外文文献及翻译基于视觉的矿井救援机器人场景识别Scene recognition for mine rescue robotlocalization.docx
《外文文献及翻译基于视觉的矿井救援机器人场景识别Scene recognition for mine rescue robotlocalization.docx》由会员分享,可在线阅读,更多相关《外文文献及翻译基于视觉的矿井救援机器人场景识别Scene recognition for mine rescue robotlocalization.docx(10页珍藏版)》请在冰豆网上搜索。
![外文文献及翻译基于视觉的矿井救援机器人场景识别Scene recognition for mine rescue robotlocalization.docx](https://file1.bdocx.com/fileroot1/2023-1/1/daf6755b-cd1b-49f6-8930-9b77cfa4c56f/daf6755b-cd1b-49f6-8930-9b77cfa4c56f1.gif)
外文文献及翻译基于视觉的矿井救援机器人场景识别Scenerecognitionforminerescuerobotlocalization
附录英文原文
Scenerecognitionforminerescuerobot
localizationbasedonvision
Abstract:
AnewscenerecognitionsystemwaspresentedbasedonfuzzylogicandhiddenMarkovmodel(HMM)thatcanbeappliedinminerescuerobotlocalizationduringemergencies.Thesystemusesmonocularcameratoacquireomni-directionalimagesofthemineenvironmentwheretherobotlocates.Byadoptingcenter-surrounddifferencemethod,thesalientlocalimageregionsareextractedfromtheimagesasnaturallandmarks.TheselandmarksareorganizedbyusingHMMtorepresentthescenewheretherobotis,andfuzzylogicstrategyisusedtomatchthesceneandlandmark.Bythisway,thelocalizationproblem,whichisthescenerecognitionprobleminthesystem,canbeconvertedintotheevaluationproblemofHMM.Thecontributionsoftheseskillsmakethesystemhavetheabilitytodealwithchangesinscale,2Drotationandviewpoint.Theresultsofexperimentsalsoprovethatthesystemhashigherratioofrecognitionandlocalizationinbothstaticanddynamicmineenvironments.
Keywords:
robotlocation;scenerecognition;salientimage;matchingstrategy;fuzzylogic;hiddenMarkovmodel
1Introduction
Searchandrescueindisasterareainthedomainofrobotisaburgeoningandchallengingsubject[1].Minerescuerobotwasdevelopedtoenterminesduringemergenciestolocatepossibleescaperoutesforthosetrappedinsideanddeterminewhetheritissafeforhumantoenterornot.Localizationisafundamentalprobleminthisfield.Localizationmethodsbasedoncameracanbemainlyclassifiedintogeometric,topologicalorhybridones[2].Withitsfeasibilityandeffectiveness,scenerecognitionbecomesoneoftheimportanttechnologiesoftopologicallocalization.
Currentlymostscenerecognitionmethodsarebasedonglobalimagefeaturesandhavetwodistinctstages:
trainingofflineandmatchingonline.
Duringthetrainingstage,robotcollectstheimagesoftheenvironmentwhereitworksandprocessestheimagestoextractglobalfeaturesthatrepresentthescene.Someapproacheswereusedtoanalyzethedata-setofimagedirectlyandsomeprimaryfeatureswerefound,suchasthePCAmethod[3].However,thePCAmethodisnoteffectiveindistinguishingtheclassesoffeatures.Anothertypeofapproachusesappearancefeaturesincludingcolor,textureandedgedensitytorepresenttheimage.Forexample,ZHOUetal[4]usedmultidimensionalhistogramstodescribeglobalappearancefeatures.Thismethodissimplebutsensitivetoscaleandilluminationchanges.Infact,allkindsofglobalimagefeaturesaresufferedfromthechangeofenvironment.
LOWE[5]presentedaSIFTmethodthatusessimilarityinvariantdescriptorsformedbycharacteristicscaleandorientationatinterestpointstoobtainthefeatures.Thefeaturesareinvarianttoimagescaling,translation,rotationandpartiallyinvarianttoilluminationchanges.ButSIFTmaygenerate1000ormoreinterestpoints,whichmayslowdowntheprocessordramatically.
Duringthematchingstage,nearestneighborstrategy(NN)iswidelyadoptedforitsfacilityandintelligibility[6].Butitcannotcapturethecontributionofindividualfeatureforscenerecognition.Inexperiments,theNNisnotgoodenoughtoexpressthesimilaritybetweentwopatterns.Furthermore,theselectedfeaturescannotrepresentthescenethoroughlyaccordingtothestate-of-artpatternrecognition,whichmakesrecognitionnotreliable[7].
Sointhisworkanewrecognitionsystemispresented,whichismorereliableandeffectiveifitisusedinacomplexmineenvironment.Inthissystem,weimprovetheinvariancebyextractingsalientlocalimageregionsaslandmarkstoreplacethewholeimagetodealwithlargechangesinscale,2Drotationandviewpoint.Andthenumberofinterestpointsisreducedeffectively,whichmakestheprocessingeasier.FuzzyrecognitionstrategyisdesignedtorecognizethelandmarksinplaceofNN,whichcanstrengthenthecontributionofindividualfeatureforscenerecognition.Becauseofitspartialinformationresumingability,hiddenMarkovmodelisadoptedtoorganizethoselandmarks,whichcancapturethestructureorrelationshipamongthem.SoscenerecognitioncanbetransformedtotheevaluationproblemofHMM,whichmakesrecognitionrobust.
2Salientlocalimageregionsdetection
Researchesonbiologicalvisionsystemindicatethatorganism(likedrosophila)oftenpaysattentiontocertainspecialregionsinthescenefortheirbehavioralrelevanceorlocalimagecueswhileobservingsurroundings[8].Theseregionscanbetakenasnaturallandmarkstoeffectivelyrepresentanddistinguishdifferentenvironments.Inspiredbythose,weusecenter-surrounddifferencemethodtodetectsalientregionsinmulti-scaleimagespaces.Theopponenciesofcolorandtexturearecomputedtocreatethesaliencymap.
Follow-up,sub-imagecenteredatthesalientpositioninSistakenasthelandmarkregion.Thesizeofthelandmarkregioncanbedecidedadaptivelyaccordingtothechangesofgradientorientationofthelocalimage[11].
Mobilerobotnavigationrequiresthatnaturallandmarksshouldbedetectedstablywhenenvironmentschangetosomeextent.Tovalidatetherepeatabilityonlandmarkdetectionofourapproach,wehavedonesomeexperimentsonthecasesofscale,2Drotationandviewpointchangesetc.Fig.1showsthatthedoorisdetectedforitssaliencywhenviewpointchanges.Moredetailedanalysisandresultsaboutscaleandrotationcanbefoundinourpreviousworks[12].
3Scenerecognitionandlocalization
Differentfromotherscenerecognitionsystems,oursystemdoesn’tneedtrainingoffline.Inotherwords,ourscenesarenotclassifiedinadvance.Whenrobotwanders,scenescapturedatintervalsoffixedtimeareusedtobuildthevertexofatopologicalmap,whichrepresentstheplacewhererobotlocates.Althoughthemap’sgeometriclayoutisignoredbythelocalizationsystem,itisusefulforvisualizationanddebugging[13]andbeneficialtopathplanning.Solocalizationmeanssearchingthebestmatchofcurrentsceneonthemap.InthispaperhiddenMarkovmodelisusedtoorganizetheextractedlandmarksfromcurrentsceneandcreatethevertexoftopologicalmapforitspartialinformationresumingability.
Resembledbypanoramicvisionsystem,robotlooksaroundtogetomni-images.From
Fig.1Experimentonviewpointchanges
eachimage,salientlocalregionsaredetectedandformedtobeasequence,namedaslandmarksequencewhoseorderisthesameastheimagesequence.ThenahiddenMarkovmodeliscreatedbasedonthelandmarksequenceinvolvingksalientlocalimageregions,whichistakenasthedescriptionoftheplacewheretherobotlocates.InoursystemEVI-D70camerahasaviewfieldof±170°.Consideringtheoverlapeffect,wesampleenvironmentevery45°toget8images.
Letthe8imagesashiddenstateSi(1≤i≤8),thecreatedHMMcanbeillustratedbyFig.2.TheparametersofHMM,aijandbjk,areachievedbylearning,usingBaulm-Welchalgorithm[14].Thethresholdofconvergenceissetas0.001.
Asfortheedgeoftopologicalmap,weassignitwithdistanceinformationbetweentwovertices.Thedistancescanbecomputedaccordingtoodometryreadings.
Fig.2HMMofenvironment
Tolocateitselfonthetopologicalmap,robotmustrunits‘eye’onenvironmentandextractalandmarksequenceL1′−Lk′,thensearchthemapforthebestmatchedvertex(scene).Differentfromtraditionalprobabilisticlocalization[15],inoursystemlocalizationproblemcanbeconvertedtotheevaluationproblemofHMM.Thevertexwiththegreatestevaluationvalue,whichmustalsobegreaterthanathreshold,istakenasthebestmatchedvertex,whichindicatesthemostpossibleplacewheretherobotis.
4Matchstrategybasedonfuzzylogic
Oneofthekeyissuesinimagematchproblemistochoosethemosteffectivefeaturesordescriptorstorepresenttheoriginalimage.Duetorobotmovement,thoseextractedlandmarkregionswillchangeatpixellevel.So,thedescriptorsorfeatureschosenshouldbeinvarianttosomeextentaccordingtothechangesofscale,rotationandviewpointetc.Inthispaper,weuse4featurescommonlyadoptedinthecommunitythatarebrieflydescribedasfollows.
GO:
Gradientorientation.Ithasbeenprovedthatilluminationandrotationchangesarelikelytohavelessinfluenceonit[5].
ASMandENT:
Angularsecondmomentandentropy,whicharetwotexturedescriptors.
H:
Hue,whichisusedtodescribethefundamentalinformationoftheimage.
Anotherkeyissueinmatchproblemistochooseagoodmatchstrategyoralgorithm.Usuallynearestneighborstrategy(NN)isusedtomeasurethesimilaritybetweentwopatterns.ButwehavefoundintheexperimentsthatNNcan’tadequatelyexhibittheindividualdescriptororfeature’scontributiontosimilaritymeasurement.AsindicatedinFig.4,theinputimageFig.4(a)comesfromdifferentviewofFig.4(b).ButthedistancebetweenFigs.4(a)and(b)computedbyJeffereydivergenceislargerthanFig.4(c).
Tosolvetheproblem,wedesignanewmatchalgorithmbasedonfuzzylogicforexhibitingthesubtlechangesofeachfeatures.Thealgorithmisdescribedasbelow.
Andthelandmarkinthedatabasewhosefusedsimilaritydegreeishigherthananyothersistakenasthebestmatch.ThematchresultsofFigs.2(b)and(c)aredemonstratedbyFig.3.Asindicated,thismethodcanmeasurethesimilarityeffectivelybetweentwopatterns.
Fig.3Similaritycomputedusingfuzzystrategy
5Experimentsandanalysis
Thelocalizationsystemhasbeenimplementedonamobilerobot,whichisbuiltbyourlaboratory.ThevisionsystemiscomposedofaC