DorusalOlmayanlmlDurumUzayModelleriiinKalman

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

Cebirsel gsterimler grnlerine gre

farkllk gstermektedir. Normal gsterimler
(a) skalar byklkleri, kaln gsterimler (a)
vektrlerivebykharfli gsterimler(A)ise
matrisleri simgelemektedir. Alt simgeler (Xa)
kesikli zaman simgeler. Koullu alt simgeler
(Xa|b ),verilmiolanbannakadarkalanolan
lmlerileaanndakixdurumunusimgeler.^
st simgesi ( a ) kestirilen (tahmini) deeri
simgeler.

1Giri

2GenelBilgi

Durum kestirimi problemi, sadece grltl

ve/veya doru olmayan lmlerine
eriilebilen bir srecin durum kestirimi ile
ilgilenir. Bu problem ortamna fen ve
mhendislik dallarndaki hemen hemen her
disiplindeouzamanrastlanr.

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.

En yaygn kullanlan durum kestiricisi Kalman

filtresidir. Kalman filtresi dorusal sistemler
iin optimal bir kestiricidir, ancak gerek
dnyadaokazsaydadorusalsistemvardr.
Bu problemin stesinden gelmenin yaygn bir
yaklam,Kalmanfiltresinikullanmadannce
sistemi dorusal hale getirmektir ve bu

2.1GenelDurumUzayModeli
Durum uzay modellerinin en genel biimi
dorusal olmayan modeldir. Bu model tipik
olarakikifonksiyonu(fveh)ierir.

f ve h fonksiyonlar srasyla durumu ve

lmleri
(gzlemleri)
belirleyen
fonksiyonlardr. u sre girdisini, w ve v
srasyla durum ve lm (gzlem) grlt
vektrlerinivekisekesiklizamanifadeeder.

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.

Dorusal durum modelleri ya gerek dorusal

sreler, ya da dorusal olmayan srelerin
basitbirekildebirincidereceTayloryaklam
aracl ile dorusal hale getirilmi biimleri
zerinekurulmumodellerdir.

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

ndirgemeli Bayes kestirimi, sreci simule

Bu metotun pratik uygulamas ounlukla ok
etmeveeldeedilenbusrecigereksreten
boyutlu durum vektrleri iin byk durum
elde edilen yeni lmleri (z) hesaba katarak
uzaylar sebebiyle olduka gtr. Bu tarz bir
simulasyon sreci ile ayn anda dzeltme
durum uzaynda her bir noktada bir nceki
prensibiilealr.Hesaplamalarikiaamalbir
olasln hesaplanmas, durum uzay
ilem ile indirgemeli olarak yaplr. lk olarak
bydke zm gittike zorlaan ok
bir sonraki durum, f fonksiyonundan
boyutlu
bir
integral ierir. Bilgisayarlar da
elde edilen durum ilerleme kans 2 p (xk|xk1)
durum uzaynn kesikli zamanda olaslk
kullanlarak mevcut durumun bir sonraki
younluk fonksiyonu hesaplamada snrl
durum zerine d deerlemesi ile tahmin
kalmaktadr. Ayrca bu hesaplamann
edilir. kinci aamada, yeni lmler gz
yaplabilmesiiindurumuzaynnnkesiklihale
nnde bulundurularak h fonksiyonundan
getirilmesi gerekmektedir. Bu yzden genelde
elde edilen lm olabilirlii 3 p (zk|xk)
bu yntem durum kestirimin teorik temeli
kullanlarak
yaplan
bu
tahmin
olarakkabuledilir.BilgisayarvastasylaBayes
dzeltilir/gncellenir.
kestirimi ya durum uzaynn kesikli hale
getirilebildiiyadamodelebellikstlamalarn
lk olarak, xk durumunun bir nceki zaman
uygulanddurumlardammknolmaktadr.
dilimineaitolaslkyounlukfonksiyonunuk1
anna kadar olan lmler baz alnarak
3.2KalmanFiltresi
hesaplamak iin ChapmanKolmogorov eitlii
Durum kestirim problemi, sre modeli
kullanlr:
zerinde belirli zorlama ve kstlamalar
uygulandnda kolay ilenir hale gelir. Bu
zorlamavekstlamalar;fvehfonksiyonlarnn

dorusal olmas, ve grlt terimleri w ve

vnin birbirleriyle ilikisiz, Gauss dalml, sfr
Daha sonra, k anna kadar olan lmler
ortalamal beyaz grlt olmalardr. Yukarda
alndktan sonra bu lmler dikkate alnarak
ifadeedilenlerimatematikselolarakuekilde
xkdurumunungncellenmiolaslkyounluk
gsterebiliriz:
fonksiyonunu elde etmek iin Bayes kural
kullanlr:

Burada Q ve R srasyla durumun ve lm

grltsnn ikinci derece zelliklerini
belirten kovaryans matrisleridir. Yukarda
tanmlananlar durum modelini aadaki
biimeindirger:

ekil3:ndirgemeliBayesKestiricisidngs

BuradaF,BveHzamanabalmatrislerdir.

2
3

ing.:StatePropagationBelief
ing.:MeasurementLikelihood

Dorusal bir modele Gauss dalml bir girdi

verildiindemodelindurumvektsdaGauss
dalml olur [12]. Bu yzden, durum ve kt
olaslk younluk fonksiyonu daima ortalama

ve kovaryansn yeterli istatistikte olduu

normaldalmasahiptir.Budagsterirki,bu
durumda btn durum olaslk younluk
fonksiyonun hesaplamaya gerek yoktur. x
ortalamavektrnvedurumiinPkovaryans
matrisinihesaplamakyeterliolacaktr.

ekil4:Kalmanfiltresidngs

filtresinin ikinci derecen hata metriine sahip

sre durumlar iin optimal bir kestirici
olduukantlanabilir[2].
3.3GeniletilmiKalmanFiltresi
Gerek dnyadaki ou sre maalesef
dorusal deildir; dolaysyla Kalman filtresi
vastas ile tahmin edilebilmeleri iin dorusal
hale getirilmelidir. Geniletilmi Kalman
Filtresi (GKF) bu problemi, f ve h
fonksiyonlarnn tahmini durum etrafnda
Jacobian matrisini hesaplayarak zer. Bu
matris durumun etrafna merkezlenen model
fonksiyonunun erisini verir. Jacobian matrisi
bir vektrn btn ksmi trevlerini
iermektedir.

ndirgemeli Bayes kestirimi yntemi F, B ve H

matrislerinin f ve h fonksiyonlarnn yerini
ald Kalman filtresine dnr. Kalman
filtresiikiaamadanoluanBayeskestiricisidir.
Bu iki aama, tahmin etme ve gncelleme
aamalardr. Gerekli hesaplamalar aada
srasylagsterilmitir:

lmler alnmadan nce bir sonraki durumu

tahminetme:

lmler
alndktan
gncelleme:

sonra

durumu

ekil 5: GKFnin dorusal olmayan bir

fonksiyonu Gauss dalmnn ortalamas
etrafnda dorusal hale getirmesi ve sonra
dorusal hale getirilmi bu model vastas ile
ortalamavekovaryansyaymas.

Burada K Kalman kazanc matrisi ve P ise

kestirimin hassasiyeti ile ilgili bilgiyi iinde
barndran durum kestirim kovaryans
matrisidir. Bu filtre ile daha fazla bilgi ve
detaylar[2]nolumakaledeneldeedilebilir.

Kalman filtresi ounlukla dorusal bir yapya

sahip olduu iin matris ters alma ilemleri
harikolayhesaplamalarierir.AyrcaKalman

Geniletilmi Kalman filtresi tahmini duruma

dayal olarak zamanla deiiklik gsteren
(zamana bal) F ve H dnda normal bir
Kalman filtresi gibi alr. Gerekli
hesaplamalaraadasrasylagsterilmitir:

lmler alnmadan nce bir sonraki durumu

tahminetme:

lmler
alndktan
gncelleme:

sonra

durumu

Burada K Kalman kazanc matrisi ve P ise

kestirimin hassasiyeti ile ilgili bilgiyi iinde
barndran durum kestirim kovaryans
matrisidir.

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

Wyte tarafndan [10] ve [11] nolu

makalelerde ortaya konmutur. Julier,
Uhlmann ve DurrantWyte bu metotta N
boyutlu rastgele bir Gauss deikenin sigma
noktalar olarak adlandrlnan 2N+1 adet
rnek ile nasl ifade edilebildiinin taslak
olarak ortaya koymulardr. Yaklam
saladklar Gauss dalm ile ayn kovaryansa
olacak ekilde noktay seebilmek iin
karekk matrisi ve kovaryans tanmlanndan
faydalanmlardr. Bu noktalar simetrik olacak
ekildeseilerekarpklknlenir.Buyntem
ile yaklatrm hatas dk olur ve ancak
drdnc ve daha yksek momentlerden
kaynaklanr.
Unscented dnmn Kalman filtrelemede
kullanm daha sonra Julier ve Uhlmann
tarafndan Unscented Kalman Filtresi
formunda [8] nolu makalede ortaya
koyulmutur. Unscented Kalman Filtresi bir
sonraki durumu sigma noktalar kullanarak
tahmin etmektedir. Bu filtre, [16] nolu
makalededahadetaylolarakanalizedilmitir.
UnscentedKalmanFiltresiileilgilikstlama;bu
filtrenindurumuzayndakinoktalararasndaki
uzakl ifade eden sigma noktalarnn gvenli
yaylmnn daha dk bir limite sahip
olmasdr. Sigma noktas yaylmlarnn bu
limitin altnda olmas pozitif yartanml
korelasyonmatrislereldeetmeyigarantialtna
almaz. Ayrca bu uzaklk durum uzaynn
boyutuna bal olarak artar ve bu ekilde
ortaya kan yksek sigma yaylm lokal
olmayan zelliklerin rneklemesine yol
aabileceinden yksek derecede dorusal
olmayan modellerde eitli problemler
meydanagetirebilir.
Bu yzden burada ortaya konan yntem, asl
unscented dnmden farkl olarak ek bir
ayarlama parametresi ,, ieren lekli
unscented dnm [6] zerine yaplanmtr.
Bu parametre, sigma noktalarnn yaylmn
pozitif yartanml korelasyon matrisler
salamaygarantialtnaalarakkontroletmede
kullanlr. Bu sayede yksek boyutlu modeller
iinbiledarbirsigmanoktasyaylmilelokal
olmayanetkilerdenkanlabilir.

ekil 6: UKFnin Gauss dalml sigma

noktalarnn dorusal olmayan bir fonksiyon
zerinde ilerletilmesi,ve sonularn ortalama
ve kovaryanslarn hesaplayarak yeniden bir
Gaussdalmoluturulmas

Bir sonraki aama geniletilen (ekli) durumun

ortalama ve kovaryansn yakalayan 2L+1
sigmanoktasnnoluturulmasniermektedir.
a matrisi bu noktalar ierecek ekilde seilir
vesutnlaruekildehesaplanr:

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:

Simetrik bir matrisin karekk tipik olarak dk

olangenCholeskydalmvastasilehesaplanr.
P matrisinin karekk matrisi A ise P=AAT
formundadr.

Daha sonra filtre, sigma noktalarn durum ve

lm (gzlem) modelleri zerinde ilerleterek
ve elde edilen sonularn arlkl ortalama ve
kovaryans matrislerini hesaplayarak bir
sonrakidurumutahmineder:

problemin zmnde kullanlabilir ; fakat,

sadece kovaryans matrisi karekklerinin
yaylmn kapsayan dorudan karekk
yaklam hesaplama asndan daha etkin bir
zm sunar. [15] nolu makalede Merwe bu
ekildebiryaklamortayakoymaktadr.

3.5DierYaklamlar

Tahminler gelen yeni lmlerle gncellenir.

Gncelleme ileminde ilk olarak lm
kovaryans ve durumlm apraz korelasyon
matrisleri hesaplanr daha sonra bu matrisler
Kalmankazancnbelirlemedekullanlr:

Deneysel sonular Unscented Kalman

filtrelerinindurummodelininncderecede
Taylorserisialmnayaknsonularverdiini
;bunakarn,geniletilmiKalmanfiltrelerinin
sadece birinci derece dorusallatrma
seviyesinde bir dorulua sahip olduunu
gstermektedir[16]. [14] nolu makalede bu
iki tip filtrenin doruluk performanslar
karlatrlmtr.
Unscented Kalman filtresinde hesaplama
asndanenokdikkatgerektirenksmsigma
noktalarnn hesaplamasnda kullanlan matris
karekk alma ksmdr. Kovaryans matrisinin
Cholesky arpanlarna ayrlmas ya da matris
kegenletirme ileminin yaplmas bu

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

Bu modelde radarn (0,0) koornidat

noktasnda konuland ve n1, n2 lm
grltleri varyanslarnn srasyla 200 ve
0.003olduuvarsaylmtr.

Dorusal olarak modellenmi ve beyaz

grltl ivmeye sahip bir uak bu ksmda
elealnacakdeneyintemelinioluturmaktadr.
Dorusal model, konum ve hz iin ikili
Wiener 5 srecibiimindeoluturulmutur.

p& x = v x , p& y = v y , v x = a x ve v& y = a x modeli

aadaki Ts=1 zaman aralkl kesikli zaman
sremodeliniortayakarr:

ekil7:Radarlauaktakibi

4.3genlemeMetoduileTakip 6

Buradaki ax ve ay ivmeleri 0.5 varyanslilkisiz

beyazgrltlolarakmodellenmitir.
Srecin, aadaki
varsaylmtr:

durumda

balad

genleme ya da nirengi metodu ile takip; iki

gzlemevitarafndanhedefinkendilerineolan
uzaklklar (d1 ve d2) gzlemleyen dorusal
olmayanbirlmmodeliilemodellenebilir:

Kestirim doruluu iin karesel hata metrii

ise;

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

Bu tarz bir yaplan, belirli bir konumdan

gelen lmlerin tamamnn her seferinde
x ekseni zerinde ayn konuma karlk
geleceinden bir belirsizlik problemi ortaya
karr. Fakat sre model tahmininin bu
belirsizlikprobleminizmeyeyardmcolmas
yksek ihtimalle bu durumu byk bir sorun
olmaktankaracaktr.

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

kili simulasyon/kestirim deneyi 10000 kere

gerekletirilmitir. Her seferinde simule
edilenuanizlediirota80zamanadmnda
hem GKF hem de UKF gzlem modelleri
kullanlarak kestirilmitir. Kestirim doruluu
sonular Tablo 1de , doruluk dalm
grafikleriiseekil10dagsterilmitir.

Yksek derecede dorusal olmayan arkus

tanjant ieren lmlere sahip radar modeli
sadecePythagoraslmleresahipgenleme
modeline gre GKF ve UKF kestirim
doruluklar arasnda daha byk farkllk
gstermektedir. Bu durum, her iki modelde
de hatann 1000in zerinde olduu birka
kestirim iin UKFnin daha grbz bir yap
sergilediinigsterir.

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.

[2] Robert Grover Brown and Patrick Y. C. Hwang.

Introduction to Random Signals and Applied Kalman
Filtering,3rdEdition.PrenticeHall,1996.

[3] ChiTsong Chen. Linear System Theory and Design,

thirdedition.OxfordUniversityPress,1999.

[4] K. Ito and K. Xiong. Gaussian filters for nonlinear

filtering problems. Kazufumi Ito and Kaiqi Xiong,
Gaussian Filters for Nonlinear Filtering Problems, IEEE
Transactions on Automatic Control, vol. 45, no. 5, pp.
910927,May2000.

[5] Joseph J.and LaViola Jr. A comparison of unscented

andextendedkalmanfilteringforestimatingquaternion
motion. Proceedings of the 2004 American Control
Conference,IEEEPress,21902195,June2004.

[6] Simon Julier. The scaled unscented transform.

Proceedings of the 2002 American Control
Conference,IEEEPress,May2002.

[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.

[10] Simon Julier, Jeffrey Uhlmann, and Hugh Durrant

Whyte. A new approach for filtering nonlinear systems.
Proceedings of the 1995 American Control Conference,
IEEEPress,June1995.

[11] Ben Quine, Jeffrey Uhlmann, and Hugh Durrant

Whyte.Implicit jacobiansforlinearisedstateestimation
innonlinearsystems.Proceedingsofthe1995American
ControlConference,IEEEPress,June1995.

[12] Charles W. Therrien. Discrete Random Signals and

StatisticalSignalProcessing.PrenticeHall,1992.

[13]RudolphvanderMerwe.Sigmapointkalmanfilters
forprobabilisticinferenceindynamicstatespacemodels.
WorkshoponAdvancesinMachineLearning, Montreal.,
June2003.

[14] Rudolph van der Merwe and Eric Wan. Sigmapoint

kalmanfiltersforintegratednavigation.

[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.

[16] Eric Wan and Rudolph van der Merwe. The

unscentedkalmanfilterfornonlinearestimation.Proc.of
IEEE Symposium 2000 (ASSPCC), Lake Louise, Alberta,
Canada,October2000.

ing.:robust

Comparison of Kalman Filter Estimation Approaches for

State Space Models with Nonlinear Measurements
Fredrik Orderud
Sem Slands vei 7-9, NO-7491 Trondheim

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

The problem of state estimation concerns the task of

estimating the state of a process while only having
access to noisy and/or inaccurate measurements from
that process. It is a very ubiquitous problem setting,
encountered in almost every discipline within science
and engineering.
The most commonly used type of state estimator is the
Kalman filter. It is an optimal estimator for linear systems, but unfortunately very few systems in real world
are linear. A common approach to overcome this problem is to linearize the system first before using the
Kalman filter, resulting in an extended Kalman filter.
This linearization does however pose some problems,
e.g. it can result in nonstable estimates [8]. The development of better estimator algorithms for nonlinear systems has therefore attracted a great deal of interest in the scientific community, because improvements here will undoubtedly have great impact in a
wide range of engineering fields.

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

A state space model is a mathematical model of a

process, where the process state x is represented by
a numerical vector. State-space models actually consists of two separate models: the process model, which
describes how the state propagates in time based on external influences, such as input and noise; and the measurement model, which describe how measurements z
are taken from the process, typically simulating noisy
and/or inaccurate measurements.

2.1

General State Space Model

The most general form of state-space models is the

nonlinear model.
This model does typically consist of two functions, f
and h:
xk+1 = f (xk , uk , wk )
zk = h(xk , vk )
which govern state propagation and measurements, respectively. u is process input, and w and v are state and
measurement noise vectors, respectively while k is the
discrete time.
State-space models are remarkably usable for modelling almost all sorts of processes. f and h are usually
based upon a set of discretized differential equations,
governing the dynamics of and observations from the
process.

vk

x k+1

uk
f(x,u,w)

xk

zk

z -1

h(x,v)

xk

Figure 1: A general state-space model. z1 is the unit

delay function known from the Z-transform in digital
signal processing.

2.2

Linear State Space Model

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

timal way of predicting a state pdf for any process,

given a system and a measurement model. In this section we will discuss this estimator, which recursively
calculates a new estimate for each time-step, based on
the estimate for the previous timestep and new measurements.
Recursive Bayesian estimation works by simulating
the process, while at the same time adjusting it to account for new measurements z, taken from the real
process. The calculations are performed recursively
in a two step procedure. First, the next state is predicted, by extrapolating the current state onto next
time step using state propagation belief p(xk |xk1 ) obtained from function f . Secondly, this prediction is
corrected using measurement likelihood p(zk |xk ) obtained from function h, taking new measurements into
account.
The Chapman-Kolmogorov equation is first used to
calculate a prior pdf for state xk , based on measurements up to time k 1:
Z

zk = Hk xk + vk
p(xk |zk1 ) =
w

vk

x k+1

uk
Bk

xk
z -1

zk
Hk

Fk

p(xk |xk1 )p(xk1 |zk1 ) dxk1

Bayes rule is then used to calculate the updated pdf

for state xk , after taking measurements up to time k
into account:
p(xk |zk ) = R

p(zk |xk )p(xk |zk1 )

p(zk |xk )p(xk |zk1 ) dxk

Figure 2: A linear state-space model

This linear model is easier both to calculate and analyze. Enabling modellers to investigate properties
such as controllability, observability and frequency response [3].
Linear state models are either based on inherently linear processes, or simply a linearized versions of a nonlinear process by means of a first order Taylor approximation.

State Estimation

State estimation concerns the problem of estimating

the probability density function (pdf ) for the state of
a process which is not directly observable. This involves both predicting the next state based on the current, and updating/correcting this prediction based on
noisy measurements taken.

3.1

Recursive Bayesian Estimation

The most general form for state estimation is known

as recursive Bayesian estimation [1]. This is the op-

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

Update state, after measurements are taken:

The problem of state estimation can be made tractable

Kk = Pk|k1 HkT (Hk Pk|k1 HkT + Rk )1
if we put certain constrains on the process model, by
x k|k = x k|k1 + Kk (zk Hk x k|k1 )
requiring both f and h to be linear functions, and the
Pk|k = (I Kk Hk )Pk|k1
noise terms w and v to be uncorrelated, Gaussian and
white with zero mean. Put in mathematical notation,
where K is the Kalman gain matrix, used in the upwe then have the following constraints:
date observer, and P is the covariance matrix for the
f (xk , uk , wk ) = Fk xk + Bk uk + wk
state estimate, containing information about the accuracy of the estimate. More details and background for
h(xk , vk ) = Hxk + vk
this filter can be found in [2].
wk N(0, Qk ) vk N(0, Rk )
The Kalman filter is quite easy to calculate, due to the
E(wi wTj ) = Qi (i j) E(vi vTj ) = Ri (i j)
fact that it is mostly linear, except for a matrix inverE(wk vTk ) = 0
sion. It can also be proved that the Kalman filter is an
optimal estimator of process state, given a quadratic
where Q and R are covariance matrices, describing the
error metric [2].
second-order properties of the state- and measurement
noise. The constraints described above reduces the
3.3 Extended Kalman Filter
state model to:
Most processes in real life are unfortunately not linear,
and therefore needs to be linearized before they can be
zk = Hk xk + vk
estimated by means of a Kalman filter. The extended
where F, B and H are matrices, possible time depen- Kalman filter (EKF) solves this problem by calculating
the Jacobian1 of f and h around the estimated state,
dent.
As the model is linear and input is Gaussian, we know which in turn yields a trajectory of the model function
that the state and output will also be Gaussian [12]. centered around this state.
The state and output pdf will therefore always be nory
mally distributed, where mean and covariance are sufy=f(x)
ficient statistics. This implies that it is not necessary
to calculate a full state pdf any more, a mean vector x
and covariance matrix P for the state will suffice.
xk+1 = Fk xk + Bk uk + wk

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

Pk|k1 = Fk Pk1|k1 FkT + Qk

1 The Jacobian is the matrix of all partial derivatives of a vector

The extended Kalman filter works almost like a regular

Kalman filter, except for F and H, which vary in time
based on the estimated state x . The actual calculations
required are:
Predict next state, before measurements are taken:
x k|k1 = f (xk1|k1 , uk , 0)
Pk|k1 = Fk Pk1|k1 FkT + Qk
Update state, after measurements are taken:
Kk = Pk|k1 HkT (Hk Pk|k1 HkT + Rk )1
x k|k = x k|k1 + Kk (zk h(xk|k1 , 0))
Pk|k = (I Kk Hk )Pk|k1
where K is the Kalman gain matrix, used in the update
observer, and P is the covariance matrix for the state
estimate, containing information about the accuracy of
the estimate.

3.4
3.4.1

Unscented Kalman Filter

Introduction

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

The underlying method of unscented transform was

first proposed by Uhlmann et al. in [11] and [10],
where they laid out the framework for representing
a Gaussian random variable in N dimensions using
2N + 1 samples, called sigma points. They utilized
the matrix square root and covariance definitions to select these points in such a way that they had the same

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

The unscented transform approach also has another

advantage, namely that noise can be treated in a nonlinear fashion to account for non-Gaussian or nonadditive noises. The strategy for doing so involves
propagation of noise through the functions by first augmenting the state vector to also include noise sources,
a technique first introduced by Julier in [7], and later
refined more in depth by Merwe in [16]. Sigma point
samples are then selected from the augmented state,
xa , which also includes noise values. The net result
is that any nonlinear effects of process and measurement noise are captured with the same accuracy as the
rest of the state, which in turn increases accuracy for
non-additive noise sources.
3.4.4

Filter Formulation

The filter starts by augmenting the state vector to L

dimensions, where L is the sum of dimensions in the
original state-vector, model noise and measurement
noise. The covariance matrix is similarly augmented

and the vk1 rows, which contains sampled measurement noise.

Each sigma-point is also assigned a weight. These
weight are derived by comparing the moments of the
sigma-points with a Taylor series expansion of the
models while assuming a Gaussian distribution, as derived in [9]. The resulting weights for mean (m) and
covariance (c) estimates then becomes:

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 = E{(xk1 x k1 )(xk1 x k1 ) }

i=0

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

The next step consists of creating 2L + 1 sigma-points

in such a way that they together captures the full mean
and covariance of the augmented state. The a matrix
is chosen to contain these points, and its columns are
calculated as follows:
a0,k1 = xak1

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

Pzz = wi [Zi,k|k1 z k|k1 ][Zi,k|k1 z k|k1 ]T

i=0
2L

(c)

Pxz = wi [xi,k|k1 x k|k1 ][Zi,k|k1 z k|k1 ]T

(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

used to solve this problem, but a more direct square

0
px
1 0 1 0
px
root approach, propagating only the square-roots of
0 1 0 1 py 0
py

the covariance matrices, offers higher computationally

=
0 0 1 0 vx + ax
vx
efficiency. Merwe et al. proposes a approach for doing
ay k
vy k
0 0 0 1
vy k+1
this in [15].

3.5

Other Approaches

Central difference Kalman filter (CDKF) proposed in

[4] offers an alternative Kalman formulation for propagating Gaussian pdf s through nonlinear functions.
This formulation, although it is different, remains very
much like UKF and is reported to perform indistinguishable from UKF in all tests performed [13]. The
CDKF is therefore not addressed by this paper.
Particle filters [1]: While the various types of Kalman
filters often offers superb estimation accuracy, there
are situations where they are not suited for the task.
This problem relates to the fact that all Kalman filters are constrained to only model Gaussian probabilities, and are therefore incapable of handling skewed
or multimodal distributions. A more general approach is therefore needed when trying to estimate
non-Gaussian distributions. Particle filters, which are
based on sequential Monte-Carlo simulations using
importance sampling, can then often be a good alternative. Particle filters are not addressed here, since
this paper focuses on estimating only the most probable state of a process, and not the entire process pdf,
thus rendering particle filters superfluous. Particle filters are also much more computationally demanding
than Kalman filters, often making them intractable for
usage in real-time settings.

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

Radar tracking can be modelled with a measurement

model observing distance and angle to the target:
# 

 " q

d
n1
p2x + p2y
=
+

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.

Figure 7: Tracking of plane motion by means of a

radar

The experiments described here aims at determining

whether there are any difference between EKF and 4.3 Tracking by Triangulation
UKF for practical tracking applications, having linear
Tracking by triangulation can similarly be modelled
process models and nonlinear measurement models.
with a measurement model observing distances to the
target from two observators:
4.1 Process Model

  p
 

2 + (p p )2
d1
(p

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

Error distribution for radar tracking

1000

Figure 8: Tracking of plane motion by means of triangulation

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

Error distribution for triangulation tracking

1000
ekf
ukf
500

0
1
10

10

10

Figure 10: Estimation accuracy distributions

Observed, true and estimated trajectory
350

assuming validity of the central limit theorem.

The radar model, having measurements involving the
highly nonlinear arcus-tangent, shows a wider difference in the estimation accuracy between EKF and
UKF, compared to the triangulation model which only
has Pythagoras measurements, being significantly
more linear. It can further be seen that UKF seems
to show a higher degree of robustness, having fewer
estimates with errors above 1000 for both of the models.

observations
true trajectory
estimated trajectory

300

250

200

150
200

150

100

50

50

100

Conclusion

Figure 9: Example illustration of observations, true

trajectory and estimated trajectory for the experiments Several papers have investigated the accuracy of UKF
using the UKF
for nonlinear process models [16], [5], but none has
addresses the accuracy for nonlinear measurement
models in particular.
4.4 Results
This paper did therefore compare the relative estimation accuracy of UKF compared to EKF for linear state
A dual simulation/estimation experiment was run
space models with nonlinear measurements. The em10000 times. Each time, a simulated plane trajectory
pirical results shows a significant difference for the
were estimated over 80 time steps, by both EKF and
radar model, but not for the tracking model. This is
UKF for both of the observation models. The estimabelieved to be caused by the difference in nonlineartion accuracy results are shown in the table below, with
ity between the two models, having a highly nonlinaccuracy distribution plots in the accompanying figure.
ear radar model and a relatively more linear tracking
The MSE estimate variance is calculated from the emmodel. The relative advantage of using UKF does
pirical error distribution, using
therefore seem to increase with the degree of nonlinearity in the measurement model. This finding is conVAR(MSE) = VAR(error)/N

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.

[3] Chi-Tsong Chen. Linear System Theory and

Design, third edition. Oxford University Press, [15] Rudolph van der Merwe and Eric Wan. The
1999.
square-root unscented kalman filter for state and
parameter-estimation. Proceedings of the In[4] K. Ito and K. Xiong. Gaussian filters for nonlinternational Conference on Acoustics, Speech,
ear filtering problems. Kazufumi Ito and Kaiqi
and Signal Processing (ICASSP), Salt Lake City,
Xiong, Gaussian Filters for Nonlinear Filtering
Utah, May 2001.
Problems, IEEE Transactions on Automatic Control, vol. 45, no. 5, pp. 910927, May 2000.

[16] Eric Wan and Rudolph van der Merwe. The

unscented kalman filter for nonlinear estima[5] Joseph J. and LaViola Jr. A comparison of untion. Proc. of IEEE Symposium 2000 (AS-SPCC),
scented and extended kalman filtering for estiLake Louise, Alberta, Canada, October 2000.
mating quaternion motion. Proceedings of the
2004 American Control Conference, IEEE Press,
2190-2195, June 2004.
[6] Simon Julier. The scaled unscented transform.
Proceedings of the 2002 American Control Conference, IEEE Press, May 2002.
[7] Simon Julier and Jeffrey Uhlmann. A general
method for approximating nonlinear transformations of probability distributions. University of
Oxford, November 1996.
[8] Simon Julier and Jeffrey Uhlmann. A new extension of the kalman filter to nonlinear systems. Int.
Symp. Aerospace/Defense Sensing, Simul. and
Controls, Orlando, FL, 1997.