决策层特征融合decision level identity fusion

Decision-LevelIdentityFusionTanXinLab5,SystemEngineeringDept.Contents1.Introduction2.Classicalinference3.Bayesianinference4.Dempster-Shafer’smethod*5.GeneralizedEvidenceProcessing(GEP)Theory6.Heuristicmethodsforidentityfusion7.Implementationandtrade-offsIntroductionDecision-levelfusionSeekstoprocessidentitydeclarationsfrommultiplesensorstoachieveajointdeclarationofidentity.(Featureextraction,identitydeclaration)Data-levelfusionFeature-levelfusionDecision-levelfusion(Datafused)(jointidentitydeclaration)IntroductionSensorASensorBSensorNFeatureExtractionIdentityDeclarationIdentityDeclarationIdentityDeclarationAssociationDecisionLevelFusion–IdentityFusionIntroductionDecision-LevelFusionTechniquesClassicalinferenceBayesianinferenceDempster-Shafer’smethodGeneralizedevidenceprocessingtheoryHeuristicmethodsClassicalinferenceStatisticalinferencetechniquesseektodrawconclusionsaboutanunderlyingmechanismordistribution,basedonanobservedsampleofdata.Classicalinferencetypicallyassumesanempiricalprobabilitymodel.Empiricalprobabilityassumesthattheobservedfrequencydistributionwillapproximatetheprobabilityasthenumberoftrials.lim()()nnPAfA()nkfAnherentrials,occurrenceofktimesTheoreticalbaseClassicalinferenceOnedisadvantageStrictlyspeaking,empiricalprobabilitiesareonlydefinedforrepeatableevents.Classicalinferencemethodsutilizeempiricalprobabilityandhencearenotstrictlyapplicabletononrepeatableevents,unlesssomemodelcanbedevelopedtocomputetherequisiteprobabilities.ClassicalinferenceMaintechnique–hypothesistestingDefinetwohypothesis1.Anullhypothesis,H0(原假设)2.Analternativehypothesis,H1(备择假设)Testlogic1.Assumethatthenullhypothesis(H0)istrue;2.ExaminetheconsequencesofH0beingtrueinthesamplingdistributionforstatistic;3.Performahypothesistest,iftheobservationhaveahighprobabilityofbeingobservedifH0istrue,thedeclarethedatadonotcontradictH0.4.Otherwise,declarethatthedatatendtocontradictH0.ClassicalinferenceMaintechnique–hypothesistestingTwoassumptionsarerequired1.anexhaustiveandmutuallyexclusivesetofhypothesiscanbedefined2.wecancomputetheprobabilityofanobservation,givenanassumedhypothesis.ClassicalinferenceGeneralizetoincludemultidimensionaldatafrommultiplesensors.Requiresaprioriknowledgeandcomputationofmultidimensionalprobabilitydensityfunctions.(aseriousdisadvantage)ClassicalinferenceAdditionaldisadvantages1.Onlytwohypothesescanbeassessedatatime;2.Complexitiesariseformultivariatedata;3.Donottakeadvantageofapriorilikelihoodassessment.Usage:identificationofdefectivepartsinmanufacturingandanalysisoffaultsinsystemdiagnosisandmaintenance.BayesianinferenceBayesianinferenceupdatesthelikelihoodofahypothesisgivenapreviouslikelihoodestimateandadditionalevidence(observations).Thetechniquemaybebasedoneitherclassicalprobabilities,orsubjectiveprobabilities.Subjectiveprobabilitiessufferalackofmathematicalrigororphysicalinterpretation.Nevertheless,ifusedwithcare,itcanbeusefulinadatafusioninferenceprocessor.Bayesianinference•Bayesianformulationii(/)()(,)(/)(/)()()iiiiiPEHPHPEHPHEPEHPHPESupposeH1,H2,…,Hi,representmutuallyexclusiveandexhaustivehypotheses()1iiPHBayesianinferenceFeatures1.provideadeterminationoftheprobabilityofahypothesisbeingtrue,giventheevidence.Classicalinferencegiveustheprobabilitythatanobservationcouldbeascribedtoanobjectorevent,givenanassumedhypothesis.2.allowincorporationofaprioriknowledgeaboutthelikelihoodofahypothesisbeingtrueatall.3.usesubjectiveprobabilitiesforaprioriprobabilitiesforhypothesis,andfortheprobabilityofevidencegivenahypothesis.Bayesianinference•Multisensorfusion•Foreachsensor,aprioridataprovideanestimateoftheprobabilitythatthesensorwoulddeclaretheobjecttobetypeigiventhattheobjecttobeoftypej,notedasP(Di|Oj).•ThesedeclarationsarethencombinedviaageneralizationofBayesianformulationdescribedbefore.Thisprovidesanupdated,jointprobabilityforeachpossibleentityOj.•InputtoBayesformulation:P(Di|Oj).foreachsensorandentityorhypothesisHi;P(Oj)aprioriprobabilities1(/,...,),1,2,...,jnPODDjMBayesianinferenceSensor#1ObservablesClassifierDeclarationSensor#2ETCSensor#nETCP(D1|Oj)P(D2|Oj)P(Dn|Oj)BayesianCombinationFormulaDecisionLogic:MAPThresholdMAPetcD1D2Dn12(/...)1,2,...,jnPODDDjMFusedIndentityDeclarationBayesianinferenceDisadvantages1.Difficultyindefiningpriorifunctions:P(Oj)2.Complexitywhentherearemultiplepotentialhypothesisandmultipleconditionallydependentevents3.Requirementsthatcompetinghypothesisbemutuallyexclusive:cannotassignevidencetoobjectOiandOj.4.Lackofanabilitytoassigngeneraluncertainty.Bayesianinference•AnIFFNExample111212122212(/)(/)...(/)(/)(/)...(/)............(/)(/)...(/)MMnnnMPDOPDOPDOPDOPDOPDOPDOPDOPDO11212()(/)...(/)(/&...)()[(/)(/)...(/)]jjnjjniiiniiPOPDOPDOPODDDPOPDOPDOPDO1(){(),...,()}jMPOPOPOIdentification-friend-foe-neutralsystemdevelopedbyFerrante,Inc.oftheU.K.Thissystemusesmultiplesensorsdesignedtooperateonboardanaircrafttoperformjointdeclarationsofidentitytodeterminewhetherobservedaircraftarefriendly,potentialenemies,orneutral.Dempster-Shafer’smethod