ProceedingsofSPIEConferenceonSignalandDataProcessing ofSmallTargets,SanDiego,CA,USA,July-August2001.(4473-41) A Survey of Maneuvering Target Tracking—Part III: Measurement Models (cid:3) X.RongLi and VesselinP.Jilkov DepartmentofElectricalEngineering UniversityofNewOrleans NewOrleans,LA70148,USA 504-280-7416(phone),504-280-3950(fax),[email protected],[email protected] Abstract Thisisthethirdpartofaseriesofpapersthatprovideacomprehensivesurveyofthetechniquesfortrackingmaneuvering targetswithoutaddressingtheso-calledmeasurement-originuncertainty.PartI[1]andPartII[2]dealwithgeneraltargetmo- tionmodelsandballistictargetmotionmodels,respectively. Thispartsurveysmeasurementmodels,includingmeasurement model-based techniques, used intargettracking. ModelsinCartesian,sensormeasurement, theirmixed, andothercoordi- natesarecovered.Thestressisonmorerecentadvances—topicsthathavereceivedmoreattentionrecentlyarediscussedin greaterdetails. KeyWords: TargetTracking,MeasurementModel,Survey 1 Introduction Thispaperisthethirdpartofaseriesthatprovidesacomprehensivesurveyoftechniquesformaneuveringtargettracking withoutaddressingtheso-calledmeasurement-originuncertainty. Mostmaneuveringtargettrackingtechniquesaremodelbased; thatis, theyrelyonexplicitlytwo descriptions: onefor thebehaviorsofthe target,usuallyintheformofamotion(ordynamics)model,andtheotherforourobservationsofthe target, knownas anobservationmodel. A surveyoftargetmotionmodelsin generaland ballistic targetmotionmodelsin particular has been reported in Part I [1] and Part II [2], respectively. This part surveys the measurement models and the relevantmodelingtechniques. Moreprecisely,thispapersurveysthemodelsofmeasurementscharacterizedbythefollowing: theyaretrulyoriginated fromthe“pointtarget”undertrack(i.e.,thereisnooriginuncertainty);andtheyaremeasurements,ratherthanobservations inamoregeneralsense, whichmaycontainotherinformation,includingtargetfeaturesasprovidedbyanimagingsensor. Also,thissurveyisconcernedwiththemathematicalmodelsasabasisformaneuveringtargettracking. Forexample,their other applications are not addressed and the actual sensor models are not of concern. Further, this surveyincludes some aspectsofestimationandfilteringtechniquesthatarehighlydependentonandthushardlyseparablefromthemeasurement models. As in our survey of motion models [1, 2], we highlight underlying ideas and try to clarify both explicit and implicit assumptionsinvolvedineachmodel,inanattempttorevealprosandconsofthemodels.Aconsiderableamountofdiscussion isgiventowardsthisend,muchofwhichcannotbefoundelsewhere. Weremindthereader,however,thatthesediscussions, althoughintendedtobeaccurateandbalanced,areobviouslynotnecessarilyfreeofourpersonalpreferenceandbias. Also, wefocusonmorerecentadvancesinmeasurementmodels—topicsthathavereceivedmoreattentionrecentlyarediscussed ingreaterdetails.Furthermore,somemodelsandtechniquespresentedinthispaperhavenotyetappearedelsewhere. AsstatedinPartI,weappreciatereceivingcommentsandanymissingmaterialthatshouldbeincludedinthispart. Therestofthepaperisorganizedasfollows.Sec.2describesmeasurementmodelsintheoriginalsensorcoordinates.Sec. 3providesageneralviewoftherolesofcoordinatesystemsinmaneuveringtargettracking.Linearizedmeasurementmodels in(Cartesian-sensor)mixedcoordinatesarepresentedinSec.4.MeasurementmodelsconvertedtoCartesiancoordinatesare coveredinSec.5.PseudomeasurementmodelingtechniquesaresurveyedinSec.6.Finally,concludingremarksaregivenin Sec.7. ResearchsupportedbyONRGrantN00014-00-1-0677andNSFGrantECS-9734285. (cid:3) 1 2 ModelsinSensorCoordinates Sensorsusedfortargettrackingprovidemeasurementsofa targetina naturalsensorcoordinatesystem(CS) orframe. In many cases (e.g., with a dish radar), this CS is spherical in 3D or polar in 2D with range , bearing (or azimuth) , elevation (Fig. 1)1, andpossiblyrangerate(orDoppler) . (Wedonotexplicitlyconsiderdirerctmeasurementsoftargebt height,aseprovidedby,e.g.,ModeC.Suchmeasurementsarer_usuallynotavailablefornon-cooperativetargets.)Notallthese measurementcomponentsare available fromall sensors. For example, some active sensors may not providerangerate or elevationangle,whilepassivesensorsprovideonlyangles(althoughpassiverangingispossible). Weconsidergenerallythe 3Dcase—therespective2Dcasefollowsinastraightforwardmanner. Inthesensorcoordinates,thesemeasurementsaregenerallymodeledinthefollowingformofadditivenoise (1) r = r+vr (2) b = b+vb (3) e = e+ve (4) r_ = r_+vr_ where denotestheerror-freetruetargetpositioninthesensorsphericalcoordinates,and aretherespective random(rm;be;aes)urementerrors. Weassumethesemeasurementsaremadeattime (or forshvorrt;)vbbu;vtew;evr_willomitthetime index wheneverpossiblewithoutambiguities. ItisnormallyassumedthatthetskemeaksurementerrorsinthesensorCSare zero-mkean,Gaussiandistributed,anduncorrelated: with cov diag (5) 2 2 2 2 vk (0;Rk) Rk = (vk)= ((cid:27)r; (cid:27)b; (cid:27)e; (cid:27)r_) where isth(cid:24)eeNrrorvector2attime and isawhitenoisesequence. 0 vk =[vr;vb;ve;vr_]k k vk f g 6 z r (cid:0)1 (cid:0)1 cos v cos u e y - b x = Fig.1: Sensorcoordinatesystems. Theabovemeasurementmodelinsphericalcoordinatesismostcommonfortrack-while-systems,e.g.,rotatingsurveil- lanceradars[3,4,5].Forscan-while-tracksurveillancesystems[6],suchasphasedarrayradars[7,8,9],thesensorprovides measurementsintermsofthedirectioncosines and —insteadofthebearing andelevation —ofthetargetposition (Fig.1)relativetoreferenceaxes.TheRUVmeausuremventmodelisasgivenabovebwith(2)–(3)repelacedby (6) u = u+vu (7) v = v+vv where and denote the error-free target position direction cosines, and are the respective measurement errors. Sometiumesatvhirddirectioncosinemeasurement isusedforconvevnui;evnvce,albeitredundant( ). w =w+vw w = p1 u2 v2 1Otherconventionsarealsoused.Forexample,thebearingmaybedefinedastheanglefromthe axis,ratherthanfromthe axis. (cid:0) (cid:0) 2WeuseSansSerifletters(e.g., )todenoteidealerror-freequantitiesandbold-faceletterstodenyotevectors. x z 2 Itisalsonormallyassumedthattheerrorsarezero-mean,Gaussiandistributed,anduncorrelated: with cov diag (8) 2 2 2 2 vk (0;Rk) Rk = (vk)= ((cid:27)r; (cid:27)r_; (cid:27)u; (cid:27)v) (cid:24)N with . 0 Nvokte=th[avtrt;hveuR;vUvV;-vCr_]Skisnotorthogonal. Nevertheless,theaboveuncorrelatednessassumptioncov diag , 2 2 iswelljustifiedbythefactthat , ,and aremeasuredbythreephysicallyindependentsystem(vs.k) = ((cid:27)r; (cid:27)u 2 2 (cid:27)v;T(cid:27)hr_e)abovetwomodelsarisenaturallryfuromthvemeasurementprocess. TheyarelinearandGaussianandcanbewritten compactlyinvector-matrixnotation (9) z=Hx+v; v (0;R) where or , (cid:24)oNr , or 0 0 0 0 0 z = [r;,b; e; r_] ,zan=d [ri;suan; vid;er_n]tityxm=atr[irx;abn;de; r_i;s:a::z]eromxat=rix.[r; u; v; r_; :::] v = [vr;vb;ve;vr_] v = 0 [vr;Rvuan;vgve;avnr_d] anHgl=em[Ie;aOsu]remenItsmayhavevastlydifferenOtaccuracies. Forexample,aphasedarrayradarhasrangemea- surementsmuchmoreaccuratethananglemeasurements—itserrorellipsoidlookslikeapancakenormaltotherangevector; thesituationis reversedfora continuous-waveradar,whichoftenhas a cigar-shapederrorellipsoidalongtherangevector [8]. Thislinearmodeliscompletelyuncoupledacrossdifferentcoordinates. Thisishighlydesirableforestimationandfil- teringinanumberofaspects. Forexample,efficientparallelprocessingmaybeaccomplishedwithlittleornoperformance degradation. More important, coordinate-decoupledfilters may be implementedthat mitigate the possible ill conditioning arisingfromthevastlydifferentaccuraciesinmeasuringrangeandangles[8]. Thesedecoupledfiltersmaypossiblyoutper- formthetheoreticallysuperiorfull-blown“optimal”filtersinthepresenceofreallyillconditioning. 3 TrackinginVariousCoordinates Various coordinatesystems (CS) have been used in target tracking, including the Earth-centeredinertial (ECI), Earth- centered (Earth) fixed (ECF, ECEF, or ECR), East-North-Up (ENU), and radar face (RF) coordinate systems. A concise description of these coordinate systems has been given in Part II [2]. Many factors affect the selection of a coordinate systems[8,10,11,9]. TheENU-CSisacommonchoicefortacticalsystemswithrelativelylimitedsensormotion,suchas inaplatform-centricsystem. TheECI-CS,alongwithitsvariantECF-CS,isatypicalchoiceforastrategicsysteminvolving multipleplatforms. As far as tracking accuracy is concerned, the probably best choice of a coordinate system in principle is to align its coordinatesto the principalaxes of the trackingerror ellipsoid [8]. This will avoid the corruptionof an accurate estimate componentbyinaccurateones.Italsoprovidesagoodframeworkagainstillconditioning.Sincetheseprincipalaxesusually vary with respect to time in a complex way, a sensible strategy is to align the coordinates to the principal axes of either the measurement error ellipsoid or dynamic error ellipsoid, depending on which error has more important directionality properties. Following this principle, several coordinate systems were discussed in [8] in the context of a single sensor, includingradar-oriented,target-oriented,andtheircombinations. A measurementis oftendescribedin a sensorreferenceframe, whichis usuallystabilized relativeto the motionof the sensor,evenifadifferentCSisselectedfortrackingpurposes.Itisingeneraldifferentfromaplatform(orsite)CS(e.g.,the ENU-CS)whenmultiplesensorsareinvolvedintheplatform(e.g.,aship). WhilethegeographicalENUframecenteredata sensorisconvenientforarotatingtrack-while-scanradar,thesensor-specificradarfaceCSismoreoftenusedwithaphased arrayradar,where and axesareintheradarfaceplaneand axisalongtheboresightdirection. Fordetailedcoxnsideryationsofthecoordinatessystemsandztherespectivetransformations,thereaderisreferredto[7,8, 10,12,11,9]. In the sequel, by a Cartesian coordinate system, we mean a generic one unless otherwise is stated explicitly; and by a sensor frame (or CS), we mean non-Cartesian (spherical or RUV) CS in which the measurements are available directly withoutcoordinatetransformation. TargetmotionisbestdescribedinaCartesianCS,butmeasurementsareavailablephysicallyinasensorCS.Assuch,there arebasicallyfourpossibilitiesofdotracking:trackinginmixedcoordinates,inCartesiancoordinates,insensorcoordinates, andinothercoordinates.Thesearedescribednext. 3 3.1 TrackinginMixedCoordinates Thisisthemostpopularapproach.Thetargetdynamicsandmeasurementsaremodeledby (10) z = h(x)+v wherethetargetstate andprocessnoiseareintheCartesiancoordinates,butmeasurement anditsadditivenoise arein thesensorcoordinatesx3. Let bethetruepositionofthetargetintheCartesiancoordiznates.Forthecaseofspvherical measurements,wehave (x;y;z) and with 0 0 0 z=[r; b; e; r_] h(x)=[r;b;e;r_] =[hr;hb;he;hr_] (11) hr = r= x2+y2+z2 (12) hb = b=ptan(cid:0)1 y x (13) he = e=tan(cid:0)1 z 2 2 x +y (14) xx_ +yy_+zz_ hr_ = r_ = p 2 2 2 x +y +z ForRUVmeasurements,wehave and with z=[r; u; v; r_]0 h(x)p=[r;u;v;r_]0 =[hr;hu;hv;hr_]0 (15) x hu = u= 2 2 2 x +y +z (16) y hv = v= p 2 2 2 x +y +z (and ). Clearly the measurement models are nonlinear and coupled across Cartesian coordinates, hw = w = 2 z2 2 p although the measuprexm+eynt+nzoise remains zero mean, Gaussian, and uncorrelated because measurements are in the sensor coordinates. Most nonlinearestimation and filtering techniques, such as the extendedKalman filters (EKF), formaneuveringtarget trackinghavebeenappliedinthisframework. Thosebasedonmeasurementmodelsareaddressedinsubsequentsections. Manyothertechniquesarecoveredinsubsequentpartsofthissurveyseries. AtypicalimplementationoftheEKFinmixed coordinates(Cartesianstateandsphericalmeasurements)canbefoundin[13]. 3.2 TrackinginCartesianCoordinates Inthisapproach,themeasurementsinthesensorcoordinatesareconvertedtotheCartesiancoordinatesfortracking.Clearly, anymeasurementexpressedinthesensorcoordinateshasanexactandequivalentrepresentationintheCartesiancoordinates. Let betheequivalentrepresentationintheCartesiancoordinatesoftheerror-freesensormeasurement 0 xp =or[x;y;z] ,=wHithxtarget state and some , for example, if . Clearly, is in fact 0 (thr;ebt;reu)e pos(irt;ioun;vo)f the target in the Cxartesian cooHrdinates, not knoHwn=to[Iu;s.O]Oncxe=the[xn;oyis;yz;m:e:a:s]urements4 xofpthe target positionare convertedto the Cartesian coordinates(i.e., the noisymeasurementsoriginallyavailablein sensor coordinates are expressed in the Cartesian coordinates), the measurement equation takes the following “linear” form in the Cartesian coordinates: ,thatis, zc =xp+vc (17) zc =Hx+vc Thismeasurementissometimesreferredtoasapseudolinearmeasurement. Thismodelapparently“eliminates”theneedto handlenonlinearmeasurements,incontrasttotheaboveapproachoftrackinginmixedcoordinates. Themajoradvantageof thisapproachisthatalinearKalmanfilterthencanbeappliedifthedynamicsislinear. Priorto[14]themeasurementnoise wascrudelytreatedtohavezeromeanandacovariancedeterminedbyafirst-order Taylorseriesexpansion.Sincethensevevracltechniqueshavebeendevelopedtocomputeoraccountforthenonzeromeanand thecovariancemoreaccurately(see,e.g.,[14,15,16,17,18,19]).ThesetechniquesaresurveyedinSec.5. Weemphasizethatthemeasurementnoise isingeneralnotonlycoupledacrosscoordinates,non-Gaussian,butalso state dependent. This state dependency is provbacbly more important but is largely ignored or overlooked in the literature 3Thismeasurementmodelwithadditivenoisecorrespondsto(9). Thenoise isnotnecessarilyinthesensorframeifthemoregeneralmodel isused. v z = h(x4;Tvh)atis,theactualobservedvalues(i.e.,realizations),nottheobservablesas(random)variables. 4 beyonditsimplicituseinthecomputationofthefirsttwomomentsof .Duetothenonlineardependencyof onthestate ,thismeasurementmodelisinfactnonlinear. Asaresult,eveniftvhecmeasurementconversionisdoneideavllcywithexact xknpowledgeofthe(state-dependent)firsttwomomentsof ,itisstillanillusionthattheapplicationoftheKalmanfilterhere inthecaseoflineardynamicsyieldsoptimalresults. Nevvecrtheless,sincethestatedependence(i.e.,nonlinearity)existsonly inthemeasurementnoise ,ratherthaninthemeasurementfunction ,itseemsreasonabletoexpectthatitsimpacton trackingperformanceisrevlactivelysmaller,ascomparedtothenonlineahri(ty)in whenhandledbymostpopularnonlinear (cid:1) filteringtechniques. Ontheotherhand,amajordrawbackofthisapproachstemhs(fr)omalackofavailabletechniquestohandle (cid:1) measurementswithstate-dependent,non-Gaussianerrors,whereasabundanttechniquesareavailableformeasurementswith nonlinear . WehavedevelopedanextensionoftheKalmanfilterforsuchproblemsthatexplicitlyaccountsforthestate dependenche(o)fthemeasurementnoisemoreeffectively,asreportedin[20]. (cid:1) Another weakness of this approach is that the conversion from sensor to Cartesian coordinates requires knowledge of range.Forangle-onlymeasurements,anestimatedrangecanbeused.However,theconvertedmeasurementshaveadegraded accuracywhenaninaccuraterangeisused,suchasforpassivesensorsorrange-denialcountermeasures. Indeed,angle-only measurementsarerarelyconvertedtotheCartesiancoordinateswithfewexceptions,oneofwhichisgivenin[21].Inaddition, itisdifficulttodevelopcoordinate-decoupledfiltersinpureCartesiancoordinatesformitigatingthepossibleillconditioning duetothelargedifferenceintheaccuraciesofrangeandanglemeasurements,aswellasforhighefficiency. Themeasurementconversionasdescribedabovedoesnotdealwithrangeratemeasurements . Itissubstantiallymore complexwhentherangeratemeasurementsareinvolved. Inthiscase,theconvertedmeasuremenr_tsarenonlinear(noteven pseudolinear). The use of as a measurementof position and velocity for (cid:1) trackingintheCartesiancodor=dinrar_te=swxaxs_s+ugygye_s+tedzzi_n+[2v2d].Notethatthismeasurementisq(xua;dyr;azt)icinthestate,w(xh_;icy_h;zi_s)not highlynonlinear. Thisisclearlysuperiortoconvertingtherangeratemeasurements directly,whichis xx_+yy_+zz_ highlynonlinear. Foruncorrelatedrangeandrangerateerrors,however,themeasurer_m=entpxh2+asy2a+nze2rr+orvr_ with zeromeanbutvariance ,whichcanbequitelargeforlong-rangetargetsd. vd = rr_ rr_ r2(cid:27)2r_ +r_2(cid:27)2r+(cid:27)2r(cid:27)2r_ (cid:0) 3.3 TrackinginSensorCoordinates Alternatively,targetdynamicscanbeconvertedfromtheCartesiantothesensorcoordinatessothatthedesirablemeasurement structureisunaltered,incontrasttoconvertingmeasurementsfromsensortoCartesiancoordinates.However,expressingtyp- ical targetmotionsin sensorcoordinates(sphericalorRUV) leadsto highlynonlinear,coordinate-coupled,andsometimes cumbersomemodels. For example, a constant-velocity(CV) motion has a simple Cartesian description with two or three independenttwo-state one-dimensionalCV models. The same motion in the spherical coordinatesis rather nonlinearand complicated,anexplicitmodelofwhichcanbefoundin,e.g.,[3]. Nontrivial,variableaccelerations(knownaspseudoaccel- erations)[10,4,11]areinducedinthesensorcoordinatesbysuchaconversion,evenforaperfectCVmotion,andthusastate vectorincludingaccelerationcomponentsisneeded.Inshort,itisimpossibletodescribetypicaltargetmotionsinthesensor coordinatesinasimple,coordinate-uncoupledway.Further,theconvertedprocessnoiseisnon-Gaussianandstatedependent evenifitisGaussianandcoordinate-uncorrelatedintheoriginalCartesiancoordinates. Nevertheless, this approachhas certain advantages. The foremost oneis that the linear, uncoupled,Gaussian structure ofthemeasurementmodelis maintained. Alargenumberoftrackingfiltersthatoperatepurelyinsensorcoordinateshave appeared in the literature. Their common feature is the use of the above linear-Gaussian measurement model. Their key differenceliesinhowthetargetdynamicsaremodeled.Adetailedcoverageofthesemodelsisbeyondthescopeofthispart, whichissupposedtocovermeasurementmodels.Forcompleteness,however,wementionbrieflybelowthesetechniquesand directthereadertothespecificreferences. Asimplisticapproachistodirectlyemploysomedecoupled1Dtargetdynamicsmodels,suchastheCV,CA,andSinger models(see Part I),forrange(rangerate)and othermeasurements(angles ordirectioncosines) separately. This approach accountsforthetargetdynamicsinthesensorcoordinatesinacrudeway; itdoesnotreallyconvertthetargetdynamicsto thesensorcoordinates.Asexplainedabove,however,high-ordermodelsthatincludeaccelerationsareneededto“cover”the actualhighlynonlineardynamicsinthesensorcoordinatesevenforatrulyCVmotion. Thisleadstoaccuracydegradation. The geometry-inducedpseudoaccelerationsare clearly not accountedfor if a two-state “CV” modelis used in each of the sensorcoordinatesindependently.Anengineeringfixistocompensatetheresultingbias[23,24]. Amoreeffectiveapproachistousethetargetdynamicsmodelactuallyconvertedinthesensorcoordinates.Thisleadsto afilteringproblemwithlinearuncoupledmeasurementsinGaussiannoisebutnonlineardynamicsandnon-Gaussian,state- dependentprocess noise. Albeit theoreticallyand computationallychallenging, this approachis beneficial in a number of cases[7,25,26,27,5]. Decoupledfirst-orderMarkovmotionmodelsinpolarcoordinatesformaneuveringaircrafttracking wereusedin[28,29,30]. Asimilar,decoupledmotionmodelinsphericalcoordinateswasproposedin[31]. Itscompletely 5 coupled version was derived from the Cartesian version in [32]. For the developments in the context of ballistic target tracking, the reader is referredto Part II. For example, [5] reportedthe developmentof a tracking filter where the reentry vehicledynamicsaremodeleddirectlyinthesphericalcoordinates. Thesefiltersoperateentirelyinthesensorcoordinates. Sec.4.2.1ofPartIIcontainsamoredetaileddiscussionofcomparisonbetweenreentry-vehicletrackinginCartesianandin sensorcoordinates. 3.4 TrackinginOtherCoordinates AlthoughtargetmotionandmeasurementsarebestdescribedinCartesianandsensorcoordinates,respectively,itisclearly notnecessaryto dotrackingentirelyin oneorbothofthese coordinates. ThemodifiedCartesian coordinates[33,34] and the better-knownmodifiedpolar coordinates[35] for angle-onlytrackingare goodexamples. Also, it is fairly commonto propagatethe targetstate in a Cartesian frameand then convertthe predictedstate, alongwith the errorcovariance,to the sensor coordinates for state update there (see, e.g., [7, 36, 37, 12, 27]). As such, while state update is decoupled across coordinates, state predictionis in generalcoupled. This approachrelies heavily on coordinatetransformation: In addition to the conversionof the predictedstate, the updatedstate andits errorcovariancemust be convertedbackto theCartesian coordinates. The covariance conversions usually rely on linearization of the error models and is possibly biased, not to emphasizethestatedependencyinherentintheapproach. Alternatively,theuseoftheso-calledradarprincipalCartesiancoordinateswassuggestedin[8],whichisanintegration of the Cartesian and the original sensor coordinates in that the range vector in the original sensor frame is retained as an axisinthisorthogonalCartesianframe. (Theothertwoaxesquantifyangularcomponents,oneparalleltotheradarface,the otherintheplanenormaltotheradarfaceandcontainingtherangevector.) Inthesamespirit,aschemebasedonrangeand angularmodelswasdevelopedin[10,11]intheorthogonalrange-horizontal-“vertical”(RHV)frame5,whichinvolverange (andrangerate)and, in Cartesian (H andV) coordinates, angularvelocityandacceleration. Similar approacheswere also takenin[38,22,39,40,41,42]. Thankstotheweakcouplingbetweentherange(rangerate)andthenon-rangecoordinates, a merit of this approachis that range and non-rangecoordinatesare processed in a quasi-independentmanner, capable of alleviatingillconditioningandhavinghighefficiency. However,themeasurementmodelsinthenon-rangecoordinatesare nolongerlinear.Inessencethisapproachcombines,inasensibleway,theframeworksinpurelysensorcoordinatesdescribed abovebyusingthe rangecoordinate,andin mixedcoordinatesas describedin Sec. 3.1forthenon-rangecoordinates. An additionaladvantageofusingrangeasacoordinateisthatrangeratemeasurementscanbeincorporatednicelyandeasily. 4 Linearized ModelsinMixedCoordinates Throughoutthispaper,weconsideronlyagenericfilteringcyclefrom to (i.e.,from to )andwrite for thepredictedstate and fortheupdatedstate . Theassociatedtke(cid:0)rr1orcotvkariancematkrices1aredkenotedby x(cid:22)and ,respectively. x^kjk(cid:0)1 x^ x^kjk (cid:0) P(cid:22) P The“standard”techniqueforhandlingthenonlinearmeasurementmodel(10)istheextendedKalmanfilters(EKF)(see, e.g.,[43,44,45])6.Ingeneral,itreliesonapproximatingthenonlinearmeasurementbythefirstfewtermsofitsTaylorseries expansion.Specifically,thecornerstoneofitsfirst-orderversion,whichismostwidelyused,islinearizationofthenonlinear model,resultinginaderivative-basedlinearizedmodel.Otherlinearizedmodelshavealsobeenproposedtohandlenonlinear measurements.Wedescribetheselinearizedmodelsnext. 4.1 Derivative-Based Themostwidelyusedtechniqueforlinearizinganonlinearmeasurementmodelintheformof(10)istoexpandthemeasure- mentfunction atthepredictedstate andignoreallnonlinearterms7: h(x) x(cid:22) (18) @h h(x) h(x(cid:22))+ (x x(cid:22)) (cid:25) @x x=x(cid:22) (cid:0) (cid:12) 5TheHaxisisreallyhorizontal,buttheVaxisisnotreallyvertical.Theyareb(cid:12)othperpendiculartotherangevector. 6ThefirstapplicationoftheKalmanfiltertoareal-lifeproblemwasinfactint(cid:12)heformofanEKF(see,e.g.,[46]). (cid:12) 7Moregenerally,ifacompletelygeneralnonlinearmodel isconsidered,wewouldhave z=h(x;v) h(x;v)(cid:25)h(x(cid:22);v(cid:22))+ @@hx(cid:12)(cid:12)(cid:12)(cid:12)x=x(cid:22)(x(cid:0)x(cid:22))+ @@hv(cid:12)(cid:12)(cid:12)(cid:12)v=v(cid:22)(v(cid:0)v(cid:22)) Suchamodeldoesnotnecessarilyhaveadditivenoiseinthesensorcoordinates;forexample,itmayhaveadditivenoiseintheCartesiancoordinates. 6 Thisamountstoapproximatingthenonlinearmodel(10)bythelinearmodel (19) z=H(x(cid:22))x+d(x(cid:22))+v where istheJacobianof and . @h ForHth(ixs(cid:22))m=ode@lxthxe=px(cid:22)redictedstateandithsc(oxv)ariancde(ax(cid:22)r)e=updha(txe(cid:22)d)usiHng(tx(cid:22)h)ex(cid:22)linearKalmanfilterequations (cid:0) (cid:12) (cid:12) (20) K = P(cid:22)H0 HP(cid:22)H0+R (cid:0)1 (21) x^ = x(cid:22)+K(z (cid:22)z) (cid:0) (cid:0) (cid:1) (22) P = (I KH)P(cid:22) = (I(cid:0)KH)P(cid:22)(I KH)+KRK0 (23) (cid:0) (cid:0) where and . Notethat isusedonlyinthecovarianceupdateand filtergaHin=coHmp(ux(cid:22)t)ationaz(cid:22)n=dtHhat(x(cid:22)(2)3x(cid:22))+isvda(lix(cid:22)d)f+orEar[bvi]tr=aryhg(ax(cid:22)in)+Ean[vd] . ThesefHactsareusedinsometechniquesdiscussed later. Althoughappearsalmosteverywhere,covarianceupdateKby(22H)shouldbeavoidedforatleasttworeasons: Itinvites horriblenumericalproblemsandit is theoreticallyvalidonlywhenthe gain is trulyoptimal, whichis rarelythecasein practice. Ithas beenlargelyoverlookedthatthegaingivenby(20)is nolonKgeroptimalandthuscanbeimprovedsinceit ignoresthelinearizationerrors. Thislinearizedmodelisadequateonlywhen issufficientlysmall,whichcanrarelybeguaranteedsincetheaccuracy of reliesonthatoftargetstatepropagxatiox(cid:22)n(i.e.,dynamicsmodel)andthepreviousstateestimate .This (cid:0) inax(cid:22)cc=urax^ckyjkm(cid:0)a1ybuildupandresultinfilteringdivergence,asreportedinnumerousexamples(see,e.g.,[43,47x^])k.(cid:0)T1ejkc(cid:0)hn1iques aimedatreducinglinearizationerrorsarediscussedinSec. 4.4. 4.2 Difference-Based We present now a new linearized model, proposed in [48], that is not only expectably more accurate but also potentially simplerthantheabovewidelyusedderivative-basedmodel. h((cid:1)) x(cid:22) x(cid:3) x h(cid:0)1(z) Fig.2: Variouslinearizations. Considerfirstascalarnonlinearmeasurement forsimplicity.Let z =h(x)+v (24) h(x) h(x(cid:22)) H(x;x(cid:22))= (cid:0) ; x=x(cid:22) x x(cid:22) 8 6 Clearly, istheslopeofthestraightlineconnecting (cid:0) and (seeFig.2).Forconvenience,wedenote (cid:3) (cid:3) H(x ;x(cid:22)) h(x ) h(x(cid:22)) (25) h(x) h(x(cid:22)) @h H(x(cid:22);x(cid:22))= lim (cid:0) =H(x(cid:22))= x!x(cid:22) x x(cid:22) @x x=x(cid:22) whichistheslopeofthetangentof at .If isabett(cid:0)erestimatethan ,itis(cid:12)(cid:12)reasonabletoexpectthat h(x) x(cid:22) x(cid:3) x(cid:22) (cid:12) (cid:12) (26) (cid:3) z =h(x(cid:22))+H(x ;x(cid:22))(x x(cid:22))+v (cid:0) 7 isabetterlinearizedmodelthanthederivative-basedmodel . Inthevectorcase,thisdifference-basedlinearizedmodelzo=f(h10(x(cid:22)))is+H(x(cid:22);x(cid:22))(x x(cid:22))+v (cid:0) (27) (cid:3) z=h(x(cid:22))+H(x ;x(cid:22))(x x(cid:22))+v where (cid:0) H(x(cid:3);x(cid:22))=[Hij]; Hij = hi(x(cid:3)j;(cid:3)x(cid:22))(cid:0)hi(x(cid:22)); hi =ithrowofh; hi(x(cid:3)j;x(cid:22))=hi(x)x=[x(cid:22)1;:::;x(cid:22)j(cid:0)1;x(cid:3)j;x(cid:22)j+1;:::;x(cid:22)n]0 xj x(cid:22)j j (28) (cid:0) Clearlyitisextremelyeasytoimplementthislinearizedmodel. ItdoesnotinvolvecomputationofanyJacobian,which could be theoretically and/or computationally challenging for a complicated nonlinear function . It is expectably more accuratein generalthan the derivative-basedlinearizedmodel, as widelyused in the EKF, providhed is a moreaccurate (cid:3) estimateof than . x Severalxways ox(cid:22)f determining are possible. First, without loss of generality for tracking applications, assume that (cid:3) , where is invertxible. Let . In the case of a 3D measurement of the target position, for 0 0 0 (cid:0)1 hex=am[phl1e;,h2]wouldbeht1he3Dtargetpositionx.1W=e cahn1th(ezn)choose . Forthecomponents (cid:3) 0 0 0 (cid:0)1 0 0 corresponxdi1ngto ,thederivatives canbeuxse=d.[Axl1te;rx(cid:22)n2a]ti=ve[lhy,1we(zc)a;nx(cid:22)fi2]rstupdatethestateestimHatiej (cid:3) @hi from to asinxajn=EKx(cid:22)Fj (noneedforcovHairjia=nce@uxpdxa(cid:3)jt=ex(cid:22)hjerethough)andthenuse intheabovelinearizedmodel. (cid:3) Theux(cid:22)seofx^this modelwill leadtoat least anexpecta(cid:12)blymoreaccuratecovarianceuxpdat=efx^ora state updateinthe formof (cid:12) [48]: x^ =x(cid:22)+K(z z(cid:22)) (29) (cid:0) P =[I KH(x(cid:3);x(cid:22))]P(cid:22)[I KH(x(cid:3);x(cid:22))]0+KRK0 (cid:0) (cid:0) 4.3 OptimallyLinearizedModel Theabovederivative-basedanddifference-basedlinearizationmodelsingeneralhavenooptimalityandcanbequitebadin manycases. We nowoutlinealinearizedmodelthatisoptimalinthemean-squareerror(MSE)sense[44],aspresentedin [48]fortracking. Anonlinearfunction canbeapproximatedoptimallyaround byalinearone: h(x) x(cid:22) (30) h(x) a+H(x x(cid:22)) inthesenseofhavingtheminimumMSE,denoting (cid:25) , (cid:0) x~ =x x(cid:22) (cid:0) (31) 0 J =E[(h(x) a Hx~)(h(x) a Hx~)] Itcanbeshownthat[48] (cid:0) (cid:0) (cid:0) (cid:0) (32) a = E[h(x)] E[h(x)x~0]P(cid:22)(cid:0)1E[x~] =(1 E[x~]0P(cid:22)(cid:0)1E[x~]) f (cid:0) g (cid:0) (33) H = E[h(x)x~0] E[h(x)]E[x~0] P(cid:22)(cid:0)1(I E[x~]E[x~]0P(cid:22)(cid:0)1)(cid:0)1 where . Inthecaseof f ,itred(cid:0)ucesto g (cid:0) P(cid:22) =E[x~x~0] E[x~]=0 (34) a=E[h(x)]; H =E[h(x)x~0]P(cid:22)(cid:0)1 Consideran exampleofa scalar nonlinearmeasurement . Assume that . Thenthe optimally linearizedmodelis z = x3 +v x (x(cid:22);P(cid:22)) (cid:24) N (35) z =x(cid:22)3+3P(cid:22)x(cid:22)+(3x(cid:22)2+3P(cid:22))(x x(cid:22))+v since and . Com(cid:0)paredwiththederivative-basedlinearizedmodel a = E[x3] = x(cid:22)3 +3,Pw(cid:22)x(cid:22)hichalHwa=ysEun[dxe3rx~e]sPt(cid:22)im(cid:0)1at=est3hx(cid:22)e2v+ari3aPt(cid:22)ion ,thismodelappearsmoreappealingfor 3 2 zma=nyx(cid:22)sit+ua3tix(cid:22)on(sx. x(cid:22))+v h(x) h(x(cid:22)) (cid:0) (cid:0) Whilethederivative-basedlinearizationreliesontruncationoftheTaylorseries,whichwillincurlargeerrorsif isnot small,thisoptimallylinearizedmodelaccountsforlargeerrorswithintheexpectationsbytheprobabilisticweightsanx~dthus tendstogiveamoreconservativefiltergainandbetterperformanceforcasesinvolvinglarge . Anotherpossibleadvantage ofthismodelisthat neednotbedifferentiable.Basically,ittradesintegrationwithdifferentiax~tion.Thismaybeparticularly usefulinsuchcaseshwherehardlimiters(orsaturations)areinvolved. In the calculationof the requiredexpectations, one may usually assume . It has a small tail probability, whichisappealingbecausethegoalislocallinearization. Insomesituations,xonemay(x(cid:22)w;aP(cid:22)nt)touseadistributionflatterthan (cid:24) N theGaussian,buttheheavytailsshouldbetruncated;thatis, maybeassumedmoreevenlydistributedthanGaussian,but onlyovera“small”neighborhoodof . Ingeneral,thelargertx~heneighborhood,themoreconservativethefiltergain. x(cid:22) 8 4.4 Linearization-ErrorReductionTechniques Sequential Processing. A well-known simple means to reduce linearization errors is sequential processing of the mea- surementcomponents(see,e.g.,[49,50,51]). Itiswellknownthatthenonlinearmeasurementsshouldbeprocessedinthe orderoftheir accuracy— moreaccuratefirst — (see, e.g., [51, 11]). A particularproblemwas consideredin [52], where thesphericalmeasurementswereprocessedsequentiallyintheorderofdecreasingaccuracy: azimuth,elevation,andrange. Performancecomparisonresultsbetweensequentialprocessingandtheconventionalvectorprocessingweregivenin[53]. Iterative EKF. Once the updated state is obtained, the nonlinear measurement model can be re-linearized at . This willingeneralreducethelinearizationerrox^rcomparedwithlinearizationat . Thestateanditserrorcovariancecanx^thenbe re-updatedbasedonthere-linearizedmodel. Thisprocesscanberepeated,x(cid:22)resultinginanalgorithmknownasaniterated EKF(IEKF)inthecontextofKalmanfiltering[43]. FollowingTheorem8.2of[43]theiterationalgorithmis (36) 0 x^ = x(cid:22) (37) i+1 i i i i x^ = x(cid:22)+K(x^ ) z h(x^ ) H(x^ )(x x^ ) ; i=0;1; :::;L (cid:0) (cid:0) (cid:0) (38) L+1 x^ = x^ (cid:2) (cid:3) (39) P = I K(x^L)H(x^L) P(cid:22) I K(x^L)H(x^L) 0+K(x^L)RK(x^L)0 (cid:0) (cid:0) where , , ifaKalmanfilteringisusedwitha derivatHive(-x^bia)se=dl@@inhxeaxr=izx^eidKmo(xd^ei)l(=i(cid:2).eP.(cid:22),Hin(ax^ifi)r0s[Ht-o(rx^die)rP(cid:22)E(cid:3)HK(F(cid:2)x^).i)A0+simRi]l(cid:0)a1ritier=ati0o;(cid:3)n1;ca:n::b;eLwrittenifadifference-basedmodel (Sec.4.2)isused.Incontrasttosomeotheriterationschemes(see,e.g.,[54,47]),thecomputationoftheupdatedcovariance (cid:12) bytheJoseph’sfo(cid:12)rm(39)isbettertobeoutsideoftheiterationloop.Thiswasusedin[43],andemphasizedanddiscussed Pin[55]. A theoreticalconsiderationoftheEKF andIEKFmeasurementupdatesas Gauss-Newtoniterationschemesanda demonstrationof the superiority of the IEKF were presented in [56]. The simulations of [7] (Table I) and [55] show that suchre-linearizationiterationscanindeedimproveaccuracyatalevelthatisscenariodependent[55]. Itshouldbewarned, however,thatanimprovementisnotguaranteed.Therearereportsthattherelinearizationiterationdegradestheperformance. Thereaderisreferredto[57]forsomeinsightthatsubstantiatesthiswarning. Higher-OrderPolynomialModels. Anotherstraightforwardideatoincreasetheaccuracyofapolynomialapproximation ofanonlinearmeasurementmodelistousequadraticterm(andpossiblyhigher-orderterms)intheTaylorseriesexpansion. IntheKalmanfilteringcontext,thisleadstowhatissometimesreferredtoasasecond-order(andhigher-order,respectively) EKF[43,47,58]. Thesimulationresultsreportedin[7]showaconsiderableimprovementinperformanceofasecond-order EKFoverafirst-orderEKF. However,second-orderEKFsarenotveryoftenusedinpracticemainlybecauseoftheirrather burdensomecomputationandlimitedormarginalperformanceimprovement. Followingthe sameideas it is alsopossibleto develophigher-orderversionsofthedifference-basedandoptimallylin- earizedmodels. Many other techniques are available for mitigating the performancedegradationdue to linearization, such as artificial inflationoferrorcovariance[58].[8]includesashortlistofsuchtechniquesandabriefdiscussion. 5 ModelsinCartesianCoordinates Sincetargetmotionis best describedinCartesian coordinatesbutmeasurementsareavailableinsensorcoordinates,as explainedin Sec. 3.2, a commonlyused methodis to convertmeasurementsfrom sensor to Cartesian coordinates, and do tracking entirely in the Cartesian coordinates. As before, we will assume a generic Cartesian frame since measurements betweenCartesianframescanbeconvertedeasilyandexactly(assumingnosensorregistrationorgridlockerrors). 5.1 ConversionofMeasuredPositions Thespherical-to-Cartesiantransformation with isgivenby (cid:0)1 0 ’=h h=[hr;hb;he] xc rcosbcose (40) zc = yc =’(z)=’(r; b;e)= rsinbcose 2 3 2 3 zc rsine 9 4 5 4 5 where and are one and the same noisy measurement, expressed in the original spherical 0 0 coordinzate=s a[rn;dbt;hee] convezrcted=C[axrcte;syica;nzcc]oordinates, respectively. The RUV-to-Cartesian transformation with (cid:0)1 isgivenby (cid:30) = h 0 h=[hr;hu;hv;hw] ru (41) zc =(cid:30)(z)=(cid:30)(r; u;v;w)= rv 2 3 rw Ifrangemeasurementsarenotavailable, asinthecaseofpassivesensorsandrange-denialcountermeasures,theabove 4 5 rangemeasurement canbereplacedbyanestimatedrange [21,11]andtherangemeasurementerror (anditsbias andvariance)shouldrbereplacedbytherangeestimationerrorr^ (anditsbiasandvariance). r r Inthesequel,wedescribetechniquesforconvertingmeasurr^emerntsfromsphericaltoCartesiancoordinat(cid:0)es.Thereaderis (cid:0) referredto[59,20]forconversionfromRUVtoCartesiancoordinates. 5.2 StandardModelofConvertedMeasurements Afterconversion,themeasurementmodelinCartesiancoordinateshastheform (42) zc =Hx+vc where isthepositionsubvectorofthestatevector and standsfortheresultingmeasurementerror. ByxTpay:=lorHsexriesexpansionof aroundthenoisymeaxsuremevnct ,wehave ’(z) z HOT (43) xp =’(z)=’(z v)=’(z) J(z)v+ (v) (cid:0) (cid:0) where istheerror-freetruetargetpositioninsphericalcoordinatesandHOT standsforthehigherorder( ) 0 terms,zan=d[trh;ebJ;aec]obian isevaluatedatthenoisymeasurement (v) 2 (cid:21) J(z) z cosbcose rsinbcose rcosbsine (44) @’ (cid:0) (cid:0) J(z)= = sinbcose rcosbcose rsinbsine @z(cid:12)z=z 2 sine 0 (cid:0) rcose 3 (cid:12) Thentheexactlyconvertedmeasurement(cid:12)model(442)canbewrittenas 5 (cid:12) HOT (45) zc =’(z)=xp+vc =xp+J(z)v (v) (cid:0) vc Clearly,thisexpansionissuperiortoexpanding aroundtheerror-freemeasurement ’(z) | {z } z HOT (46) zc =’(z)=’(z+v)=xp+vc =xp+J(z)v+ (v) vc whichinvolvestheunknown . z Evidently,thetrueconvertedmeasurementerror ismeasurementdep|endent({ozrstated}ependent),non-Gaussian,cor- relatedacrosscoordinates,andhasnonzeromean. Tvhucs,theconvertedmeasurement hasabias andaconditional bias . zc E[vc] CEle[avrlcy,z]conversion of the measured position values per se is straightforward and nothing is really subtle here. The j maintaskintheso-calledmeasurementconversion,whichisbettercalledmeasurementmodelconversion,reallyliesinthe conversionoftheassociatednoisestatistics. AsexplainedinSec.3.2,themajoradvantageofexpressingmeasurementsin Cartesiancoordinatesistheattractive“linear”structureofthecorrespondingmeasurementmodel(42). Inorderforalinear filter(e.g.,theKalmanfilter)totakeadvantageofthis“linear”structure,thefirsttwomomentsof mustbedetermined. vc 5.3 LinearizedConversion The“standard”approachistotreat (approximately)aszero-meanwithcovariance[50,22,10,45,11] vc (47) L 0 R =J(z)RJ(z) 10