复杂背景条件下红外弱小点目标实时检测技术研究

上海交通大学硕士学位论文复杂背景条件下红外弱小点目标实时检测技术研究姓名：曾明申请学位级别：硕士专业：控制理论与控制工程指导教师：李建勋20060201II:12K-out-of-NK-out-of-NMHTK-out-of-NKNIIIBayesianK-out-of-NIVRESEARCHONDIMPOINTTARGETREALTIMEDETECTIONINCOMPLICATEDBACKGROUNDABSTRACTLongrangeinfrareddimpointtargetdetectionisthekeytechnologyinsuchsystemsasinfraredsurveillanceandtrackingsystem(IRST),precisionguidancesystem,infraredearly-warningsystem,widefield-of-viewtargetsurveillancesystem,satelliteremotesensingsystem.Infraredsystemisusedinmilitaryfieldwidelyforitsworkingdayandnightpassivedetectionresistinterferencestronghidedetectingcamouflagetargetandsoon.Therefore,studiesonreal-timedimpointtargetdetectionalgorithmsinlowSNRimagesequencescanresultinlongercombatrangeandresponsetime,whichisofgreatimportancetosurvivalprobability.Totheresearchofinfrareddimtargetdetectionandrecognitioninternalandexternal,someproblemsneedtohandleasfollowing:(1)Mostresearchersemploymorphologicalfilterswithfixedstructuralelementforsingleframedetectionatpresent.Unfortunatelytheydoagoodjobonlyinsomecorrespondingimagemodel.Howeverusuallytheimagesignalsareextremelycomplicatedandchangeallthetime.Sothesefilterscannotworkwellforinfrareddimtargetinlongdistance,resultinginleakdetectionandfalsealarm.(2)Dimpointtargetdetectionusuallymakesuseofmultipleframesinformation.Atpresent,someofmethodsformulti-framesassociationdetectionareadoptedincluding“K-out-of-N”ruleassociationdetectionandMHTalgorithmmainly.Althoughthe“K-out-of-N”ruleassociationalgorithmhaslowcomputation,itworkswithpoordetectionperformance;likewise,althoughMHTalgorithmcanworkwell,itisdifficulttopracticeforthehugecomputation.Therefore,anewpracticalmethodwhichworkswellbothatdetectionperformanceandalgorithmcomputationisneededurgentlyforassociationdetection.VInfrareddimpointtargetdetectionincludingsingleframeandmulti-frameinlowSNRsequentialimagesisresearchedinthispaper.Themajorstudiesarefollowing.(1)Combiningmorphologicalfilterwithneuralnetandgeneticalgorithmseparatelyfortargetdetectionofsingleframe.Then,theneuralnetandgeneticalgorithmsoptimizethestructuralelementandmorphologicaloperatoroffilter.Thus,themorphologicalfiltercanworkwithspecificintelligenceandhasoptimizationprocessfordimpointtargetdetectioninsophisticatedbackground.Meanwhile,adaptivethresholdisadoptedinsingleframefilteringinordertoimprovedetectionperformance.(2)“K-out-of-N”sequentialimagesassociationdetectionalgorithmisstudiedextensivelyandinformationfusingisintroduced.Newtoninterpolationisintroducedfirsttimeforoptimizationthefusionschemesoastogettheoptimalperformancefor“K-out-of-N”regulation.InadditionconfidencelocalassociationandMHTareintroducedinto“K-out-of-N”ruleassociationdetectionalgorithmseparately.Eventually,thedetectionperformanceandcomputationarederivedsimulatedandcompareddetailedly.(3)Studyingdistributedfuzzydecisionfusion.Inordertosolvedistributeddecisioninformation-fusingproblems,fuzzytheoryisintroducedintodecisionfusionsystembasedonminimumBayesianriskcriterionandthedistributedfuzzydecisioninformation-fusingmethodisdesigned,thustheoptimalfusionruleisderivedatthefusioncenter.Thenthisdistributedfuzzydecisionfusiontheoryisintroducedintotargetdetectioninsequentialimages.Keywords:infraredpointtargetdetection,infraredassociationdetectioninsequentialimages,neuralnet,geneticalgorithm,adaptivethreshold,“K-out-of-N”rulealgorithm,distributeddecisioninformationfusion.I160304007203F570033ATR51483020203JW0303451476010604JW03035III2006220IV2006220200622011.120600.10.050.1;21.21980201,Top-Hat[1-5]P.T.Jackway[6]Top-Hat3,2B)1(Mibi≤≤0/133×55×77×000110Top-Hat3ParallelGeneticAlgorithmPGABhandarkaEhrgardt4Huttunen4Blostein(1.1)k1.1Figure1.1Treedatastructureoftargettrajectory5k1illy=∑1im×2iσ×1im×2iσ×1m0mαβ10(/)PHHα=1.110(/)PHHα=1.21001()()()iiiYuYuuYuλλλ≥≤1.3()iYλ22110102210(/)()ln(/)2iiillifYHmmmmYYifYHλσσ=−−==−∑1.411111illilliililYYbbYaα===≥≤∑∑∑1.52101102immauimmσ+=+−1.62100102immbuimmσ+=+−1.760u1u0ln1uaβ=−1.811lnuaβ−=1.9iN0H1H1NiN11NiNiττ==≥≤∑∑1.10111010[()()]Nmmmmστφαφβ−−=+−1.11φ1N11222110[()()]/()Nammφφβσ−−=+−1.12Blosteink57()()kkYHYH|Pr|Prmax01()kYH|Pr1()kYH|Pr0i1−+GiMarkov1+kkk()kkknsfs,1=+n()()()()()kkKskkKYHYHYHYHk|Pr|Prmaxmax|Pr|Prmax011)(01)(−=K1−KkGM/M()2/MGTonissen199668ReedPoratXiongReed4N(N)Bar-Shalom1.320192,34K-out-of-NK-out-of-NMHT1.403F570036030400751483020203JW031051476010604JW030310K-out-of-NKNBayesianmatlab112.1[7],[8]Top-Hat[1-5]P.T.Jackway[6]Top-Hat,Top-Hat[1-5],,[9-13][14-16]Won[11]RitterGX[12]GranaM[13][3][10]HarveyNR[14][15-16]12Top-HatKoivistoP[17]ZhaoChunhui[18]Kraft[19]Top-HatTop-Hat2.21964SerraMatheron[20]3×3B3×3FFB13)}1,1(,),1,0(),1,1(max{1,11,01,1−++−+=⊕−−BFBFBFBF(2.1)F)}1,1(,),1,0(),1,1(min{1,11,01,1−−−−−=−−BFBFBFB(2.2),,xyFFx,yFBFBF(=BB⊕)(2.3)BBFBFΘ⊕=•)((2.4)Top-HatTop-HatTop-HatTop-Hat))(()(,xBFFxOTHBF−=(2.5)))(()(,xFBFxCTHBF−•=(2.6)Top-HatTop-HatTop-HatTop-HatTop-HatTop-HatTop-Hat)(nnBi×)(mmBo×oiBB⊂ioBBA−=2.7Top-Hat))()(()(,xBAFFxTHiBF⊕Θ−=2.8[8]Top-Hat142.3,---80Top-HatTop-Hat9LkX,BkkY∑=−=LkkkdYLE12)(21(2.9)kdk0λkYTop-Hat−•−=)))(max(()))(