[雙語翻譯]手勢識別外文翻譯--光照不變的實時魯棒手勢識別（英文）

1、Optik159(2018)283–294ContentslistsavailableatScienceDirectOptikjournalhome page:www.elsevier.de/ijleoOriginalresearcharticleLightinvariantreal-timerobusthandgesturerecognitionAnkitChaudhary a,?, J.L.Raheja ba DataScienc

2、eDivision,SchoolofComputerScience,NorthwestMissouriStateUniversity,MO,USAb Cyber-PhysicalSystems,CEERI-CSIR,RJ,Indiaa r t i cl e i nfoArticlehistory:Received25August2016Accepted22November2017Keywords:Gesturerecogn

3、itionOrientationhistogramLightintensityinvariantsystemsExtremechangeinlightintensityNaturalcomputingRobustskindetectiona bs t r ac tComputervisionhas spread over differentdomainsto facilitatedifficultoperations.

4、It worksasthe artificialeye for manyindustrialapplicationsto observeelements,process, automa-tionand to find defects.Vision-basedsystemscan also be appliedto normalhumanlifeoperationsbut changinglight condi

5、tionsis a big problemfor thesesystems.Hand ges-turerecognitioncan be embeddedwith many existing interactiveapplications/gamestomakeinteractionnaturaland easy but changingilluminationand non-uniformbackground

6、smakeit verydifficultto perform operationswith goodimage segmentation.Ifa visionbasedsystemisinstalledin publicdomain,different peoplearesupposedto work on theapplication.Thispaper demonstratesa light intens

7、ityinvarianttechniquefor hand gesture recog-nitionwhich can be easily appliedto othervision-basedapplicationsalso. The techniquehasbeen tested on differentpeople in differentlight conditionswith the extr

8、emechangeinintensity.This was done as one skin colorlooksdifferentin changedlightintensityanddifferentskin colors maylook same in changedlight intensity.Theorientationhistogramwasused to identifyuniquefeatu

9、res ofa hand gestureand itwas comparedusing super-visedANN. The overall accuracyof 92.86%is achievedin extremelight intensitychangingenvironments.©2017 ElsevierGmbH.All rightsreserved.1.IntroductionCo

10、mputervisionapplicationshavebeenpartofindustryoperationsformorethanfourdecades.Theyarehelpfulinfastingtheindustrialprocess,automatedmanydifficulttasksandalsohelpinfindingminordefects[1].Manyapplicationswereusinghandgestu

11、rerecognitiontechniquesfordifferentpurposesashandgestureprovidesanaturalwaytocommunicatewithmachines[2–5].Theseapplicationswereinitiallybasedonwiredgloves,colorstripsorchemicalstodetectaregionofinterest(ROI)smoothly.Asur

12、veyofdifferentdevicesandtechniquesusedforhandgesturerecognitioncanbefoundoutin[6].Tomakehuman-machinecommunicationmoreeffective,gesturerecognitionofbarehandhadintroducedwhereanypersoncouldusehishandinnaturalposition[7–10

13、].Alotofworkhasbeendoneintheareaofnaturalhandgesturerecognitiontomakeitmorerobust.Currently,thiskindofapplications[11]andgamesaremorepopularasauserfeelcomfortableanddon’tneedanythingtooperatethevision-basedsystem.Recentl

14、ytherehasbeenagrowinginterestinthefieldoflightintensity-invariantobjectrecognition.Foradvancedapplicationsinthisarea,onecansetupasysteminthelaboratorywithidealconditions.However,inpracticalscenarios,the? Correspondingaut

15、hor.E-mailaddresses:dr.ankit@ieee.org(A.Chaudhary),jagdish@ceeri.res.in(J.L.Raheja).https://doi.org/10.1016/j.ijleo.2017.11.1580030-4026/©2017ElsevierGmbH.Allrightsreserved.A.Chaudhary,J.L.Raheja/Optik159(2018)283–2

16、94285OrientationHistogram(OH)techniqueforfeatureextractionwasdevelopedbyMcConnell[23].Themajoradvantageofthistechniqueisthatitissimpleandrobusttolightingchanges[24].Ifwefollowpixel-intensitiesapproach,certainproblemsaris

17、eduetovaryingillumination[16].Ifpixelbypixelproximityforthesamegestureistakenfromtwodifferentimages,whiletheilluminationconditionsaredifferent,thedistancebetweenthemwouldbelarge.Insuchscenarios,thepictureitselfactsasafea

18、turevector.Themainmotivationforusingtheorientationhistogramistherequirementforlightningandpositioninvariance.Anotherimportantaspectofthegesturerecognitionisthatirrespectiveoftheorientationofthehandindifferentimages,forth

19、esamegesturewemustgetthesameoutput.Thiscanbedonebyformingalocalhistogramforlocalorientations[25].Hence,thisapproachmustberobustforilluminationchangesanditmustalsooffertranslationalinvariance.Wewouldalsoneedthegesturestob

20、ethesameregardlessofwheretheyoccurintheimage.Thepixellevelsofthehandwouldvaryconsiderablywithrespecttolight,ontheotherhand,theorientationvaluesremainfairlyconstant.We needtocalculatethelocalorientationfromthedirectionof

21、theimagegradient.Thelocalorientationangle?willbeafunctionofpositionxandy,andtheimageintensitiesI(x,y).Theangle?isdefinedas:?(x,y)=arctan[I(x,y)-I(x-1, y),I(x, y)-I(x,y-1)](1)NowformavectorФofNelements,withtheith elemen

22、tshowingthenumberoforientationelements?(x,y)betweentheangles 360 ?N [i? 12]and 360 ?N [i+ 12 ].WhereФisdefinedas:Ф(i)=?x,y{ 10 if|?(x,y)? 360 ?N i|< 360 ?N otherwise (2)3.LightinvariantsystemThehandgesturerecognitio

23、nsystemworksontheprincipleofthe2Dcomputervision.Thesystemhasaninterfacewithasmallcamerawhichcapturesusers’gestures.Theinputtothesystemisimageframeofmovinghandinfrontofacameracapturedasalivevideo.Thepreprocessingofimagefr

24、amewasdoneasdiscussedin[26]withreal-timeconstraint.TheresultingimagewouldbetheROI,onlyhandgestureimage.Nowwedoneedtofindoutfeaturevectorsfromtheinputimagetorecognizeitwiththehelpofclassifier.Asthissystemwasforresearchpur

25、poseonly,we tookonlysixdifferentgesturesinthedatasetasmanyresearchersalsohavetestedtheirmethodswithsixgesturesinthepast[21].Thesesixdifferentgestureswhichwereusedinthisresearch,areshowninFig.2.Thesystemisexpandabletohav

26、emanydifferenttypesofgestures,ifneeded.Theimagesofeachgesturewerecollectedwithdifferentskincolorandlightintensity.Oncethegesturewouldgetrecognizedthecorrespondingactiontakesplacewhichwasassociatedwithit.Inoursystem,theau

27、diodescriptionofthematchedgesturewasattachedasthecorrespondingaction.Inrecognitionofthegesture,theaudiofilecorrespondingtotherecognizedgesturewouldbeplayed.Theimplementationofthesystemisdiscussedindifferentsteps:3.1.Data

28、collectionfortrainingpurposeThetrainingimagesfortheANNwerecollectedfromdifferentsourcesincludingonlinesearchandmanuallycollection.Thiswastoensuretherobustnessofthemethodasimagesfromdifferentsourceswouldcontaindifferentsk

29、incolor,differentlightintensity,anddifferenthandshape.Skincolorhasthepropertythatitlooksdifferentindifferentlightintensities.Weused14differentimagesforeachgesturetotrainANN.3.2.Pre-processingofimagesWe needtogettheROIfr

30、omtheimageswiththerandombackgroundforthetrainingpurposeandfortherecognition.IftheimageshaveonlyROIthenthetrainingoftheANNwouldbebetter.Allimages,usedfortraining,wereconvertedintosameresolutionasthesystemcamerawascapturin

31、gtheuser’sgesture[26].3.3.FeatureextractionTotraintheANNandforgesturerecognition,thefeaturesneedtobeextractedfromthepre-processedimages.Thealgorithmusedforfeatureextractionresultsinanorientationhistogramforagivengesture.

32、Thesamealgorithmwasappliedforallthegesturespresentinthedatabaseinordertogenerateatrainingpattern.Thesetrainingpatternswerestoredandappliedtotheneuralnetworktotrainit.Forgesturerecognitionpurposethesamealgorithmwasapplied