matlab图像处理 外文翻译 外文文献 英文文献 基于视觉的矿井救援Word格式文档下载.docx

matlab图像处理 外文翻译 外文文献 英文文献 基于视觉的矿井救援Word格式文档下载.docx

《matlab图像处理 外文翻译 外文文献 英文文献 基于视觉的矿井救援Word格式文档下载.docx》由会员分享，可在线阅读，更多相关《matlab图像处理 外文翻译 外文文献 英文文献 基于视觉的矿井救援Word格式文档下载.docx（14页珍藏版）》请在冰豆网上搜索。

canbeconvertedintotheevaluationproblemofHMM.Thecontributionsoftheseskills

makethesystemhavetheabilitytodealwithchangesinscale,2Drotationandviewpoint.

Theresultsofexperimentsalsoprovethatthesystemhashigherratioofrecognitionand

localizationinbothstaticanddynamicmineenvironments.

Keywords:

robotlocation;

scenerecognition;

salientimage;

matchingstrategy;

fuzzy

logic;

hiddenMarkovmodel

1Introduction

Searchandrescueindisasterareainthedomainofrobotisaburgeoningand

challengingsubject[1].Minerescuerobotwasdevelopedtoenterminesduring

emergenciestolocatepossibleescaperoutesforthosetrappedinsideanddetermine

whetheritissafeforhumantoenterornot.Localizationisafundamentalprobleminthis

field.Localizationmethodsbasedoncameracanbemainlyclassifiedintogeometric,

topologicalorhybridones[2].Withitsfeasibilityandeffectiveness,scenerecognition

becomesoneoftheimportanttechnologiesoftopologicallocalization.

Currentlymostscenerecognitionmethodsarebasedonglobalimagefeaturesand

havetwodistinctstages:

trainingofflineandmatchingonline.

Duringthetrainingstage,robotcollectstheimagesoftheenvironmentwhereit

worksandprocessestheimagestoextractglobalfeaturesthatrepresentthescene.Some

approacheswereusedtoanalyzethedata-setofimagedirectlyandsomeprimaryfeatures

werefound,suchasthePCAmethod[3].However,thePCAmethodisnoteffectivein

distinguishingtheclassesoffeatures.Anothertypeofapproachusesappearancefeatures

includingcolor,textureandedgedensitytorepresenttheimage.Forexample,ZHOUet

al[4]usedmultidimensionalhistogramstodescribeglobalappearancefeatures.This

methodissimplebutsensitivetoscaleandilluminationchanges.Infact,allkindsof

globalimagefeaturesaresufferedfromthechangeofenvironment.

LOWE[5]presentedaSIFTmethodthatusessimilarityinvariantdescriptorsformed

bycharacteristicscaleandorientationatinterestpointstoobtainthefeatures.The

featuresareinvarianttoimagescaling,translation,rotationandpartiallyinvariantto

illuminationchanges.ButSIFTmaygenerate1000ormoreinterestpoints,whichmay

slowdowntheprocessordramatically.

Duringthematchingstage,nearestneighborstrategy（NN）iswidelyadoptedforits

facilityandintelligibility[6].Butitcannotcapturethecontributionofindividualfeature

forscenerecognition.Inexperiments,theNNisnotgoodenoughtoexpressthe

similaritybetweentwopatterns.Furthermore,theselectedfeaturescannotrepresentthe

scenethoroughlyaccordingtothestate-of-artpatternrecognition,whichmakes

recognitionnotreliable[7].

Sointhisworkanewrecognitionsystemispresented,whichismorereliableand

effectiveifitisusedinacomplexmineenvironment.Inthissystem,weimprovethe

invariancebyextractingsalientlocalimageregionsaslandmarkstoreplacethewhole

imagetodealwithlargechangesinscale,2Drotationandviewpoint.Andthenumberof

interestpointsisreducedeffectively,whichmakestheprocessingeasier.Fuzzy

recognitionstrategyisdesignedtorecognizethelandmarksinplaceofNN,whichcan

strengthenthecontributionofindividualfeatureforscenerecognition.Becauseofits

partialinformationresumingability,hiddenMarkovmodelisadoptedtoorganizethose

landmarks,whichcancapturethestructureorrelationshipamongthem.Soscene

recognitioncanbetransformedtotheevaluationproblemofHMM,whichmakes

recognitionrobust.

2Salientlocalimageregionsdetection

Researchesonbiologicalvisionsystemindicatethatorganism（likedrosophila）often

paysattentiontocertainspecialregionsinthescenefortheirbehavioralrelevanceor

localimagecueswhileobservingsurroundings[8].Theseregionscanbetakenasnatural

landmarkstoeffectivelyrepresentanddistinguishdifferentenvironments.Inspiredby

those,weusecenter-surrounddifferencemethodtodetectsalientregionsinmulti-scale

imagespaces.Theopponenciesofcolorandtexturearecomputedtocreatethesaliency

map.

Follow-up,sub-imagecenteredatthesalientpositioninSistakenasthelandmark

region.Thesizeofthelandmarkregioncanbedecidedadaptivelyaccordingtothe

changesofgradientorientationofthelocalimage[11].

Mobilerobotnavigationrequiresthatnaturallandmarksshouldbedetectedstably

whenenvironmentschangetosomeextent.Tovalidatetherepeatabilityonlandmark

detectionofourapproach,wehavedonesomeexperimentsonthecasesofscale,2D

rotationandviewpointchangesetc.Fig.1showsthatthedoorisdetectedforitssaliency

whenviewpointchanges.Moredetailedanalysisandresultsaboutscaleandrotationcan

befoundinourpreviousworks[12].

3Scenerecognitionandlocalization

Differentfromotherscenerecognitionsystems,oursystemdoesn'

tneedtraining

offline.Inotherwords,ourscenesarenotclassifiedinadvance.Whenrobotwanders,

scenescapturedatintervalsoffixedtimeareusedtobuildthevertexofatopologicalmap,

whichrepresentstheplacewhererobotlocates.Althoughthemap'

sgeometriclayoutis

ignoredbythelocalizationsystem,itisusefulforvisualizationanddebugging[13]and

beneficialtopathplanning.Solocalizationmeanssearchingthebestmatchofcurrent

sceneonthemap.InthispaperhiddenMarkovmodelisusedtoorganizetheextracted

landmarksfromcurrentsceneandcreatethevertexoftopologicalmapforitspartial

informationresumingability.

Resembledbypanoramicvisionsystem,robotlooksaroundtogetomni-images.

From

Experimentonviewpointchanges

Fig.1

asnamedasequence,detectedandformedtobeareeachimage,salientlocalregions

hiddenThenaastheimagesequence.samelandmarksequencewhoseorderisthe

salientlocalimageMarkovmodeliscreatedbasedonthelandmarksequenceinvolvingkourIntherobotlocates.theasdescriptionoftheplacewherewhichregions,istaken

weeffect,overlap.Consideringthe170°

viewEVI-D70systemcamerahasafieldof±

toget8images.sampleenvironmentevery45°

≤8）,thecreatedHMMcanbeillustratedby（1≤SiiLetthe8imagesashiddenstate

areachievedbylearning,usingBaulm-WelchbjkaijFig.2.TheparametersofHMM,and

algorithm[14].Thethresholdofconvergenceissetas0.001.

Asfortheedgeoftopologicalmap,weassignitwithdistanceinformationbetween

twovertices.Thedistancescanbecomputedaccordingtoodometryreadings.

HMMofenvironment

Fig.2

Tolocateitselfonthetopologicalmap,robotmustrunits‘eye'

onenvironmentand

extractalandmarksequenceL1′?

Lk′,thensearchthemapforthebestmatchedvertex

（scene）.Differentfromtraditionalprobabilisticlocalization[15],inoursystem

localizationproblemcanbeconvertedtotheevaluationproblemofHMM.Thevertex

withthegreatestevaluationvalue,whichmustalsobegreaterthanathreshold,istaken

asthebestmatchedvertex,whichindicatesthemostpossibleplacewheretherobotis.

4Matchstrategybasedonfuzzylogic

Oneofthekeyissuesinimagematchproblemistochoosethemosteffectivefeatures

ordescriptorstorepresenttheoriginalimage.Duetorobotmovement,thoseextracted

landmarkregionswillchangeatpixellevel.So,thedescriptorsorfeatureschosenshould

beinvarianttosomeextentaccordingtothechangesofscale,rotationandviewpointetc.

Inthispaper,weuse4featurescommonlyadoptedinthecommunitythatarebriefly

describedasfollows.

GO:

Gradientorientation.Ithasbeenprovedthatilluminationandrotationchanges

arelikelytohavelessinfluenceonit[5].

ASMandENT:

Angularsecondmomentandentropy,whicharetwotexture

descriptors.

H:

Hue,whichisusedtodescribethefundamentalinformationoftheimage.

Anotherkeyissueinmatchproblemistochooseagoodmatchstrategyoralgorithm.

Usuallynearestneighborstrategy（NN）isusedtomeasurethesimilaritybetweentwo

patterns.ButwehavefoundintheexperimentsthatNNcan'

tadequatelyexhibitthe

individualdescriptororfeature'

scontributiontosimilaritymeasurement.Asindicatedin

Fig.4,theinputimageFig.4（a）comesfromdifferentviewofFig.4（b）.Butthedistance

betweenFigs.4（a）and（b）computedbyJeffereydivergenceislargerthanFig.4（c）.

Tosolvetheproblem,wedesignanewmatchalgorithmbasedonfuzzylogicfor

exhibitingthesubtlechangesofeachfeatures.Thealgorithmisdescribedasbelow.

Andthelandmarkinthedatabasewhosefusedsimilaritydegreeishigherthanany

othersistakenasthebestmatch.ThematchresultsofFigs.2（b）and（c）aredemonstrated

byFig.3.Asindicated,th