Pattern Recognition.docx
《Pattern Recognition.docx》由会员分享,可在线阅读,更多相关《Pattern Recognition.docx(35页珍藏版)》请在冰豆网上搜索。
PatternRecognition
PatternRecognition49(2016)102–114
YiminZhoua,n,1,GuolaiJianga,b,1,YaorongLinb
a
ShenzhenInstitutesofAdvancedTechnology,ChineseAcademyofSciences,China
bSchoolofElectronicandInformationEngineering,SouthChinaUniversityofTechnology,China
Articlehistory:
Received17March2014
Receivedinrevisedform
8August2014
Accepted29July2015
Availableonline8August2015
Thispaperpresentsahigh-levelhandfeatureextractionmethodforreal-timegesturerecognition.Firstly,thefingersaremodelledascylindricalobjectsduetotheirparalleledgefeature.Thenanovelalgorithmisproposedtodirectlyextractfingersfromsalienthandedges.Consideringthehandgeometricalcharacteristics,thehandpostureissegmentedanddescribedbasedonthefingerpositions,palmcenterlocationandwristposition.Aweightedradialprojectionalgorithmwiththeoriginatthe
article
info
abstract
Keywords:
Computervision
Fingermodelling
Salienthandedge
Convolutionoperator
Real-timehandgesturerecognition
wristpositionisappliedtolocalizeeachfinger.Thedevelopedsystemcannotonlyextractextensionalfingersbutalsoflexionalfingerswithhighaccuracy.Furthermore,handrotationandfingeranglevariationhavenoeffectonthealgorithmperformance.Theorientationofthegesturecanbecalculatedwithouttheaidofarmdirectionanditwouldnotbedisturbedbythebarearmarea.Experimentshavebeenperformedtodemonstratethattheproposedmethodcandirectlyextracthigh-levelhandfeatureandestimatehandposesinreal-time.
&2015ElsevierLtd.Allrightsreserved.
1.Introduction
Handgesturerecognitionbasedoncomputervisiontechnologyhasbeenreceivedgreatinterestsrecently,duetoitsnaturalhuman-computerinteractioncharacteristics.Handgesturesaregenerallycomposedofdifferenthandposturesandtheirmotions.However,humanhandisanarticulatedobjectwithover20degreesoffreedom(DOF)[12],andmanyself-occlusionswouldoccurinitsprojectionresults.Moreover,handmotionisoftentoofastandcomplicatedcomparedwithcurrentcomputerimageprocessingspeed.Therefore,real-timehandpostureestimationisstillachallengingresearchtopicwithmulti-disciplinaryworkincludingpatternrecognition,imageprocessing,computervision,artificialintelligenceandmachinelearning.
Inhuman–machineinteractionhistory,keyboardinput&charactertextoutputandmouseinput&graphicwindowdisplayaremaintraditionalinteractionforms.Withthedevelopmentofcomputertechniques,thehuman–machineinteractionviahandpostureplaysanimportantroleunderthreedimensionalvirtualenvironment.Manymethodshavebeendevelopedforhandposerecognition[3,4,10,18,24,29].
AgeneralframeworkforvisualbasedhandgesturerecognitionisillustratedinFig.1.Firstly,thehandislocatedandsegmentedfromtheinputimage,whichcanbeachievedviaskin-colorbasedsegmentationmethods[27,31]ordirectobjectrecognitionalgorithms.Thesecondstepistoextractusefulfeatureforstatichandpostureandmotionidentification.Thenthegesturecanbeidentifiedviafeaturematching.Finally,differenthumanmachineinteractioncanbeappliedbasedonthesuccessfulhandgesturerecognition.
Therearealotofconstraintsanddifficultiesinaccuratehandgesturerecognitionfromimagessincehumanhandisanobjectwithcomplexandversatileshapes[25].Firstly,differentfromlessremarkablemetamorphosisobjectssuchashumanface,humanhandpossessesover20freedegreeplusvariationsinhandgesturelocationandrotationwhichmakehandpostureestimationextremelydifficult.Evidenceshowsthatatleast6-dimensioninformationisrequiredforbasichandgestureestimation.Theocclusionalsocouldincreasethedifficultyinposerecognition.Sincetheinvolvedhandgestureimagesareusuallytwodimensionedimages,itwouldresultinocclusionofsomekeypartsofthehandontheplaneprojectduetovariousheightsofthehandshapes.
Besides,theimpactofthecomplexenvironmenttothebroadlyappliedvisual-basedhandgesturerecognitiontechniqueshasto
nCorrespondingauthor.
E-mailaddresses:
ym.zhou@(Y.Zhou),gl.jiang@(G.Jiang).
1firstauthorandsecondauthorcontributeequallyinthepaper.
The
beconsidered.Thelightnessvariationandcomplexbackgroundsuchfactorsmakeitmoredifficultforthehandgesturesegmentation.Uptonow,thereisnouniteddefinitionfordynamichand
http:
//dx.doi.org/10.1016/j.patcog.2015.07.0140031-3203/&2015ElsevierLtd.Allrightsreserved.
Fig.2.Handgesturemodelswithdifferentcomplexities(a)3Dstripmodel;(b)3Dsurfacemodel;(c)papermodel[36];(d)gesturesilhouette;and(e)gesturecontour.
gesturerecognition,whichisalsoanunsolvedproblemtoaccommodatehumanhabitsandfacilitatecomputerrecognition.Itshouldbenotedthathumanhandhasdeformableshapeinfrontofacameraduetoitsowncharacteristics.Theextractionofahandimagehastobeexecutedinreal-timeindependentoftheusersanddevice.Humanmotionpossessesafastspeedupto5m/sfortranslationand3001C/sforrotation.Thesamplingfrequencyofadigitalcameraisabout30–60Hz,whichcouldresultinfuzzificationonthecollectedimageswithnegativeimpactonfurtheridentification.Ontheotherhand,withthehandgesturemoduleaddedinthesystem,thedealtframenumberpersecondforthecomputerwillbeevenless,whichwillbringmoreseriouspressureontherelativelylowersamplingspeed.Moreover,alargeamountofdatahavetobedealtincomputervisualsystem,especiallyforhighcomplexversatileobjects.Undercurrentcomputerhardwareconditions,alotofhigh-precisionrecognitionalgorithmsaredifficulttobeoperatedinreal-time.
Ourdevelopedalgorithmfocusesonsinglecamerabasedrealtimehandgesturerecognition.Someassumptionsaremadewithoutlossofgenerality:
(a)thebackgroundisnottoocomplexwithoutlargeareaskincolordisturbance;(b)lightnessshouldavoidtoolowortoolightsuchworseconditions;(c)thepalmisrightfacedtothecamerawithdistanceintheranger0:
5m.Thesethreelimitationsarenotdifficulttoberealizedintheactualapplicationscenarios.
Firstly,anewfingerdetectionalgorithmisproposed.Comparedtopreviousfingerdetectionalgorithms,thedevelopedalgorithmisindependentofthefingertipfeaturebutcanextractfingersdirectlyfromthemainedgeofthewholefingers.Consideringthateachfingerhastwomain“parallel”edges,afingerisdeterminedfromconvolutionresultofsalienthandedgeimagewithaspecificoperatorG.Thealgorithmcannotonlyextractextensionalfingersbutalsoflexionalfingerswithhighaccuracy,whichisthebasisforcompletehandposehigh-levelfeatureextraction.Afterthefingercentralareahasbeenobtained,thecenter,orientationofthehandgesturecanbecalculated.Duringtheprocedure,anovelhigh-levelgesturefeatureextractionalgorithmisdeveloped.Throughweightedradiusprojectionalgorithm,thegesturefeaturesequencecanbeextractedandthefingerscanalsobelocalizedfromthelocalmaximaofangularprojection,thusthegesturecanbeestimateddirectlyinreal-time.
Theremainderofthepaperisorganizedasfollows.Section2describeshandgesturerecognitionprocedureandgenerallyusedmethods.FingerextractionalgorithmbasedonparalleledgecharacteristicsisintroducedinSection3.Salienthandimagecanalsobeachieved.ThespecificoperatorGandthresholdisexplainedindetailinSection4.High-levelhandfeatureextractionthroughconvolutionisdemonstratedinSection5.ExperimentsindifferentscenariosareperformedtoprovetheeffectivenessoftheproposedalgorithminSection6.ConclusionsandfutureworksaregiveninSection7.
2.Methodsofhandgesturerecognitionbasedoncomputer
vision
2.1.Handmodelling
Handposturemodellingplaysakeyroleinthewholehandgesturerecognitionsystem.Theselectionofthehandmodelisdependentontheactualapplicationenvironments.Thehandmodelcanbecategorizedasgestureappearancemodellingand3Dmodelling.GenerallyusedhandgesturemodelsaredemonstratedinFig.2.
3Dhandgesturemodelconsidersthegeometricalstructurewithhistogramorhyperquadricsurfacetoapproximatefingerjointsandpalm.Themodelparameterscanbeestimatedfromsingleimageorseveralimages.However,the3Dmodelbasedgesturemodellinghasquiteahighcalculationcomplexity,andtoomanylinearizationandapproximationwouldcauseunreliableparameterestimation.Asforappearancebasedgesturemodels,theyarebuiltthroughappearancecharacteristics,whichhavetheadvantagesoflesscomputationloadandfastprocessingspeed.Theadoptionofthesilhouette,contourmodelandpapermodelcanonlyreflectpartialhandgesturecharacteristics.Inthispaper,basedonthesimplifiedpapergesturemodel[36],anewgesturemodelisproposedwhereeachfingerisrepresentedbyextensionandflexionstatesconsideringgesturecompletenessandreal-timerecognitionrequirements.
Manyhandposerecognitionmethodsuseskincolor-baseddetectionandtakegeometricalfeaturesforhandmodelling.Handposeestimationfrom2Dto3Dusingmulti-viewpointsilhouetteimagesisdescribedin[35].Inrecentyears,3Dsensors,suchasbinocularcameras,Kinectandleapmotion,havebeenappliedforhandgesturerecognitionwithgoodperformance[5].However,handgesturerecognitionhasquitealimitation,since3Dsensorsarenotalwaysavailableinmanysystems,i.e.,GoogleGlasses.
2.2.Descriptionofhandgesturefeature
Thefeatureextraction