无线传感网络中的目标分类融合

Vol.32,No.6ACTAAUTOMATICASINICANovember,2006ClassificationFusioninWirelessSensorNetworks1)LIUChun-Ting1HUOHong1FANGTao1LIDe-Ren2SHENXiao11(InstituteofImageProcessing&PatternRecognition,ShanghaiJiaotongUniversity,Shanghai200240)2(StateKeyLaboratoryofInformationEngineeringinSurveying,MappingandRemoteSensing,WuhanUniversity,Wuhan430079)(E-mail:[e-mail],[e-mail])AbstractInwirelesssensornetworks,targetclassificationdiffersfromthatincentralizedsensingsystemsbecauseofthedistributeddetection,wirelesscommunicationandlimitedresources.Westudytheclassificationproblemofmovingvehiclesinwirelesssensornetworksusingacousticsignalsemittedfromvehicles.Threealgorithmsincludingwaveletdecomposition,weightedk-nearest-neighborandDempster-Shafertheoryarecombinedinthispaper.Finally,weuserealworldexperimentaldatatovalidatetheclassificationmethods.Theresultshowsthatwaveletbasedfeatureextractionmethodcanextractstablefeaturesfromacousticsignals.ByfusionwithDempster′srule,theclassificationperformanceisimproved.KeywordsWirelesssensornetworks,classificationfusion,waveletdecomposition,weightedk-nearest-neighbor,Dempster-Shafertheory1IntroductionTheconvergenceofsensors,microelectromechanicalsystem(MEMS)andwirelesscommunicationtechnologiesmakeswirelesssensornetworks(WSNs)viable.WSNsarecomposedofalargeamountofsensornodes.Theyaredenselydeployed,spatiallydistributedandself-organized.Sensornodesareautomaticallyorganizedtodetect,processandcommunicatethephysicalinformationofthenetworkcoveredarea.Sensorsmayhavemodestcapabilities,butmayachievehighandlargescaleperformancethroughnodescoordination[1].TheflexibilitiesandextendedcoveragemakeWSNsaccomplishdetectiontasksthattraditionalsinglesensorsystemcannotdo.Theappealingapplicationareasincludemilitary,commercial,health,environmentalmonitoringanddetection[2].ThelargevolumeofdatainsenornetworksandthelimitedcommunicationandstoragecapabilitiesmakedatafusioninWSNsbecomenecessary.Fusioninsensornetworksisnotexactlythesameasthatintraditionalmulti-sensorsystems.Sensornodesmusttransportthelocalinformationtoothernodesorsinknodetoachievethefusionprocess.SomenewissuesrelatedtofusioninWSNshavebeengreatcon-cerned,suchasroutingandclusteringprotocolsforefficientdatatransmission[3,4],fusionstructures[5,6],nodetaskallocationformalism[7],collaborativesignalprocessing[8]andtimesynchronization[5,9].TargetclassificationisatypicalapplicationofWSNs.ThelimitationofWSNsshowsthatclassi-ficationmethodsadoptedinthisareaneedthemeritsoflowcostofcomputationandstorageresource,aswellasreductionofcommunicationbetweennodes.TargetclassificationinWSNsalwaysadoptsthedecisionfusionstrategy.Sensornodesprocesstherawdataofsometypeofsensingmodalitiestomaketheclassificationdecisionandfusethedecisionsofeachlocalnodetoreachaglobalresult.Methodstoclassifymovingvehiclesusingpassiveacousticsignalshavebeenwidelyanalysised[10∼12].Passiveacousticsensorshavetheadvantagessuchaspassive,affordable,smallsize,andnon-line-of-sightdetection.Soundsemittedfromdifferentmovingvehiclescanexhibitdistinguishablecharacteristics.In[13],ahierarchicalclassificationandfusionprocessinWSNshasbeenproposed.Itincludesfea-tureextraction,classification,multi-modalandtemporalfusioninlocalsensornodesanddistributedmulti-sensorfusion.In[14],theauthorsconsideredtherelationshipofsensortotargetdistanceandsignaltonoiseratioandproposedadistancebasedfusionscheme.Amaximumaposterior(MAP)probabilityfusionalgorithmbasedonBayesianestimationwasproposed.ThismethodcanimprovetheclassificationaccuracybyusingthedistributedinformationofWSNs.Inthispaper,wefocusontwoaspectsofthetargetclassificationproblem:Analyzingpassiveacousticsignalstoextractdistinctfeaturesforefficientmovingvehicleclassification;studyingthefu-sionmethodtoimprovetheclassificationperformanceinsensornetworksenvironment.Fourierbasedacousticfeatureextractionmethodscontainthefrequencydiscrepancyofdifferentvehicletypes,buttheequalresolutionofthewholefrequencydomainwillhaveacertainredundancy.Principlecomponentanalysis(PCA)offrequencyvectorcanreducethecorrelationoffeatureelementsandexpressreliability1)SupportedinpartbyScience&TechnologyDepartmentofShanghai(05dz15004)ReceivedDecember8,2005;inrevisedformMarch21,2006948ACTAAUTOMATICASINICAVol.32butthecomputationofeigenvalueandeigenvectorisdifficult.Italsoneedsstablerecordconditionfortrainingsamples[15]whichisunavailableinWSNssurveillanceapplications.Waveletbasedmethodpro-videsatime-frequencyandmulti-resolutionanalysismeasure,notonlyextractsfrequencycomponentofsignals,butalsolocatesthefrequenciesintimedomain.Waveletbasedmethodforvehicleclassificationhasbeenstudiedbymanyresearchersandgotsatisfyingclassificationperformance[10,12].Weadoptthewaveletbasedmethodforacousticsignalextractioninthispaper.Afterpatternextraction,toclassifythenewfeaturevectorx,weuseaweightedk-NNclassificationrule.k-NNruleisanonparametricclassificationmethod.Itissimpleandeasytouse.Anextensionofk-NNclassifierwasproposedin[16],whichgivestheknearestneighborsdifferentweightsaccordingtothedistancebetweenxandtheneighbors.InWSNs,thedisparatesen