Professional Documents
Culture Documents
EKFMehmet
EKFMehmet
FiltresiKestirimiYaklamlarnnKarlatrlmas
FredrikOrderud
SemSlandsvei79,NO7491Trondheim
Trkeyeeviren:M.MehmetNEFES,Ankaraniversitesi
zet
Geniletilmi Kalman Filtresi (GKF) , esasen
basit,salamvegerekzamanluygulamalara
uygun olmasndan dolay dorusal olmayan
durumuzaymodellerikestirimiiinfiiliolarak
uygulanan bir standart haline gelmitir [11].
Bununla birlikte, son yllarda Unscented
Kalman Filtresi (UKF) olarak adlandrlan
alternatif bir yaklam da ortaya kmtr. Bu
filtre, dorusal olmayan modeller iin daha
yksek doruluk ve salamllk salamay
amalamaktadr.
UKFnin
doruluunu
dorusal olmayan sre modelleri iin
aratran bir ka makale olmasna karn
bunlarn hi biri zellikle dorusal olmayan
lm modellerinin doruluuna hitap
etmemektedir. Bu makale, dorusal olmayan
lmlere sahip iki takip modeli iin GKF ve
UKF performanslarn karlatrarak bu
boluudoldurmayamalamaktadr.
yaklamgeniletilmiKalmanfiltresininiteler.
Bununla birlikte, dorusallatrma kararsz
kestirimler gibi baz problemler ortaya
karabilir[8].Dorusalolmayansistemleriin
daha iyi kestirici algoritmalarn gelitirilmesi
abalar, ortaya kacak yenilikliklerin
mhendislik alanlarnda geni bir yelpazede
byk
yank
bulmasnn
kanlmaz
olacandan bilim dnyasnca byk ilgi
grmektedir.
Notasyon
1Giri
2GenelBilgi
Durumuzaymodeli,durumunun(x)saysalbir
vektr ile ifade edildii bir srecin
matematiksel modelidir. Durum uzay
modelleri iki ayrk model ierirler. Bu
modeller; girdi ve grlt gibi d etkenler ile
durumun zaman ierisinde nasl ilerlediini
anlatan sre modeli ve sreten lmlerin
nasl alnacan ifade eden genellikle de
grltl ve/veya doru olmayan lmleri
kullananlmmodelidir.
2.1GenelDurumUzayModeli
Durum uzay modellerinin en genel biimi
dorusal olmayan modeldir. Bu model tipik
olarakikifonksiyonu(fveh)ierir.
ekil2:DorusalDurumUzayModeli
Bu dorusal modelde hesaplama ve analiz
kolaydr.
Kullancya
denetlenebilirlik,
gzlenebilirlik, frekans yant gibi zellikleri
irdelemeimknsunar.
Durumuzaymodellerihemenhemenheresit
srecin modellenmesinde olduka kullanlr.
f ve h fonksiyonlar srecin dinamiini ve
gzlemlerini belirten kesikli hale getirilmi
diferansiyel denklemler zerine kurulmu
fonksiyonlardr.
3DurumKestirimi
ekil1:GenelDurumUzayModeli.Z :Birim
gecikmefonksiyonu(Saysalsinyalilemede
kullanlanZdnm)
2.2DorusalDurumUzayModeli
Dorusal durum uzay modeli, f ve h
fonksiyonlarnn durum ve girdide dorusal
olduu modeldir. Dorusal olan bu
fonksiyonlar, hesaplamalar dorusal cebire
indirgeyen F, B ve H matrisleri ile ifade
edilebilir. Bu ekildeki durum uzay modeli u
ekildegsterilir:
Durum
kestirimi,
durumu
dorudan
gzlenemeyen bir srecin olaslk younluk
fonksiyonunu (oyf) tahmin etme (kestirme)
problemi ile ilgilenir. Bu olgu, hem gnceli
kullanarak bir sonraki durumu tahmin etmeyi
hem de yaplan bu tahmini, alnan grltl
lmleri kullanarak gncellemeyi/dzeltmeyi
kapsar.
3.1ndirgemeliBayesKestirimi 1
Durum kestirimi iin bilinen en genel biim
indirgemeli Bayes kestirimidir [1]. Bu biim,
sistemi ve lm modeli verilen bir srecin
durumunun olaslk younluk fonksiyonunu
tahminetmedeenuygunyoldur.Bublmde,
indirgemeliolarakherbirzamanadmndabir
nceki zaman admnda elde edilen kestirimi
ve yeni lmleri baz alarak yeni bir kestirim
yapanbukestiriciyiirdeleyeceiz.
ing.:RecursiveBayesianEstimation
ekil3:ndirgemeliBayesKestiricisidngs
BuradaF,BveHzamanabalmatrislerdir.
2
3
ing.:StatePropagationBelief
ing.:MeasurementLikelihood
ekil4:Kalmanfiltresidngs
lmler
alndktan
gncelleme:
sonra
durumu
lmler
alndktan
gncelleme:
sonra
durumu
3.4UnscentedKalmanFiltresi
3.4.1Giri
Rastgele Gauss deikenlerinin dorusal
olmayan bir fonksiyon zerinde ilerletilmesi
problemi unscented dnm olarak
adlandrlan baka bir yntemle ele alnabilir.
Bu dnm, fonksiyonlar dorusal hale
getirmek yerine belirli bir noktalar kmesini
kullanr ve bu noktalar dorusal olmayan
fonksiyon zerinde ilerletir. Bu noktalar,
ortalamas,kovaryansvemuhtemelenyksek
dereceli momentleri rastgele Gauss deikeni
ile eleecek ekilde seilir. Ortalama ve
kovaryans,
fonksiyon
zerindeki
bu
noktalardan klasik fonksiyon dorusallatrma
yntemine gre daha doru sonu verecek
ekildeyenidenhesaplanabilir.Buyntemdeki
temel fikir; fonksiyonun kendisinin yerine
olaslk dalmna yaklam salamaktr. Bu
strateji genel olarak hem hesaplama
karmakln azaltr hem de daha doru ve
hzl sonular ortaya kararak kestirim
doruluunuartrr.
3.4.2GenelBilgi
Unscented dnmn temelini tekil eden
metod ilk olarak Julier, Uhlmann ve Durrant
3.4.3TakviyeliDurum(AugmentedState)
Gauss olmayan ya da aditif olmayan
grltleri oluturmak iin grltnn
dorusal olmayan bir usul ile ele alnabilir
olmas unscented dnm yaklamnn bir
baka avantajdr. Bu stratejideki ilem
grltnn
fonksiyonlar
zerinde
ilerletilmesiniierir.lkolarakdurumvektr,
grlt
kaynaklarn
hesaba
katmak
maksadyla geniletilir, bir baka deyile
takviye edilr. Bu yntem ilk olarak Julier
tarafndan [7] nolu makalede ortaya
koyulmutur, daha sonra ise Merwe
tarafndan [16] nolu makalede daha detayl
olarak incelenmitir. Bir sonraki aamada ise
grlt deerlerini de ieren bu takviyeli
durumdan(xa)sigmanoktasrnekleriseilir.
Bu ilemin net sonucu ; sre ve lm
grltlerinin dorusal olmayan etkilerinin
durumun geri kalan ile ayn dorulukla elde
edilmesi,vedolaysylaaditifolmayangrlt
kaynaklariindoruluuartrlmasdr.
3.4.4FiltreFormlasyonu
Filtre, durum vektrnn L boyutuna kadar
geniletilmesi, takviye edilmesi ile balar.
Burada L orijinal durum vektr, model
grlts ve lm grlt boyutlarnn
toplamdr. Benzer ekilde kovaryans matrisi
de L2 boyutunda bir matris olacak ekilde
geniletilir. Bu biimde takviye edilen durum
tahmin vektr xa ve kovaryans matrisi Pa u
ekildeifadeedilir:
ialtsimgesikovaryansmatrisininkarekknn
4
inci sutnuna karlk gelmektedir. 0<1
aralnda olan parametresi sigma noktas
yaylmnbelirler.Buparametrelokalolmayan
etkilerden saknmak iin genelde ok kk
seilirvetipikdeeri0.001civarndadr.
Elde edilen ka1 matrisi bu halde durumu
ieren kx1 , rneklenmi sre grltsn
ieren kw1 verneklenmilmgrltsn
ieren kv1 stunlarnaayrtrlabilir.
Her bir sigma noktas bir bal deere atanr.
Bu bal deerler, sigma noktalar
momentlerinin Gauss dalm varsaym
altnda modellerin Taylor serisi alm ile
karlatrlarak [9] nolu makalede de
belirtildii gibi elde edilir. Elde edilen bal
deerler, ortalama (m) ve kovaryans (c)
tahminlerindeuekildekullanlr:
3.5DierYaklamlar
MerkeziFarkKalmanFiltresi(MFKF),[4]nolu
makalede Gauss dalma sahip olaslk
younluk fonksiyonlarnn dorusal olmayan
fonksiyonlar
zerinde
ilerletilmesinde
kullanlabilecek alternatif bir Kalman
formlasyonu olarak nerilmitir. [13] nolu
makalede bu formlasyonun esasen farkl bir
yapsolmasnaramen,yaplantestlerdeelde
edilen sonularn UKF sonular ile ayrt
edilemez bir ekilde benzerlik tad ifade
edilmektedir.BuyzdenMerkeziFarkKalman
Filtresi (MFKF) bu makalede kapsam d
tutulmutur.
Tanecik (Partikl) filtreleri [1]: eitli trldeki
Kalman filtreleri ounlukla yksek dorulua
sahip kestirimler salamasna karn baz
kestirim problemlerine uygun olmad
durumlar da olabilir. Bu durum, Kalman
filtrelerinin sadece Gauss olaslklarn
modelleyebilmesindenvedolaysylaarpkya
da farkl formlardaki dalmlar iin uygun
olmamasndan kaynaklanr. Bu yzden Gauss
olmayan dalmlar da kapsamak iin daha
genel bir yaklama ihtiya vardr. nem
rneklemesini kullanan ardk MonteCarlo
simulasyonlarna dayal tanecik (partikl)
filtreleri bu durumlarda sklkla kullanlan iyi
biralternatifyaklamdr.Bumakaledesrecin
olaslkyounlukfonksiyonuntamamnadeil,
sadece en olas durumunun kestirimine
odaklanlmtr ; bu nedenle tanecik (partikl)
filtresi yaklamnn burada ele alnmasna
gerek yoktur. Ayrca, tanecik (partikl)
filtreleri Kalman filtrelerine gre ounlukla
daha karmak hesaplamalar iermesinden
dolay bu tip filtrelerin gerek zamanl
dzeneklerdekullanlmasoldukagtr.
4Deneyler
Buksmdaelealnandeneyler,dorusalsre
modellerine ve dorusal olmayan lm
modellerine sahip pratik takip uygulamalar
iin GKF ve UKF performanslarnn arasndaki
farkllklarortayakarmayamalamaktadr.
4.1SreModeli
ekil7:Radarlauaktakibi
4.3genlemeMetoduileTakip 6
durumda
balad
Bumodeldegzlemevlerininsrasyla(300,0)
ve(300,0)koornidatnoktalarndakonuland
ve n1, n2 lm grltleri varyanslarnn her
ikisininde200olduuvarsaylmtr.
olaraktanmlanmtr.
4.2RadarlaTakip
Radarla takip ya da radarla izleme hedefin
uzakln (d) ve asn () gzlemleyen
dorusalolmayanbirlm/gzlemmodeliile
modellenebilir:
ekil8:genlememetoduileuaktakibi
Wienersreleribeyazgrltileentegre
srelerdir.
ing.:TrackingbyTriangulation
ekil10:KestirimDorulukDalmlar
ekil9:HedefinizlediirotaiinUKF
kullanlarakyaplandeneylerdeeldeedilen
gzlemler,gerekrotavekestirilenrota
OrtalamaKareselHata(OKH)kestirimvaryans
merkezi limit teoreminin geerlilii varsaym
altnda deneysel hata dalm kullanlarak
hesaplanmtr:
VAR(OKH)=VAR(hata)/N
4.4Sonular
Model
GKFOrtalama
KareselHata
(OKH)
Radar
174.4(5.00)
genleme 185.2(3.15)
UKFOrtalama
KareselHata
(OKH)
116.9(0.363)
183.1(2.81)
Tablo1:Kestirimleriinortalamakareselhata
deerleri(1000simulasyonsonras
doygunluaulaanhatadeerleri).Doruluk
varyanslardaparanteziindeverilmitir.
5Sonu
Unscented
Kalman
filtresinin
(UKF)
doruluunu dorusal olmayan sre
modelleriiinaratranbirkamakale[16],[5],
olmasna karn, bunlarn hi biri zellikle
dorusal olmayan lm modellerinin
doruluunahitapetmemektedir.
Bu nedenle bu makalede dorusal olmayan
lmlere sahip dorusal durum uzay
modelleri iin UKF kestirim doruluk
performans GKFye gre greceli olarak
karlatrlmtr. Deneysel sonular radar ile
takip modeli iin UKF ve GKF performanslar
arasnda ciddi bir fark olduunu, buna karn
genleme metodu ile takip modeli iin
nemlibirfarkolmadngstermektedir.Bu
durum iki model arasndaki dorusallk
derecesi
farkndan
kaynaklanmaktadr;
genleme metodu ile takip modeli radar
modeline gre daha dorusal bir yapya
sahiptir. Bu yzden, dorusallk derecesi
olduka dk dorusal olmayan lm
modelleri iin UKF kullanmn GKFye gre
daha avantajldr. Bu karm [16] nolu
makalede sunulan UKF kullanm ile ilgili
argmanlarlartmektedir.
Kestirim doruluk dalm grafikleri , iki
kestiricinin her iki model iin de byk
hatalara sahip kestirimler hari benzer
sonular verdiini ortaya koymaktadr. Buna
gre UKFnin GKFye gre daha grbz 7
(salam)birkestiriciolduusylenebilir.
Kaynaka
[1]S.Arulampalam,S.Maskell,N.Gordon,andT.Clapp.
A tutorial on particle filters for online nonlinear/non
gaussian bayesian tracking. IEEE Transactions on Signal
Processing,50(2):174188,February2002.
[7]SimonJulierandJeffreyUhlmann.Ageneralmethod
for approximating nonlinear transformations of
probabilitydistributions.UniversityofOxford,November
1996.
[8]SimonJulierandJeffreyUhlmann.Anewextensionof
the kalman filter to nonlinear systems. Int. Symp.
Aerospace/Defense Sensing, Simul. And Controls,
Orlando,FL,1997.
[9]SimonJulierandJeffreyUhlmann.Unscentedfiltering
and nonlinear estimation. Proceedings of the IEEE,
March2004.
[13]RudolphvanderMerwe.Sigmapointkalmanfilters
forprobabilisticinferenceindynamicstatespacemodels.
WorkshoponAdvancesinMachineLearning, Montreal.,
June2003.
[15]RudolphvanderMerweandEricWan.Thesquare
root unscented kalman filter for state and parameter
estimation. Proceedings of the International Conference
onAcoustics,Speech,andSignalProcessing(ICASSP),Salt
LakeCity,Utah,May2001.
ing.:robust
Abstract
The Extended Kalman Filter (EKF) has long been the
de-facto standard for nonlinear state space estimation
[11], primarily due to its simplicity, robustness and
suitability for realtime implementations. However, an
alternative approach has emerged over the last few
years, namely the unscented Kalman filter (UKF). This
filter claims both higher accuracy and robustness for
nonlinear models. Several papers have investigated the
accuracy of UKF for nonlinear process models, but
none has addresses the accuracy for nonlinear measurement models in particular. This paper claims to
bridge this gap by comparing the performance of EKF
to UKF for two tracking models having nonlinear measurements.
Introduction
Notation
Algebraic letters are distinguished based on their appearance: Normal (a) denotes scalars, bold (a) denotes
vectors and uppercase (A) denotes matrices. Subscripts (xa ) denote discrete time. Conditional subscripts (xa|b ) denote state x in time a, given measurements up to time b. A hat superscript (a)
denotes an
estimated value, known only with a certain belief.
Background
2.1
vk
x k+1
uk
f(x,u,w)
xk
zk
z -1
h(x,v)
xk
2.2
A linear state-space model is a model where the functions f and h are linear in both state and input. The
functions can then be expressed by using the matrices
F, B and H, reducing state propagation calculations
to linear algebra. Overall this results in the following
state-space model:
xk+1 = Fk xk + Bk uk + wk
zk = Hk xk + vk
p(xk |zk1 ) =
w
vk
x k+1
uk
Bk
xk
z -1
zk
Hk
Fk
State Estimation
3.1
p(x k |z k-1)
predict
zk
update
p(x k |z k)
Figure 3: Recursive Bayesian estimator loop
Unfortunately, this method does not scale very well
in practice, mainly due to the large state space for
multidimensional state vectors. Calculating the prior
probability of each point in this state space involves
a multidimensional integral, which quickly becomes
intractable as the state space grows. Computers are
also limited to calculation of the pdf in discrete point
in state space, requiring a discretization of the state
space. This technique is therefore mainly considered
as a theoretic foundation for state estimation in general. Bayesian estimation by means of computers
is only possible if either the state space can be discretized, or if certain limitations apply for the model.
3.2
Kalman Filter
x k|k-1 P k|k-1
predict
zk
update
x
x k|k
P k|k
Figure 5: Illustration of how the Extended Kalman filter linearizes a nonlinear function around the mean of
a Gaussian distribution, and thereafter propagates the
The recursive Bayesian estimation technique is then mean and covariance through this linearized model
reduced to the Kalman filter, where f and h is replaced
by the matrices F, B and H. The Kalman filter is, just
as the Bayesian estimator, decomposed into two steps:
f (x, u, w)
predict and update. The actual calculations required
Fk =
x
are:
x k|k ,uk ,0
Predict next state, before measurements are taken:
h(x, v)
Hk =
x x k|k1 ,0
x k|k1 = Fk x k1|k1 + Bk uk
Figure 4: Kalman filter loop
3.4
3.4.1
The problem of propagating Gaussian random variables through a nonlinear function can also be approached using another technique, namely the unscented transform. Instead of linearizing the functions, this transform uses a set of points, and propagates them through the actual nonlinear function, eliminating linearization altogether. The points are chosed
such that their mean, covariance, and possibly also
higher order moments, match the Gaussian random
variable. Mean and covariance can be recalculated
from the propagated points, yielding more accurate results compared to ordinary function linearizaton. The
underlying idea is also to approximate the probability distribution instead of the function. This strategy
typically does both decrease the computational complexity, while at the same time increasing estimate accuracy, yielding faster, more accurate results.
3.4.2
Background
covariance as the Gaussian they approximated. Skewness was avoided by selecting the points in a symmetric way, such that any approximation error would only
originate from the fourth and higher moments.
Usage of the unscented transform in Kalman filtering was then presented by Julier in [8], where he introduced the Unscented Kalman filter (UKF), which
approximates the state estimate using sigma points.
Later, it was analyzed more in depth in [16].
A limitation associated with the unscented Kalman filter is that it has a lower bound on the safe spread of
the sigma points, meaning the distance between the
points in state space. Sigma point spreads below this
bond are not guaranteed to yield positive semidefinite
correlation matrices. This distance also increases with
the dimension of the state space, a limitation that may
cause problems in highly nonlinear models, since high
sigma point spread may result in sampling of non-local
features.
The technique presented here is therefore based on the
scaled unscented transform [6], which provides an additional tuning parmeter, , compared to the original
unscented transform. This parameter is used to arbitrary control the spread of the sigma-points, while at
the same time guaranteeing positive semidefinite covariance matrices. Even models of high dimensonality
can then keep a tight sigma point spread to avoid nonlocal effects.
3.4.3
Augmented state
Filter Formulation
y=f(x)
1
2
1
(c)
w 0 = 4 2 2
1
(c)
(m)
wi = wi = 2
2 L
(m)
w0 = 1
i=0
i=0
Figure 6: Illustration of how the unscented Kalman
filter propagates sigma-points from a Gaussian distrii = 1..., 2L
bution through a nonlinear function, and recreates a
Gaussian distribution, by calculating the mean and co- The filter then predicts next state by propagating the
variance of the results
sigma-points through the state and measurement models, and then calculating weighted averages and covarito a L2 matrix. Together this forms the augmented state ance matrices of the results:
estimate vector xa and covariance matrix Pa :
xk|k1 = f (xk1 , uk , wk1 )
2L
(m)
xk1
x
=
k|k1
wi xk|k1
xak1 = 0w
i=0
0v
2L
(c)
P
=
wi [xk|k1 x k|k1 ][xk|k1 x k|k1 ]T
a
a
a
a
a
T
k|k1
Pk1
0
0
Zk|k1 = h(xk|k1 , vk1 )
Qk1
0
= 0
2L
(m)
0
0
Rk1
z k|k1 = w Zi,k|k1
i
i=0
q
a ),
ai,k1 = xak1 + ( LPk1
i
q
a )
ai,k1 = xak1 ( LPk1
iL ,
i = 1..., L
i=0
The predictions are then updated with new measurements by first calculating the measurement covariance and state-measurement cross correlation matrices, which are then used to determine the Kalman gain:
2L
i = L + 1..., 2L
(c)
i=0
where subscript i means the i-th column of the square
1
2
K
=
P
xz Pzz
k
root of the covariance matrix . The parameter, in
the interval 0 < 1, determines sigma-point spread.
x k|k = x k|k1 + Kk (zk z k|k1 )
This parameter is typically quite low, often around
Pk|k = Pk|k1 Kk Pyy KkT
0.001, to avoid non-local effects.
The resulting ak1 matrix can now be decomposed Experimental results indicate [16] that Unscented
vertically into the xk1 rows, which contains the state; Kalman filters yield results comparable to a third orthe wk1 rows, which contains sampled process noise der Taylor series expansion of the state-model, while
2 The square root of a symmetric matrix is typically calculated Extended Kalman filters of course only are accurate to
by means of a lower triangular Cholesky decomposition. The a first order linearization. Consult [14] for a comparisquare root A of matrix P is then on the form P = AAT .
son of accuracy between the two kinds of filters.
The most computationally demanding part of the Un- The model px = vx , py = vy , vx = ax and vy = ay yields
scented Kalman filter is the matrix square-root used the following discrete time process model when asto calculate sigma points. Matrix diagonalization or suming zero order hold with timestep Ts = 1.
Cholesky factorization of the covariance matrix can be
3.5
Other Approaches
Experiments
where the accelerations ax and ay are modelled as uncorrelated white noise with a variance of 0.5.
The process is assumed to start in the following state:
x0 = [200 200 4 0]T
with the squared error metric (x x )T (x x ) for estimation accuracy.
4.2
Tracking by Radar
n2
atan(py /px )
where the measurement model is clearly nonlinear.
The radar is assumed to be positioned in the coordinates (0, 0) with measurement noises n1 and n1 , having
a variance of 200 and 0.003, respectively.
p
)
n
x
1x
y
1y
1
The basis for the experiment is an aeroplane, modelled
+
= p
d2
n2
(px p2x )2 + (py p2y )2
linearly as a dual Wiener process3 for position and velocity respectively, driven by white noise acceleration.
where the measurement model is also clearly nonlin3 Wiener processes are integrated white noise
ear. The observators are assumed to be positioned in
Model
Radar
Triangulation
d1
EKF MSE
174.4 (5.00)
185.2 (3.15)
UKF MSE
116.9 (0.363)
183.1 (2.81)
Table 1: Mean squared error (MSE) for the estimations, saturating errors above 1000. Accuracy variance is given in parenthesis.
d2
ekf
ukf
500
the coordinates (300, 0) and (300, 0), with measurement noises n1 and n2 , both having a variance of 200
.
This configuration also poses an ambiguity problem,
since measurements coming from a given position will
be equal to the same position flipped about the x-axis.
This will probably not be a big problem, since the
process model prediction will help resolving this ambiguity.
0
1
10
10
10
0
1
10
10
10
observations
true trajectory
estimated trajectory
300
250
200
150
200
150
100
50
50
100
Conclusion
sistent with the arguments for using the UKF presented [9] Simon Julier and Jeffrey Uhlmann. Unscented
filtering and nonlinear estimation. Proceedings
in [16].
of the IEEE, March 2004.
The estimation error distribution plots show that the
two estimators yield quite similar results for both
models, with the most significant exception being the [10] Simon Julier, Jeffrey Uhlmann, and Hugh
Durrant-Whyte. A new approach for filtering
amount of estimates having severely large errors. This
nonlinear systems. Proceedings of the 1995
leads us to the conclusion of UKF being a more robust
American Control Conference, IEEE Press, June
estimator than EKF.
1995.
Statement of reproducibility: All program code required to reproduce the results shown in this paper are [11] Ben Quine, Jeffrey Uhlmann, and Hugh Durrantfreely available upon request by contacting the author.
Whyte. Implicit jacobians for linearised state estimation in nonlinear systems. Proceedings of the
1995 American Control Conference, IEEE Press,
References
June 1995.
[1] S. Arulampalam, S. Maskell, N. Gordon, and
[12] Charles W. Therrien. Discrete Random Signals
T. Clapp. A tutorial on particle filters for on-line
and Statistical Signal Processing. Prentice Hall,
non-linear/non-gaussian bayesian tracking. IEEE
1992.
Transactions on Signal Processing, 50(2):174
188, February 2002.
[13] Rudolph van der Merwe. Sigma-point kalman fil[2] Robert Grover Brown and Patrick Y. C. Hwang.
Introduction to Random Signals and Applied
Kalman Filtering, 3rd Edition. Prentice Hall,
1996.
[14]
ters for probabilistic inference in dynamic statespace models. Workshop on Advances in Machine Learning, Montreal., June 2003.
Rudolph van der Merwe and Eric Wan. Sigmapoint kalman filters for integrated navigation.