1.04 1.02 0.98 0.96 0.94 0.92 0.9



Similar documents
Whole-body dynamic motion planning with centroidal dynamics and full kinematics

Understanding and Applying Kalman Filtering

Introduction: Overview of Kernel Methods

Data Sensor Fusion for Autonomous Robotics

Least Squares Estimation

CORRELATION ANALYSIS

POLYNOMIAL AND MULTIPLE REGRESSION. Polynomial regression used to fit nonlinear (e.g. curvilinear) data into a least squares linear regression model.

Adaptive Demand-Forecasting Approach based on Principal Components Time-series an application of data-mining technique to detection of market movement

How To Build A Hybrid Fusion Module

United Arab Emirates University College of Sciences Department of Mathematical Sciences HOMEWORK 1 SOLUTION. Section 10.1 Vectors in the Plane

Vectors. Objectives. Assessment. Assessment. Equations. Physics terms 5/15/14. State the definition and give examples of vector and scalar variables.

Data Mining Part 5. Prediction

MAT 242 Test 3 SOLUTIONS, FORM A

Automated Stellar Classification for Large Surveys with EKF and RBF Neural Networks

Data Mining and Data Warehousing. Henryk Maciejewski. Data Mining Predictive modelling: regression

COMPUTATION OF THREE-DIMENSIONAL ELECTRIC FIELD PROBLEMS BY A BOUNDARY INTEGRAL METHOD AND ITS APPLICATION TO INSULATION DESIGN

A repeated measures concordance correlation coefficient

LIST OF TABLES. 4.3 The frequency distribution of employee s opinion about training functions emphasizes the development of managerial competencies

TIME SERIES ANALYSIS & FORECASTING

Example: Credit card default, we may be more interested in predicting the probabilty of a default than classifying individuals as default or not.

Regression III: Advanced Methods

Dimensionality Reduction: Principal Components Analysis

8. Linear least-squares

Phys222 Winter 2012 Quiz 4 Chapters Name

A multi-scale approach to InSAR time series analysis

System Identification for Acoustic Comms.:

{(i,j) 1 < i,j < n} pairs, X and X i, such that X and X i differ. exclusive-or sums. ( ) ( i ) V = f x f x

An Online Estimation of Rotational Velocity of Flying Ball via Aerodynamics

Numerical Analysis. Gordon K. Smyth in. Encyclopedia of Biostatistics (ISBN ) Edited by. Peter Armitage and Theodore Colton

4.3 Least Squares Approximations

Manifold Learning Examples PCA, LLE and ISOMAP

Solving DEs by Separation of Variables.

Chapter 13 Introduction to Nonlinear Regression( 非 線 性 迴 歸 )

Logistic Regression. Jia Li. Department of Statistics The Pennsylvania State University. Logistic Regression

SOLUTIONS. Would you like to know the solutions of some of the exercises?? Here you are

Week 5: Multiple Linear Regression

Sections 2.11 and 5.8

Linearly Independent Sets and Linearly Dependent Sets

Physics Kinematics Model

Mean = (sum of the values / the number of the value) if probabilities are equal

3.2 Sources, Sinks, Saddles, and Spirals

Big Data: The curse of dimensionality and variable selection in identification for a high dimensional nonlinear non-parametric system

Computer exercise 2: Least Mean Square (LMS)

Learning Outcomes. Distinguish between Distance and Displacement when comparing positions. Distinguish between Scalar and Vector Quantities

Automatic Detection of Emergency Vehicles for Hearing Impaired Drivers

Smoothing. Fitting without a parametrization

Manufacturing Equipment Modeling

Concepts in Investments Risks and Returns (Relevant to PBE Paper II Management Accounting and Finance)

Assessment of Numerical Methods for DNS of Shockwave/Turbulent Boundary Layer Interaction

Principle Component Analysis and Partial Least Squares: Two Dimension Reduction Techniques for Regression

Community Mining from Multi-relational Networks

Optimal Design of α-β-(γ) Filters

Flexible Neural Trees Ensemble for Stock Index Modeling

Further Mathematics for Engineering Technicians

Equations of Lines and Planes

The dynamic equation for the angular motion of the wheel is R w F t R w F w ]/ J w

Results of wake simulations at the Horns Rev I and Lillgrund wind farms using the modified Park model

GRAPHICAL METHOD OF FACTORING THE CORRELATION

Fitting Subject-specific Curves to Grouped Longitudinal Data

Lecture 3: Linear methods for classification

Design-Simulation-Optimization Package for a Generic 6-DOF Manipulator with a Spherical Wrist

NOWCASTING OF PRECIPITATION Isztar Zawadzki* McGill University, Montreal, Canada


Motion. Complete Table 1. Record all data to three decimal places (e.g., or or 0.000). Do not include units in your answer.

MATHEMATICS FOR ENGINEERING BASIC ALGEBRA

Course 8. An Introduction to the Kalman Filter

Time Series Data Mining in Rainfall Forecasting Using Artificial Neural Network

The Credibility of the Overall Rate Indication

Mechanics i + j notation

Modelling and Forecasting financial time series of «tick data» by functional analysis and neural networks

physics 112N magnetic fields and forces

Probability Hypothesis Density filter versus Multiple Hypothesis Tracking

On the Efficiency of Competitive Stock Markets Where Traders Have Diverse Information

Cloud Analytics for Capacity Planning and Instant VM Provisioning

MATH 185 CHAPTER 2 REVIEW

Modeling and Optimization of Performance of Four Stroke Spark Ignition Injector Engine

X X X a) perfect linear correlation b) no correlation c) positive correlation (r = 1) (r = 0) (0 < r < 1)

Field Data Recovery in Tidal System Using Artificial Neural Networks (ANNs)

The Big Picture. Correlation. Scatter Plots. Data

POTENTIAL OF STATE-FEEDBACK CONTROL FOR MACHINE TOOLS DRIVES

Typical Linear Equation Set and Corresponding Matrices

Deterministic Sampling-based Switching Kalman Filtering for Vehicle Tracking

A methodology for monitoring and leakage reduction in water distribution systems

Univariate Regression

Identification of Hybrid Systems

Transcription:

ANewExtensionoftheKalmanFiltertoNonlinear SimonJ.JulierSystems TheRoboticsResearchGroup,DepartmentofEngineeringScience,TheUniversityofOxford Oxford,OX13PJ,UK,Phone:+44-1865-282180,Fax:+44-1865-273908 siju@robots.ox.ac.ukuhlmann@robots.ox.ac.uk JereyK.Uhlmann optimality,tractabilityandrobustness.however,theapplicationofthekftononlinearsystemscanbedicult. TheKalmanlter(KF)isoneofthemostwidelyusedmethodsfortrackingandestimationduetoitssimplicity, ABSTRACT widelyusedlteringstrategy,overthirtyyearsofexperiencewithithasledtoageneralconsensuswithinthe trackingandcontrolcommunitythatitisdiculttoimplement,diculttotune,andonlyreliableforsystems whicharealmostlinearonthetimescaleoftheupdateintervals. ThemostcommonapproachistousetheExtendedKalmanFilter(EKF)whichsimplylinearisesallnonlinear modelssothatthetraditionallinearkalmanltercanbeapplied.althoughtheekf(initsmanyforms)isa sampledpointscanbeusedtoparameterisemeanandcovariance,theestimatoryieldsperformanceequivalentto thekfforlinearsystemsyetgeneraliseselegantlytononlinearsystemswithoutthelinearisationstepsrequired bytheekf.weshowanalyticallythattheexpectedperformanceofthenewapproachissuperiortothatofthe EKFand,infact,isdirectlycomparabletothatofthesecondorderGausslter.Themethodisnotrestricted Inthispaperanewlinearestimatorisdevelopedanddemonstrated.Usingtheprinciplethatasetofdiscretely toassumingthatthedistributionsofnoisesourcesaregaussian.wearguethattheeaseofimplementationand moreaccurateestimationfeaturesofthenewlterrecommenditsuseovertheekfinvirtuallyallapplications. Keywords:Navigation,estimation,non-linearsystems,Kalmanltering,sampling. mustbeestimatedfromnoisysensorinformation,somekindofstateestimatorisemployedtofusethedatafrom Filteringandestimationaretwoofthemostpervasivetoolsofengineering.Wheneverthestateofasystem 1 INTRODUCTION andobservationmodelsarelinear,theminimummeansquarederror(mmse)estimatemaybecomputedusing dierentsensorstogethertoproduceanaccurateestimateofthetruesystemstate.whenthesystemdynamics thekalmanlter.however,inmostapplicationsofinterestthesystemdynamicsandobservationequationsare nonlinearandsuitableextensionstothekalmanlterhavebeensought.itiswell-knownthattheoptimalsolution tothenonlinearlteringproblemrequiresthatacompletedescriptionoftheconditionalprobabilitydensityis anumberofsuboptimalapproximationshavebeenproposed6?8;13;16;21. maintained14.unfortunatelythisexactdescriptionrequiresapotentiallyunboundednumberofparametersand

TheEKFappliestheKalmanltertononlinearsystemsbysimplylinearisingallthenonlinearmodelssothat thetraditionallinearkalmanlterequationscanbeapplied.however,inpractice,theuseoftheekfhastwo well-knowndrawbacks: ProbablythemostwidelyusedestimatorfornonlinearsystemsistheextendedKalmanlter(EKF)20;22. 1.Linearisationcanproducehighlyunstableltersiftheassumptionsoflocallinearityisviolated. linearsystems,yetgeneraliseselegantlytononlinearsystemswithoutthelinearisationstepsrequiredbythe 2.ThederivationoftheJacobianmatricesarenontrivialinmostapplicationsandoftenleadtosignicant EKF.Thefundamentalcomponentofthislteristheunscentedtransformationwhichusesasetofappropriately InthispaperwederiveanewlinearestimatorwhichyieldsperformanceequivalenttotheKalmanlterfor implementationdiculties. chosenweightedpointstoparameterisethemeansandcovariancesofprobabilitydistributions.wearguethat theexpectedperformanceofthenewapproachissuperiortothatoftheekfand,infact,isdirectlycomparable tothatofthesecondordergausslter.further,thenatureofthetransformissuchthattheprocessand algorithmhassuperiorimplementationpropertiestotheekf.wedemonstratethedierencesinperformancein observationmodelscanbetreatedas\blackboxes".itisnotnecessarytocalculatejacobiansandsothe Kalmanltertononlinearsystems.Wearguethattheprincipleproblemistheabilitytopredictthestateof thenewlterrecommenditsuseovertheekfinvirtuallyallapplications. anexampleapplication,andwearguethattheeaseofimplementationandmoreaccurateestimationfeaturesof thesystem.section3introducestheunscentedtransformation.itspropertiesareanalysedandafullltering algorithm,whichincludestheeectsofprocessnoise,isdeveloped.insection4anexampleispresented.using realisticdata,thecomparisonoftheunscentedlterandekfforthetrackingofareentrybodyisconsidered. Thestructureofthispaperisasfollows.InSection2wedescribetheproblemstatementforapplyinga ConclusionsaredrawninSection5.Acompanionpaper10,extendsthebasicmethodandshowsthatjudiciously selectingadditionalpointscanleadtoanydesiredlevelofaccuracyforanygivenpriordistribution. 2.1ProblemStatement 2 ESTIMATIONINNONLINEARSYSTEMS WewishtoapplyaKalmanltertoanonlineardiscretetimesystemoftheform wherex(k)isthen-dimensionalstateofthesystemattimestepk,u(k)istheinputvector,v(k)istheqdimensionalstatenoiseprocessvectorduetodisturbancesandmodellingerrors,z(k)istheobservationvector (2) x(k+1)=f[x(k);u(k);v(k);k]; z(k)=h[x(k);u(k);k]+w(k); (1) andw(k)isthemeasurementnoise.itisassumedthatthenoisevectorsv(k)andw(k),arezero-meanand \predictor-corrector"structure.let^x(ijj)betheestimateofx(i)usingtheobservationinformationinformation TheKalmanlterpropagatesthersttwomomentsofthedistributionofx(k)recursivelyandhasadistinctive Ev(i)vT(j)=ijQ(i);Ew(i)wT(j)=ijR(i);Ev(i)wT(j)=0;8i;j: uptoandincludingtimej,zj=[z(1);:::;z(j)].thecovarianceofthisestimateisp(ijj).givenanestimate predictedquantitiesaregivenbytheexpectations ^x(kjk),thelterrstpredictswhatthefuturestateofthesystemwillbeusingtheprocessmodel.ideally,the P(k+1jk)=Ehfx(k+1)?^x(k+1jk)gfx(k+1)?^x(k+1jk)gTjZki: ^x(k+1jk)=ef[x(k);u(k);v(k);k]jzk (4) (3)

ofx(k),conditiononzk,isknown.however,thisdistributionhasnogeneralformandapotentiallyunbounded niteandtractablenumberofparametersneedbepropagated.itisconventionallyassumedthatthedistribution ofx(k)isgaussianfortworeasons.first,thedistributioniscompletelyparameterisedbyjustthemeanand numberofparametersarerequired.inmanyapplications,thedistributionofx(k)isapproximatedsothatonlya Whenf[]andh[]arenonlinear,theprecisevaluesofthesestatisticscanonlybecalculatedifthedistribution informative3. covariance.second,giventhatonlythersttwomomentsareknown,thegaussiandistributionistheleast thekalmanlteralinearupdateruleisspeciedandtheweightsarechosentominimisethemeansquarederror oftheestimate.theupdateruleis Theestimate^x(k+1jk+1)isgivenbyupdatingthepredictionwiththecurrentsensormeasurement.In P(k+1jk+1)=P(k+1jk)?W(k+1)P(k+1jk)WT(k+1) ^x(k+1jk+1)=^x(k+1jk)+w(k+1)(k+1); Itisimportanttonotethattheseequationsareonlyafunctionofthepredictedvaluesofthersttwomoments ofx(k)andz(k).therefore,theproblemofapplyingthekalmanltertoanonlinearsystemistheabilityto W(k+1)=Px(k+1jk)P?1 (k+1)=z(k+1)?^z(k+1jk) predictthersttwomomentsofx(k)andz(k).thisproblemisaspeciccaseofageneralproblem tobe (k+1jk): abletocalculatethestatisticsofarandomvariablewhichhasundergoneanonlineartransformation. form.supposethatxisarandomvariablewithmeanxandcovariancepxx.asecondrandomvariable,yis 2.2TheTransformationofUncertainty relatedtoxthroughthenonlinearfunction Theproblemofpredictingthefuturestateorobservationofthesystemcanbeexpressedinthefollowing WewishtocalculatethemeanyandcovariancePyyofy. Thestatisticsofyarecalculatedby(i)determiningthedensityfunctionofthetransformeddistributionand y=f[x]: (5) closedformsolutionsexist.however,suchsolutionsdonotexistingeneralandapproximatemethodsmustbe (ii)evaluatingthestatisticsfromthatdistribution.insomespecialcases(forexamplewhenf[]islinear)exact, ecientandunbiased. used.inthispaperweadvocatethatthemethodshouldyieldconsistentstatistics.ideally,theseshouldbe holds.thisconditionisextremelyimportantforthevalidityofthetransformationmethod.ifthestatisticsare Thetransformedstatisticsareconsistentiftheinequality notconsistent,thevalueofpyyisunder-estimated.ifakalmanlterusestheinconsistentsetofstatistics,it willplacetoomuchweightontheinformationandunderestimatethecovariance,raisingthepossibilitythat Pyy?Ehfy?ygfy?ygTi0 (6) ofthelefthandsideofequation6shouldbeminimised.finally,itisdesirablethattheestimateisunbiasedor thelterwilldiverge.byensuringthatthetransformationisconsistent,thelterisguaranteedtobeconsistent greatlyinexcessoftheactualmeansquarederror.itisdesirablethatthetransformationisecient thevalue aswell.however,consistencydoesnotnecessaryimplyusefulnessbecausethecalculatedvalueofpyymightbe consideringthetaylorseriesexpansionofequation5aboutx.thisseriescanbeexpressed(usingratherinformal ye[y]. Theproblemofdevelopingaconsistent,ecientandunbiasedtransformationprocedurecanbeexaminedby

notation)as: wherexisazeromeangaussianvariablewithcovariancepxx,andrnfxnistheappropriatenthorderterm f[x]=f[x+x] inthemultidimensionaltaylorseries.takingexpectations,itcanbeshownthatthetransformedmeanand =f[x]+rfx+12r2fx2+13!r3fx3+14!r4fx4+ (7) covariancearey=f[x]+12r2fpxx+12r4fex4+ Pyy=rfPxx(rf)T+1 13!r3fEx4(rf)T+: 24!r2fEx4?Ex2Pyy?EPyyx2+P2yy(r2f)T+ (8) Inotherwords,thenthordertermintheseriesforxisafunctionofthenthordermomentsofxmultipliedby thenthorderderivativesoff[]evaluatedatx=x.ifthemomentsandderivativescanbeevaluatedcorrectly (9) uptothenthorder,themeaniscorrectuptothenthorderaswell.similarcommentsholdforthecovariance byaprogressivelysmallerandsmallerterm,thelowestordertermsintheseriesarelikelytohavethegreatest equationaswell,althoughthestructureofeachtermismorecomplicated.sinceeachtermintheseriesisscaled impact.therefore,thepredictionprocedureshouldbeconcentratedonevaluatingthelowerorderterms. thisassumption, LinearisationassumesthatthesecondandhigherordertermsofxinEquation7canbeneglected.Under Pyy=rfPxx(rf)T: y=f[x]; (10) thesecondandhigherordertermsinthemeanandfourthandhigherordertermsinthecovariancearenegligible. ComparingtheseexpressionswithEquations8and9,itisclearthattheseapproximationsareaccurateonlyif However,inmanypracticalsituationslinearisationintroducessignicantbiasesorerrors.Anextremelycommon (11) andimportantproblemisthetransformationofinformationbetweenpolarandcartesiancoordinatesystems10;15. Thisisdemonstratedbythesimpleexamplegiveninthenextsubsection. 2.3Example returnspolarinformation(rangerandbearing)andthisistobeconvertedtoestimatetocartesiancoordinates. Thetransformationis: Supposeamobilerobotdetectsbeaconsinitsenvironmentusingarange-optimisedsonarsensor.Thesensor Thereallocationofthetargetis(0;1).Thedicultywiththistransformationarisesfromthephysicalproperties xy=rcos ofthesonar.fairlygoodrangeaccuracy(with2cmstandarddeviation)istradedotogiveaverypoorbearing rsinwithrf=cos?rsin sinrcos: tobeviolated. measurement(standarddeviationof15).thelargebearinguncertaintycausestheassumptionoflocallinearity

comparedwiththosecalculatedbythetruestatistics linearisation,itsvaluesofthestatisticsof(x;y)were Toappreciatetheerrorswhichcanbecausedby isticswereobtained.theresultsareshowninfigure1. usedtoensurethataccurateestimatesofthetruestat- totheslowconvergenceofrandomsamplingmethods, anextremelylargenumberofsamples(3:5106)were whicharecalculatedbymontecarlosimulation.due 1.04 Thisgureshowsthemeanand1contoursforwhich arecalculatedbyeachmethod.the1contouristhe locusofpointsfy:(y?y)p?1 1.02 graphicalrepresentationofthesizeandorientationof y(y?y)=1gandisa 1 0.98 0.96 isextremelysubstantial.linearisationerrorseectivelyintroduceanerrorwhichisover1.5timesthelatesthemeanatandtheuncertaintyellipseis 0.9 thepositionis1mwhereasinrealityitis96.7cm.this therangedirection,wherelinearisationestimatesthat Pyy.Ascanbeseen,thelinearisedtransformationis biasedandinconsistent.thisismostpronouncedinfigure1:themeanandstandarddeviationel- lipsesfortheactualandlinearisedformofthe 0.94 uncertaintyellipseissolid.linearisationcalcu- transformation.thetruemeanisatandthe 0.92 mittedeachtimeacoordinatetransformationtakes dashed. 0.3 0.2 0.1 0 0.1 0.2 0.3 place.eveniftherewerenobias,thetransformation itself,thesameerrorwiththesamesignwillbecom- isabiaswhicharisesfromthetransformationprocess standarddeviationoftherangemeasurement.sinceit True mean: x EKF mean: o isinconsistent.itsellipseisnotlongenoughintherdirection.infact,thenatureoftheinconsistencycompounds ismuchsmallerthanthetruevalue. theproblemofthebiased-ness:notonlyistheestimateorrinerror,butalsoitsestimatedmeansquarederror sizeofthetransformedcovariance.thisisonepossibleofwhyekfsaresodiculttotune sucientnoise mustbeintroducedtoosetthedefectsoflinearisation.however,introducingstabilisingnoiseisanundesirable solutionsincetheestimateremainsbiasedandthereisnogeneralguaranteethatthetransformedestimateremains consistentorecient.amoreaccuratepredictionalgorithmisrequired. Inpracticetheinconsistencycanberesolvedbyintroducingadditionalstabilisingnoisewhichincreasesthe 3.1TheBasicIdea 3 THEUNSCENTEDTRANSFORM foundedontheintuitionthatitiseasiertoapproximateagaussiandistributionthanitistoapproximatablewhichundergoesanonlineartransformation.itis Theunscentedtransformationisanew,novel methodforcalculatingthestatisticsofarandomvari- anarbitrarynonlinearfunctionortransformation23. (orsigmapoints)arechosensothattheirsamplemean TheapproachisillustratedinFigure2.Asetofpoints functionisappliedtoeachpointinturntoyieldacloud andsamplecovariancearexandpxx.thenonlinear oftransformedpointsandyandpyyarethestatisticsofthetransformedpoints.althoughthismethod baresasupercialresemblancetomontecarlo-typeform. Figure2:Theprincipleoftheunscentedtrans- Nonlinear Transformation

methods,thereisanextremelyimportantandfundamentaldierence.thesamplesarenotdrawnatrandom butratheraccordingtoaspecic,deterministicalgorithm.sincetheproblemsofstatisticalconvergencearenot anissue,highorderinformationaboutthedistributioncanbecapturedusingonlyaverysmallnumberofpoints. pointsgivenby Then-dimensionalrandomvariablexwithmeanxandcovariancePxxisapproximatedby2n+1weighted Xi+n=x?p(n+)PxxiWi+n=1=2(n+) X0 =x+p(n+)pxxiwi W0 =1=2(n+) ==(n+) where2<,p(n+)pxxiistheithroworcolumnofthematrixsquarerootof(n+)pxxandwiisthe (12) weightwhichisassociatedwiththeithpoint.thetransformationprocedureisasfollows: 1.Instantiateeachpointthroughthefunctiontoyieldthesetoftransformedsigmapoints, 2.Themeanisgivenbytheweightedaverageofthetransformedpoints, Yi=f[Xi]: 3.Thecovarianceistheweightedouterproductofthetransformedpoints, y=2nxi=0wiyi: (13) Thepropertiesofthisalgorithmhavebeenstudiedindetailelsewhere9;12andwepresentasummaryofthe Pyy=2nXi=0WifYi?ygfYi?ygT: (14) resultshere: 1.Sincethemeanandcovarianceofxarecapturedpreciselyuptothesecondorder,thecalculatedvalues ofthemeanandcovarianceofyarecorrecttothesecondorderaswell.thismeansthatthemeanis orderofaccuracy.however,therearefurtherperformancebenets.sincethedistributionofxisbeing calculatedtoahigherorderofaccuracythantheekf,whereasthecovarianceiscalculatedtothesame 2.Thesigmapointscapturethesamemeanandcovarianceirrespectiveofthechoiceofmatrixsquareroot approximatedratherthanf[],itsseriesexpansionisnottruncatedataparticularorder.itcanbeshown thattheunscentedalgorithmisabletopartiallyincorporateinformationfromthehigherorders,leadingto evengreateraccuracy. 3.Themeanandcovariancearecalculatedusingstandardvectorandmatrixoperations.Thismeansthatthe whichisused.numericallyecientandstablemethodssuchasthecholeskydecomposition18canbeused. 4.providesanextradegreeoffreedomto\netune"thehigherordermomentsoftheapproximation,and notnecessarytoevaluatethejacobianswhichareneededinanekf. algorithmissuitableforanychoiceofprocessmodel,andimplementationisextremelyrapidbecauseitis canbeusedtoreducetheoverallpredictionerrors.whenx(k)isassumedgaussian,ausefulheuristicisto selectn+=3.ifadierentdistributionisassumedforx(k)thenadierentchoiceofmightbemore appropriate.

5.Althoughcanbepositiveornegative,anegativechoiceofcanleadtoanon-positivesemideniteestimate densitydistributions8;16;21.inthissituation,itispossibletouseamodiedformofthepredictionalgorithm. Themeanisstillcalculatedasbefore,butthe\covariance"isevaluatedaboutX0(k+1jk):Itcanbe ofpyy.thisproblemisnotuncommonformethodswhichapproximatehigherordermomentsorprobability shownthatthemodiedformensurespositivesemi-denitenessand,inthelimitas(n+)!0, Inotherwords,thealgorithmcanbemadetoperformexactlylikethesecondOrderGaussFilter,but withouttheneedtocalculatejacobiansorhessians. (n+)!0y=f[x]+12r2fpxx;lim (n+)!0pyy=rfpxx(rf)t: meansand1contoursdeterminedbythedierent transformcanbeseeninfigure3whichshowsthe methods.thetruemeanliesatwithadottedcovariancecontour.thepositionoftheunscentedmean Theperformancebenetsofusingtheunscented value onthescaleofthegraph,thetwopointslieon topofoneanother.further,theunscentedtransform isindicatedbya?anditscontourissolid.thelinearisedmeanisatandusedadashedcontour.ascanbe seentheunscentedmeanvalueisthesameasthetrue thanthetruecontourintherdirection. isconsistent infact,itscontourisslightlylarger isbettersuitedthanlinearisationforlteringapplica- andeaseofimplementation,theunscentedtransform tions.indeed,sinceitcanpredictthemeanandcovari- ancewithsecondorderaccuracy,anylterwhichuses Givenitspropertiesofsuperiorestimationaccuracy Figure3:Theunscentedtransformasappliedto theunscentedtransformtothelteringproblemanddevelopstheunscentedlter. ans.thenextsubsectionexaminestheapplicationof doesnotrequirethederivationofjacobiansorhessi- theunscentedtransformwillhavethesameperformanceasthetruncatedsecondordergaussfilter1butthemeasurementexample. 3.2TheUnscentedFilter ThetransformationprocesseswhichoccurinaKalmanlterconsistofthefollowingsteps: Predictthenewstateofthesystem^x(k+1jk)anditsassociatedcovarianceP(k+1jk).Thisprediction Predicttheexpectedobservation^z(k+1jk)andtheinnovationcovarianceP(k+1jk).Thisprediction musttakeaccountoftheeectsofprocessnoise. Finally,predictthecross-correlationmatrixPxz(k+1jk): shouldincludetheeectsofobservationnoise. models.first,thestatevectorisaugmentedwiththeprocessandnoisetermstogiveanna=n+qdimensional Thesestepscanbeeasilyaccommodatedbyslightlyrestructuringthestatevectorandprocessandobservation 1.04 1.02 1 0.98 0.96 0.94 0.92 0.9 0.3 0.2 0.1 0 0.1 0.2 0.3 True mean: x EKF mean: o Kappa mean: +

1.ThesetofsigmapointsarecreatedbyapplyingEquation12totheaugmentedsystemgivenbyEquation15. 2.Thetransformedsetisgivenbyinstantiatingeachpointthroughtheprocessmodel, 3.Thepredictedmeaniscomputedas Xi(k+1jk)=f[Xai(kjk);u(k);k]: 4.Andthepredictedcovarianceiscomputedas ^x(k+1jk)=2na Xi=0WiXai(k+1jk): 5.Instantiateeachofthepredictionpointsthroughtheobservationmodel, P(k+1jk)2na Xi=0WifXi(k+1jk)?^x(k+1jk)gfXi(k+1jk)?^x(k+1jk)gT 6.Thepredictedobservationiscalculatedby Zi(k+1jk)=h[Xi(k+1jk);u(k);k] 7.Sincetheobservationnoiseisadditiveandindependent,theinnovationcovarianceis ^z(k+1jk)=2na Xi=1WiZi(k+1jk): 8.Finallythecrosscorrelationmatrixisdeterminedby P(k+1jk)=R(k+1)+2na Xi=0WifZi(kjk?1)?^z(k+1jk)gfZi(kjk?1)?^z(k+1jk)gT Pxz(k+1jk)=2na Xi=0WifXi(kjk?1)?^x(k+1jk)gfZi(kjk?1)?^z(k+1jk)gT vector, Box3.1:Thepredictionalgorithmusingtheunscentedtransform. Theprocessmodelisrewrittenasafunctionofxa(k), xa(k)=x(k) x(k+1)=f[xa(k);u(k);k] v(k): andtheunscentedtransformuses2na+1sigmapointswhicharedrawnfrom Thematricesontheleadingdiagonalarethecovariancesando-diagonalsub-blocksarethecorrelations ^xa(kjk)=^x(kjk) 0q1andPa(kjk)=P(kjk)Pxv(kjk) Pxv(kjk) Q(k): (15) betweenthestateerrorsandtheprocessnoises.althoughthismethodrequirestheuseofadditionalsigma points,itmeansthattheeectsoftheprocessnoise(intermsofitsimpactonthemeanandcovariance)are introducedwiththesameorderofaccuracyastheuncertaintyinthestate.theformulationalsomeansthat

expressionfortheunscentedtransformisgivenbytheequationsinbox3.1. correlatednoisesources(whichcanariseinschmidt-kalmanlters19)canbeimplementedextremelyeasily.the withprocessand/orobservationnoise,thentheaugmentedvectorisexpandedtoincludetheobservationterms. agivenapplication.forexample,iftheobservationnoiseisintroducedinanonlinearfashion,oriscorrelated Thissectionhasdevelopedtheunscentedtransformsothatitcanbeusedinlteringandtrackingapplications. Variousextensionsandmodicationscanbemadetothisbasicmethodtotakeaccountofspecicdetailsof ThenextsectiondemonstratesitsbenetsovertheEKFforasampleapplication. 4 EXAMPLEAPPLICATION ofthebodyistobetrackedbyaradarwhichaccurately athighaltitudeandataveryhighspeed.theposition lustratedinfigure4:avehicleenterstheatmosphere measuresrangeandbearing.thistypeofproblemhas Inthissectionweconsidertheproblemwhichisil- whichact.themostdominantisaerodynamicdrag, particularlystressfulforltersandtrackersbecauseof beenidentiedbyanumberofauthors1;2;5;17asbeing actonthevehicle.therearethreetypesofforces thestrongnonlinearitiesexhibitedbytheforceswhich tialnonlinearvariationinaltitude.thesecondtypeof whichisafunctionofvehiclespeedandhasasubstan- forceisgravitywhichacceleratesthevehicletowards atmosphereincreases,drageectsbecomeimportant thecentreoftheearth.thenalforcesarerandom buetingterms.theeectoftheseforcesgivesatrajectoryoftheformshowninfigure4:initiallythe almostvertical.thetrackingproblemismademore dicultbythefactthatthedragpropertiesofthe andthevehiclerapidlydeceleratesuntilitsmotionis trajectoryisalmostballisticbutasthedensityofthe vehiclemightbeonlyverycrudelyknown. isthesamplevehicletrajectoryandthesolidline Figure4:Thereentryproblem.Thedashedline trackanobjectwhichexperiencesasetofcomplicated,highlynonlinearforces.thesedependonthecurrent Insummary,thetrackingsystemshouldbeabletooftheradarismarkedbya. isaportionoftheearth'ssurface.theposition positionandvelocityofthevehicleaswellasoncertaincharacteristicswhicharenotknownprecisely.thelter's aerodynamicproperties(x5).thevehiclestatedynamicsare statespaceconsistsofthepositionofthebody(x1andx2),itsvelocity(x3andx4)andaparameterofits _x1(k)=x3(k) _x2(k)=x4(k) _x3(k)=d(k)x3(k)+g(k)x1(k)+v1(k) _x4(k)=d(k)x4(k)+g(k)x2(k)+v2(k) _x5(k)=v3(k) (16) whered(k)isthedrag-relatedforceterm,g(k)isthegravity-relatedforcetermandv(k)aretheprocessnoise terms.deningr(k)=px21(k)+x2(k)asthedistancefromthecentreoftheearthandv(k)=px23(k)+x24(k) x 2 (km) 500 400 300 200 100 0 100 200 6350 6400 6450 6500 x 1 (km)

asabsolutevehiclespeedthenthedragandgravitationaltermsare D(k)=?(k)exp[R0?R(k)] V(k);G(k)=?Gm0 Forthisexampletheparametervaluesare0=?0:59783,H0=13:406,Gm0=3:9860105andR0=6374and and(k)=0expx5(k): r3(k) reecttypicalenvironmentalandvehiclecharacteristics2.theparameterisationoftheballisticcoecient,(k), reectstheuncertaintyinvehiclecharacteristics5.0istheballisticcoecientofa\typicalvehicle"anditis scaledbyexpx5(k)toensurethatitsvalueisalwayspositive.thisisvitalforlterstability. andbearingatafrequencyof10hz,where Themotionofthevehicleismeasuredbyaradarwhichislocatedat(xr;yr).Itisabletomeasureranger rr(k)=p(x1(k)?xr)2+(x2(k)?yr)2+w1(k) w1(k)andw2(k)arezeromeanuncorrelatednoiseprocesseswithvariancesof1mand17mradrespectively4.the highupdaterateandextremeaccuracyofthesensormeansthatalargequantityofextremelyhighqualitydatais (k)=tan?1x2(k)?yr x1(k)?xr+w2(k) availableforthelter.thebearinguncertaintyissucientlythattheekfisabletopredictthesensorreadings accuratelywithverylittlebias. Thetrueinitialconditionsforthevehicleare x(0)= 0 B@?1:8093?6:7967 6500:4 349:14 0:6932 1 CAandP(0)= 2 64 10?610?6 10?6 Inotherwords,thevehicle'scoecientistwicethenominalcoecient. 0 0 10?60 0 0 3 75: Thevehicleisbuetedbyrandomaccelerations, Q(k)=242:406410?5 0 2:406410?50 0 Theinitialconditionsassumedbythelterare, 3 5 ^x(0j0)= 0 B@?1:8093?6:7967 6500:4 349:141CAandP(0j0)= 0 2 64 10?610?6 10?6 Thelterusesthenominalinitialconditionand,toosetfortheuncertainty,thevarianceonthisinitialestimate 0 0 0 10?60 0 3 75: is1ḃothlterswereimplementedindiscretetimeandobservationsweretakenatafrequencyof10hz.however, duetotheintensenonlinearitiesofthevehicledynamicsequations,theeulerapproximationofequation16was onlyvalidforsmalltimesteps.theintegrationstepwassettobe50mswhichmeantthattwopredictionswere

10 0 Mean squared error and variance of x 1 10 3 Mean squared error and variance of x 3 10 0 Mean squared error and variance of x 5 10 1 10 1 Position variance km 2 10 2 10 4 10 2 10 3 (a)resultsforx1. (b)resultsforx3. 10 4 10 3 10 5 10 5 Figure5:ThemeansquarederrorsandestimatedcovariancescalculatedbyanEKFandan 10 4 10 6 unscentedlter.inallthegraphs,thesolidlineisthemeansquarederrorcalculatedbytheekf, (c)resultsforx5. 10 7 10 6 10 5 0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200 Time s Time s errorandthedot-dashedlineitsestimatedcovariance. andthedottedlineisitsestimatedcovariance.thedashedlineistheunscentedmeansquared Time s madeperupdate.fortheunscentedlter,eachsigmapointwasappliedthroughthedynamicsequationstwice. error(thediagonalelementsofp(kjk))againstactualmeansquaredestimationerror(whichisevaluatedusing FortheEKF,itwasnecessarytoperformaninitialpredictionstepandre-linearisebeforethesecondstep. 100MonteCarlosimulations).Onlyx1,x3andx5areshown theresultsforx2aresimilartox1,andx4isthe sameasthatforx3.inallcasesitcanbeseenthattheunscentedlterestimatesitsmeansquarederrorvery accurately,anditispossibletobecondentwiththelterestimates.theekf,however,ishighlyinconsistent: TheperformanceofeachlterisshowninFigure5.Thisgureplotstheestimatedmeansquaredestimation thepeakmeansquarederrorinx1is0:4km2,whereasitsestimatedcovarianceisoveronehundredtimessmaller. error.finally,itcanbeseenthatx5ishighlybiased,andthisbiasonlyslowlydecreasesovertime.thispoor performanceisthedirectresultoflinearisationerrors. Similarly,thepeakmeansquaredvelocityerroris3:410?4km2s?2whichisover5timesthetruemeansquared theneedtoconsistentlypredictthenewstateandobservationofthesystem.wehaveintroducedanewltering InthispaperwehavearguedthattheprincipledicultyforapplyingtheKalmanltertononlinearsystemsis 5 CONCLUSIONS advantagesovertheekf.first,itisabletopredictthestateofthesystemmoreaccurately.second,itismuch algorithm,calledtheunscentedlter.byvirtueoftheunscentedtransformation,thisalgorithmhastwogreat lessdiculttoimplement.thebenetsofthealgorithmweredemonstratedinarealisticexample. Inacompanionpaper11,weextendthedevelopmentoftheunscentedtransformandyieldageneralframeworkfor itsderivationandapplication.itisshownthatthenumberofsigmapointscanbeextendedtoyieldalterwhich matchesmomentsuptothefourthorder.thishigherorderextensioneectivelyde-biasesalmostallcommon nonlinearcoordinatetransformations. Thispaperhasconsideredonespecicformoftheunscentedtransformforoneparticularsetofassumptions. Velocity variance (km/s) 2 Coefficient variance

[1]M.Athans,R.P.WishnerandA.Bertolini.Suboptimalstateestimationforcontinuous-timenonlinearsystemsfrom discretenoisymeasurements.ieeetransactionsonautomaticcontrol,tac-13(6):504{518,october1968. REFERENCES [2]J.W.AustinandC.T.Leondes.StatisticallyLinearizedEstimationofReentryTrajectories.IEEETransactionson [3]D.E.Catlin.Estimation,controlandthediscretekalmanlter.InAppliedMathematicalSciences71,page84. [4]C.B.Chang,R.H.WhitingandM.Athans.OntheStateandParameterEstimationforManeuveringReentry AerospaceandElectronicSystems,AES-17(1):54{61,January1981. [5]P.J.Costa.AdaptiveModelArchitectureandExtendedKalman-BucyFilters.IEEETransactionsonAerospaceand Springer-Verlag,1989. [6]F.E.Daum.Newexactnonlinearlters.InJ.C.Spall,editor,BayesianAnalysisofTimeSeriesandDynamic Vehicles.IEEETransactionsonAutomaticControl,AC-22:99{105,February1977. [7]N.J.Gordon,D.J.SalmondandA.F.M.Smith.NovelApproachtoNonlinear/non-GaussianBayesianState ElectronicSystems,AES-30(2):525{533,April1994. [8]A.H.Jazwinski.StochasticProcessesandFilteringTheory.AcademicPress,1970. Models,chapter8,pages199{226.MarcelDrekker,Inc.,1988. [9]S.J.JulierandJ.K.Uhlmann.AGeneralMethodforApproximatingNonlinearTransformationsofProbability Estimation.IEEProceedings-F,140(2):107{113,April1993. [10]S.J.JulierandJ.K.Uhlmann.AConsistent,DebiasedMethodforConvertingBetweenPolarandCartesian Distributions.WWWhttp://www.robots.ox.ac.uk/~siju,8/1994. CoordinateSystems.InTheProceedingsofAeroSense:The11thInternationalSymposiumonAerospace/Defense [12]S.J.Julier,J.K.UhlmannandH.F.Durrant-Whyte.ANewApproachfortheNonlinearTransformationofMeans [11]S.J.Julier,J.K.UhlmannandH.F.Durrant-Whyte.ANewApproachforFilteringNonlinearSystems.InThe ProceedingsoftheAmericanControlConference,Seattle,Washington.,pages1628{1632,1995. Sensing,SimulationandControls,Orlando,Florida.SPIE,1997.Acquisition,TrackingandPointingXI. [13]H.J.Kushner.Approximationstooptimalnonlinearlters.IEEETransactionsonAutomaticControl,AC-12(5):546{ andcovariancesinlinearfilters.ieeetransactionsonautomaticcontrol,1996. [15]D.LerroandY.K.Bar-Shalom.TrackingwithDebiasedConsistentConvertedMeasurementsvs.EKF.IEEE [14]H.J.Kushner.Dynamicalequationsforoptimumnon-linearltering.JournalofDierentialEquations,3:179{190, 556,October1967. [16]P.S.Maybeck.StochasticModels,Estimation,andControl,volume2.AcademicPress,1982. [17]R.K.Mehra.AComparisonofSeveralNonlinearFiltersforReentryVehicleTracking.IEEETransactionson TransactionsonAerospaceandElectronicsSystems,AES-29(3):1015{1022,July1993. [18]W.H.Press,S.A.Teukolsky,W.T.VetterlingandB.P.Flannery.NumericalRecipesinC:TheArtofScientic [19]S.F.Schmidt.Applicationsofstatespacemethodstonavigationproblems.InC.T.Leondes,editor,Advanced AutomaticControl,AC-16(4):307{319,August1971. [20]H.W.Sorenson,editor.Kalmanltering:theoryandapplication.IEEE,1985. ControlSystems,volume3,pages293{340.AcademicPress,1966. Computing.CambridgeUniversityPress,2edition,1992. [21]H.W.SorensonandA.R.Stubberud.Non-linearlteringbyapproximationoftheaposterioridensity.International [22]J.K.Uhlmann.Algorithmsformultipletargettracking.AmericanScientist,80(2):128{141,1992. [23]J.K.Uhlmann.Simultaneousmapbuildingandlocalizationforrealtimeapplications.Technicalreport,University JournalofControl,8(1):33{51,1968. ofoxford,1994.transferthesis.