b
ilistic ma
p-
ba
sed
techn
i
qu
es (S
M Ma
lagon
-So
l
da
ra
)
81
Manuel
Toled
a
n
o
-A
y
a
l
a
r
e
c
e
ive
d
his M.
of Sc
. d
e
gree
in
Autom
a
tic
Control
from
the Univ
ersi
t
y
of Querétaro and his Ph.D in 2010. His r
e
s
earch
interests include sign
al pro
cessing,
telecomunicatio
ns, ren
e
wable
en
erg
y
,
and b
i
os
y
s
t
e
m
s
.
Genaro Soto-Z
ar
azúa
rec
e
iv
ed hi
s M. of Sc. d
e
gr
ee in
Autom
a
tic
Control from
the
Universit
y
of
Querétaro
and h
i
s Ph.D. from th
e same institutio
n in 2010. His r
e
search
interests
includ
e signal
processing, auto
mati
on, and
b
i
osy
s
tems.
Roberto V. C
a
r
r
illo-Serr
a
no re
c
e
ived h
i
s engin
eering in
instru
m
e
ntation and process
control
bachelor degree, instrumentation
and automatic control master de
gree, and
doctor in
engineerin
g
degree b
y
the
Universidad Autónoma de Quer
étaro in 2000, 2008, and 2011, respectively
.
Roberto worked
in Kellogg d
e
México from 1999 to 2006.
His research
ar
eas are
robot
manipulators co
ntrol, electr
i
c
m
achin
es
contro
l,
and under
actu
a
ted mechan
ical s
y
stems contro
l.
Nowaday
s
, Rob
e
rto is professor in Universidad
Au
tónoma de Querétaro and he
is member of the
S
N
I (national s
y
s
t
em
of res
earch
ers
)
in M
é
xi
co.
His publications
are ind
e
xed
in JRC datab
a
se of
ISI-Thomson.
Evaluation Warning : The document was created with Spire.PDF for Python.