




版權(quán)說明:本文檔由用戶提供并上傳,收益歸屬內(nèi)容提供方,若內(nèi)容存在侵權(quán),請進行舉報或認領(lǐng)
文檔簡介
附錄英文原文SeenerecognitionforminerescuerobotocalizationbasedonvisionAbstract:AnewscenerecognitionsystemwaspresentedbasedonfuzzylogicandhiddenMarkovmode1(HMM)thatcanbeapp1iedinminerescuerobotlocalizationduringemergencies.Thesystemusesmonocu1arcameratoacquireomni—directionalimagesofthemineenvironmentwheretherobotlocates.Byadoptingcenter-surrounddifferencemethod,thesalientloca1imageregionsareextractedfromtheimagesasnaturallandmarks.TheselandmarksareorganizedbyusingHMMtorepresentthescenewheretherobotis,andfuzzylogicstrategyisusedtomatchthesceneand1andmark.Bythisway,theloca1izationproblem,whichisthescenerecognitionprob1eminthesystem,canbeconvertedintotheevaluationproblemofHMM.Thecontributionsoftheseski1lsmakethesystemhavetheabilitytodealwithchangesinscale,2Drotationandviewpoint.Theresultsofexperimentsa1soprovethatthesystemhashigherratioofrecognitionandloca1izationinbothstaticanddynamicmineenvironments.Keywords:robotlocation;scenerecognition;salientimage;matchingstrategy;fuzzy1ogic;hiddenMarkovmode11IntroductionSearchandrescueindisasterareainthedomainofrobotisaburgeoningandcha11engingsubject[1].Minerescuerobotwasdevelopedtoenterminesduringemergenciestolocatepossibleescaperoutesforthosetrappedinsideanddeterminewhetheritissafeforhumantoenterornot.Localizationisafundamentalprobleminthisfield.Localizationmethodsbasedoncameracanbemainlyclassifiedintogeometric,topologicalorhybridones[2].Withitsfeasibilityandeffectiveness,scenerecognitionbecomesoneoftheimportanttechnologiesoftopologicallocalization.Current1ymostscenerecognitionmethodsarebasedongloba1imagefeaturesandhavetwodistinctstages:trainingoff1ineandmatchingon1ine.Duringthetrainingstage,robotcollectstheimagesoftheenvironmentwhereitworksandprocessestheimagestoextractg1oba1featuresthatrepresentthescene.Someapproacheswereusedtoanalyzethedata-setofimagedirect1yandsomeprimaryfeatureswerefound,suchasthePCAmethod[3].However,thePCAmethodisnoteffectiveindistinguishingtheclassesoffeatures.Anothertypeofapproachusesappearancefeaturesincludingco1or,textureandedgedensitytorepresenttheimage.Forexample,ZHOUetal[4]usedmultidimensionalhistogramstodescribeg1obalappearancefeatures.Thismethodissimplebutsensitivetosca1eandi1luminationchanges.Infact,allkindsofglobalimagefeaturesaresufferedfromthechangeofenvironment.LOWE[5]presentedaSIFTmethodthatusessimilarityinvariantdescriptorsformedbycharacteristicsealeandorientationatinterestpointstoobtainthefeatures.Thefeaturesareinvarianttoimagescaling,translation,rotationandpartia1lyinvarianttoilluminationchanges.ButSIFTmaygenerate1000ormoreinterestpoints,whichmayslowdowntheprocessordramatically.Duringthematchingstage,nearestneighborstrategy(NN)iswide1yadoptedforitsfaci1ityandintel1igibility[6].Butitcannotcapturethecontributionofindividualfeatureforscenerecognition.Inexperiments,theNNisnotgoodenoughtoexpressthesimi1aritybetweentwopatterns.Furthermore,theselectedfeaturescannotrepresentthescenethoroughlyaccordingtothestate-of-artpatternrecognition,whichmakesrecognitionnotreliab1e[7].Sointhisworkanewrecognitionsystemispresented,whichismorereliableandeffectiveifitisusedinacomplexmineenvironment.Inthissystem,weimprovetheinvariancebyextractingsalientlocalimageregionsas1andmarkstoreplacethewholeimagetodealwithlargechangesinscale,2Drotationandviewpoint.Andthenumberofinterestpointsisreducedeffectively,whichmakestheprocessingeasier.FuzzyrecognitionstrategyisdesignedtorecognizethelandmarksinplaceofNN,whichcanstrengthenthecontributionofindividualfeatureforscenerecognition.Becauseofitspartialinformationresumingability,hiddenMarkovmodelisadoptedtoorganizethose1andmarks,whichcancapturethestructureorrelationshipamongthem.Soscenerecognitioncanbetransformedtotheevaluationprob1emofHMM,whichmakesrecognitionrobust.Salient1ocalimageregionsdetectionResearchesonbio1ogicalvisionsystemindicatethatorganism(likedrosophi1a)oftenpaysattentiontocertainspecialregionsinthescenefortheirbehavioralre1evanceorlocalimagecueswhileobservingsurroundings[8].Theseregionscanbetakenasnaturallandmarkstoeffectivelyrepresentanddistinguishdifferentenvironments.Inspiredbythose,weusecenter-surrounddifferencemethodtodetectsalientregionsinmulti—scaleimagespaces.Theopponenciesofcolorandtexturearecomputedtocreatethesaliencymap.Follow-up,sub-imagecenteredatthesalientpositioninSistakenasthelandmarkregion.Thesizeofthelandmarkregioncanbedecidedadaptivelyaccordingtothechangesofgradientorientationofthelocalimage[11].Mobilerobotnavigationrequiresthatnaturallandmarksshouldbedetectedstablywhenenvironmentschangetosomeextent.Tovalidatetherepeatabilityonlandmarkdetectionofourapproach,wehavedonesomeexperimentsonthecasesofscale,2Drotationandviewpointchangesetc.Fig.1showsthatthedoorisdetectedforitssaliencywhenviewpointchanges.Moredetailedanalysisandresultsaboutscaleandrotationcanbefoundinourpreviousworks[12].ScenerecognitionandlocalizationDifferentfromotherscenerecognitionsystems,oursystemdoesn'tneedtrainingoffline.Inotherwords,ourscenesarenotclassifiedinadvance.Whenrobotwanders,scenescapturedatintervalsoffixedtimeareusedtobuildthevertexofatopologicalmap,whichrepresentstheplacewhererobotlocates.Althoughthemap'sgeometriclayoutisignoredbythe1oca1izationsystem,itisusefulforvisualizationanddebugging[13]andbeneficialtopathplanning.Solocalizationmeanssearchingthebestmatchofcurrentsceneonthemap.InthispaperhiddenMarkovmodelisusedtoorganizetheextractedlandmarksfromcurrentsceneandcreatethevertexoftopologicalmapforitspartialinformationresumingabi1ity.Resembledbypanoramicvisionsystem,robotlooksaroundtogetomni-images.FromFig?1Experimentonviewpointchangeseachimage,salientlocalregionsaredetectedandformedtobeasequence,namedaslandmarksequencewhoseorderisthesameastheimagesequence.ThenahiddenMarkovmode1iscreatedbasedonthelandmarksequenceinvolvingksalientlocalimageregions,whichistakenasthedescriptionoftheplacewheretherobotlocates.InoursystemEVI-D70camerahasaviewfieldof±170°.Consideringtheoverlapeffect,wesample
environmentevery45°toget8images.Letthe8imagesashiddenstateSi(1<i<8),thecreatedHMMcanbeillustratedbyFig.2.TheparametersofHMM,aijandbjk,areachievedbylearning,usingBaulm-WelchaIgorithm[14].Thethresholdofconvergenceissetas0.001.Asfortheedgeoftopologicalmap,weassignitwithdistanceinformationbetweentwovertices.Thedistancescanbecomputedaccordingtoodometryreadings.Fig.2HMFig.2HMMofenvironmentTolocateitselfonthetopo1ogicalmap,robotmustrunits‘eye'onenvironmentandextractalandmarksequenceLI'—Lk',thensearchthemapforthebestmatchedvertex(seene).Differentfromtraditionalprobabilisticlocalization[15],inoursystemlocalizationproblemcanbeconvertedtotheeva1uationproblemofHMM.Thevertexwiththegreatesteva1uationva1ue,whichmustalsobegreaterthanathreshold,istakenasthebestmatchedvertex,whichindicatesthemostpossibleplacewheretherobotis.Matchstrategybasedonfuzzy1ogicOneofthekeyissuesinimagematchproblemistochoosethemosteffectivefeaturesordescriptorstorepresenttheoriginalimage.Duetorobotmovement,thoseextractedlandmarkregionswillchangeatpixelleve1.So,thedescriptorsorfeatureschosenshouldbeinvarianttosomeextentaccordingtothechangesofscale,rotationandviewpointetc.Inthispaper,weuse4featurescommonlyadoptedinthecommunitythatarebrieflydescribedasfollows.GO:Gradientorientation.Ithasbeenprovedthati1luminationandrotationchangesarelikelytohave1essinfluenceonit[5].ASMandENT:Angularsecondmomentandentropy,whicharetwotexturedescriptors.H:Hue,whichisusedtodescribethefundamentalinformationoftheimage.Anotherkeyissueinmatchproblemistochooseagoodmatchstrategyora1gorithm.Usuallynearestneighborstrategy(NN)isusedtomeasurethesimilaritybetweentwopatterns.ButwehavefoundintheexperimentsthatNNcan'tadequatelyexhibittheindividualdescriptororfeature'scontributiontosimilaritymeasurement.AsindicatedinFig.4,theinputimageFig.4(a)comesfromdifferentviewofFig.4(b).ButthedistancebetweenFigs.4(a)and(b)computedbyJeffereydivergenceislargerthanFig.4(c).Toso1vetheproblem,wedesignanewmatchalgorithmbasedonfuzzylogicforexhibitingthesubt1echangesofeachfeatures.Thealgorithmisdescribedasbelow.Andthelandmarkinthedatabasewhosefusedsimilaritydegreeishigherthananyothersistakenasthebestmatch.Thematchresu1tsofFigs.2(b)and(c)aredemonstratedbyFig.3.Asindicated,thismethodcanmeasurethesimilarityeffectivelybetweentwopatterns.Fig.3Similaritycomputedusingfuzzystrategy5ExperimentsandanalysisThe1ocalizationsystemhasbeenimplementedonamobilerobot,whichisbui1tbyourlaboratory.ThevisionsystemiscomposedofaCCDcameraandaframe-grabberIVC-4200.Thereso1utionofimageissettobe400x320andthesamplefrequencyissettobe10frames/s.Thecomputersystemiscomposedof1GHzprocessorand512Mmemory,whichiscarriedbytherobot.Presentlytherobotworksinindoorenvironments.BecauseHMMisadoptedtorepresentandrecognizethescene,oursystemhastheabilitytocapturethediscriminationaboutdistributionofsalientlocalimageregionsanddistinguishsimilarsceneseffectively.Table1showstherecognitionresultofstaticenvironmentsincluding51anewaysandasilo.10scenesareselectedfromeachenvironmentandHMMsarecreatedforeachscene.Then20scenesarecol1ectedwhentherobotenterseachenvironmentsubsequentlytomatchthe60HMMsabove.Inthetable,“truth”meansthatthescenetobelocalizedmateheswiththerightscene(theevaluationvalueofHMMis30%greaterthanthesecondhighevaluation).“Uncertainty”meansTOC\o"1-5"\h\zthattheevaluationvalueofHMMisgreaterthanthesecondhighevaluationunder10%.“Errormatch”meansthatthescenetobelocalizedmatcheswiththewrongscene.Inthetable,theratiooferrormatchis0.Butitispossiblethatthescenetobelocalizedcan'tmatchanyscenesandnewvertexesarecreated.Furthermore,the“ratiooftruth”aboutsi1oislowerbecausesalientcuesarefewerinthiskindofenvironment.Intheperiodofautomaticexploring,simi1arscenescanbecombined.Theprocesscanbesummarizedas:whenloca1izationsucceeds,thecurrentlandmarksequenceisaddedtotheaccompanyingobservationsequenceofthematchedvertexun-repeatedlyaccordingtotheirorientation(includingtheangleoftheimagefromwhichthesalientlocalregionandtheheadingoftherobotcome).TheparametersofHMMarelearnedagain.Comparedwiththeapproachesusingappearancefeaturesofthewho1eimage(Method2,M2),oursystem(M1)useslocalsalientregionstolocalizeandmap,whichmakesithavemoretoleranceofscale,viewpointchangescausedbyrobot'smovementandhigherratioofrecognitionandfeweramountofverticesonthetopologicalmap.So,oursystemhasbetterperformanceindynamicenvironment.ThesecanbeseeninTable2.Laneways1,2,4,5areinoperationwheresomeminersareworking,whichpuzzletherobot.6Conclusions1)Salientloca1imagefeaturesareextractedtoreplacethewhol??????q?q?*eimagetoparticipatemrecognition,whichimprovetheto1eranceofchangesinscale,2Drotationandviewpointofenvironmentimage.2)Fuzzylogicisusedtorecognizethelocalimage,andemphasizetheindividualfeature'contributiontorecognition,whichimprovesthereliabilityoflandmarks.HMMisusedtocapturethestructureorrelationshipofthoselocalimages,whichconvertsthescenerecognitionproblemintotheevaluationproblemofHMM.Theresultsfromtheaboveexperimentsde/r
溫馨提示
- 1. 本站所有資源如無特殊說明,都需要本地電腦安裝OFFICE2007和PDF閱讀器。圖紙軟件為CAD,CAXA,PROE,UG,SolidWorks等.壓縮文件請下載最新的WinRAR軟件解壓。
- 2. 本站的文檔不包含任何第三方提供的附件圖紙等,如果需要附件,請聯(lián)系上傳者。文件的所有權(quán)益歸上傳用戶所有。
- 3. 本站RAR壓縮包中若帶圖紙,網(wǎng)頁內(nèi)容里面會有圖紙預覽,若沒有圖紙預覽就沒有圖紙。
- 4. 未經(jīng)權(quán)益所有人同意不得將文件中的內(nèi)容挪作商業(yè)或盈利用途。
- 5. 人人文庫網(wǎng)僅提供信息存儲空間,僅對用戶上傳內(nèi)容的表現(xiàn)方式做保護處理,對用戶上傳分享的文檔內(nèi)容本身不做任何修改或編輯,并不能對任何下載內(nèi)容負責。
- 6. 下載文件中如有侵權(quán)或不適當內(nèi)容,請與我們聯(lián)系,我們立即糾正。
- 7. 本站不保證下載資源的準確性、安全性和完整性, 同時也不承擔用戶因使用這些下載資源對自己和他人造成任何形式的傷害或損失。
最新文檔
- 2025年廣州貨運從業(yè)資格證網(wǎng)上考試題庫及答案
- 利用志愿服務活動推動勞動教育的實踐研究
- 人力資源管理招聘與選拔實務測試題
- ××超市打印設備辦法
- ××中學訴訟管理制度
- 2025年運動場館燈具項目規(guī)劃申請報告
- 2025年公路養(yǎng)護檢測設備項目申請報告
- 2025年觀光型酒店項目提案報告模板
- 醫(yī)學微生物學案例分析題集
- 業(yè)務合作協(xié)議及其合規(guī)責任承諾約定
- 2025年華僑港澳臺學生聯(lián)招考試英語試卷試題(含答案詳解)
- ASTM-D3359-(附著力測試標準)-中文版
- 兒童口腔科接診基本流程綱要
- 公司工作流程圖
- GB∕T 34876-2017 真空技術(shù) 真空計 與標準真空計直接比較校準結(jié)果的不確定度評定
- (完整版)管理經(jīng)濟學題庫
- 車工技師論文 細長軸的加工技術(shù)方法
- 修訂版《語言學綱要》(修訂版)學習指導書練習參考答案(完整)
- BP-2B微機母線保護裝置技術(shù)說明書V1
- 普通高等學校本科專業(yè)目錄(2015年版)
- 供應鏈管理調(diào)研報告
評論
0/150
提交評論