外文翻译利用手腕麦克风和三轴加速计来进行手势定位.docx

外文翻译利用手腕麦克风和三轴加速计来进行手势定位.docx

《外文翻译利用手腕麦克风和三轴加速计来进行手势定位.docx》由会员分享，可在线阅读，更多相关《外文翻译利用手腕麦克风和三轴加速计来进行手势定位.docx（14页珍藏版）》请在冰豆网上搜索。

外文翻译利用手腕麦克风和三轴加速计来进行手势定位

GestureSpottingUsingWristWornMicrophone

and3-AxisAccelerometer

（原文2）Abstract.Weperformcontinuousactivityrecognitionusingonlytwowrist-wornsensors-a3-axisaccelerometerandamicrophone.Webuildontheintuitivenotionthattwoverydierentsensorsareunlikelytoagreeinclassicationofafalseactivity.Bycomparingimperfect,slidingwindowclassicationsfromeachofthesesensors,weareablediscernactivitiesofinterestfromnulloruninterestingactivities.Whereonesensoraloneisunabletoperformsuchpartitioning,usingcomparisonweareabletoreportgoodoverallsystemperformanceofupto70%accuracy.Inpresentingtheseresults,weattempttogiveamore-indepthvisualizationoftheerrorsthancanbegatheredfromconfusionmatricesalone.

1Introduction

Handactionsplayacrucialroleinmosthumanactivities.Asaconsequencesdetectingandrecognisingsuchactivitiesisoneofthemostimportantaspectsofcontextrecognition.Atthesameitisoneofthemostdicult.Thisisparticularlytrueforcontinuousrecognitionwhereasetofrelevanthandmotions（gestures）needtobespottedinadatastream.Thedicultiesofsuchrecognitionstemfromtwothings.First,duetoalargenumberofdegreesoffreedom,handmotionstendtobeverydiverse.Thesameactivitymightbeperformedinmanydierentwaysevenbyasingleperson.Second,intermsofmotion,handsarethemostactivebodyparts.Wemoveourhandscontinuously,mostlyinanunstructuredway,evenwhennotdoinganythingparticularwiththem.Infactinmostsituationssuchunstructuredmotionsbyfaroutnumbergesturesthatarerelevantforcontextrecognition.Thismeansthatacontinuousgesturespottingapplicationshastodealwithanzeroclassthatisdiculttomodelwhiletakingupmostofthesignal.

1.1PaperContributions

Ourgrouphasinvestedaconsiderableamountofworkintohandgesturespotting.Todatethisworkhasfocusedonusingseveralsensorsdistributedovertheuserfibodytomaximisesystemperformance.Thisincludedmotionsensors（3axisaccelerometer,3axisgyroscopesand3axismagneticsensors）ontheupperandlowerarm[3],microphone/accelerometercombinationontheupperandlowerarm[5]aswellas,morerecently,acombinationofseveralmotionsensorsandultrasoniclocationdevices.Thispaperinvestigatestheperformanceofagesturespottingsystembasedonasingle,wristmounteddevice.

Theideabehindtheworkisthatwristmountedaccessoriesarebroadlyacceptedandwornbymostpeopleondailybasis.Incontrast,systemsthatrequiretheusertoputonseveralsensorsatlocationssuchastheupperarmwouldhavemuchmoreproblemswithuseracceptance.

Thedownsideofthisapproachisthereducedamountofinformationavailablefortherecognition.Thisforexamplemeansthatthemethodofanalysingsoundintensitydierencesbetweenmicrophonesondierentpartsofthebodythatwasthecornerstoneofourprevioussignalpartitioningworkisnotfeasible.Thisproblemiscompoundedbythefactthatfortheapproachtomakesensethatwristmounteddevicecanneithercontaintoomanysensorsnorcanitrequirecomputingand/orcommunicationpowerthatwouldimplylarge,bulkybatteries.

Themaincontributionofthepaperistoshowthat,foracertainsubsetofactivities,reasonablegesturespottingresultscanbeachievedwithacombinationofamicrophoneand3axisaccelerometermountedonthewrist.Ourmethodreliesonsimplejumpingwindowsoundprocessingalgorithmsthatwehaveshown[10]torequireonlyminimalcomputational

andcommunicationperformance.FortheaccelerationweuseinferenceonHiddenMarkovModels（HMM）,againonjumpingwindowsacrossthedata.

Toourknowledgethisisthersttimethatsuchasimplesystemandastraightforwardjumpingwindowmethodhasbeensuccessfullyusedforhandgesturespottingincontinuousdatastreamwithadominant,unstructuredzeroclass.Previouslysuchsetupsandalgorithmshaveonlybeenshowntobesuccessfulleitherforsegmentedrecognitionorforscenarioswherethezeroclasswaseithereasytomodelornotrelevant（e.g.recognitionofstanding,sitting,walking,running[6,9,12]）.Wheretheseapproachesuseaccelerationsensors,intheworkof[?

?

]soundwasexploitedforperformingsituationanalysisinthewearablecomputingdomain.Also[?

]usedsoundinformationtoimprovetheperformanceofhearingaids.Complimentaryinformationfromsoundandaccelerationhasbeenusedbeforetodetectdefectsinmaterialsurfaces,e.g.in[13],butnoworkthattheauthorsareawareusestheseforrecognitionofcomplexactivities.

Inthepaperwesummarisethesoundandaccelerationalgorithmsandthenfocusontheperformanceofdierentfusionmethods.Itisshownthatappropriatefusionisthekeytoachievinggoodperformancedespitesimplesensorsandalgorithms.Weverifyourapproachondatafromawoodworkshopassemblyexperimentthathavewehaveintroducedandusedinpreviouswork[5].Wepresenttheresultsusingbothtraditionalconfusionmatrices,plusanovelvisualisationmethodthatprovidesamorein-depthunderstandingoftheerrortypes.

2RecognitionMethod

Weapplyslidingwindowsoflenghtwsecondsacrossallthedatainincrementsofw.AteachstepweapplyanwjmpLDAbasedclassicationonthesounddata,andanHMMclassicationonthesound.Thefisoftresultsofeachclassication-LDAdistancesforsoundandHMMclasslikelihoodsforacceleration-areconvertedintoclassrankings,andthesearefusedtogetherusingoneoftwomethods:

comparisonoftoprank（COMP）,andamethodusingLogisticRegression

（LR）.

2.1FramebyFrameSoundClassication

UsingLDAFrame-by-framesoundclassicationwascarriedoutusingpatternmatchingoffeaturesextractedinthefrequencydo-main.Eachframerepresentsawindowon100msofrawaudiodata.Thesewindowsarethenjumpedovertheentiredatasetin25msincrements,producinga40Hzoutput.

Theaudiostreamwastakenatasamplerateof2kHzfromthewristwornmicrophone.FromthisaFastFourierTrans-form（FFT）wascarriedoutoneach100mswindow,generatinga100binoutputvector（12fsfftwnd=122100=100bins）.

Makinguseofthefactthatourrecognitionproblemrequiresasmallnitenumberofclasses,weappliedLinearDiscriminantAnalysis（LDA）[1]toreducethedimensionalityoftheseFFTvectorsfrom100to#Classes1.

ClassicationofeachframecanthenbecarriedoutusingasimpleEuclideanminimumdistancecalculation.Wheneverwewishtomakeadecision,wesimplycalculatetheincomingpointinLDAspaceandnditsnearestclassmeanvaluefromthetrainingdataset.ThissavingincomputationcomplexitybydimensionalityreductioncomesatthecomparativelyminorcostofrequiringustocomputeandstoreasetofLDAclassmeanvaluesfromwhichtheLDAdistancesmightbeobtained.

Equally,anearestneighbourapproachmightbeused.Fortheexperimentdescribedherehowever,Euclideandistancewasfoundtobesucient.

Alargerwindow,wlen,wasmovedoverthedatainwjmpsecondincrements.Thisrelativelylargewindowwaschosentoreectthefactthatalloftheactivitiesweareinterestedinoccuratthetimescaleofatleastseveralseconds.OneachwindowwecomputeasumoftheconstituentLDAdistancesforeachclass.Fromthesetotaldistances,wethenrankeachclassaccordingtominimumdistance.Classicationofthewindowisthensimplyamatterofchoosingthetoprankingclass.

2.2HMMAccelerationClassication

Incontrasttotheapproachusedforsoundrecognition,weemployedmodelbasedclassication,specicallytheHiddenMarkovModel（HMM）,forclassifyingaccelerometerdata[8,

11].（TheimplementationoftheHMMlearningandinferenceroutinesforthisexperimentwasprovidedcourtesyofKevinP.MurphyfiHMMToolboxformatlab[7].）

ThefeaturesusedtofeedtheHMMmodelswerecalculatedfromsliding100mswindowsonthex,y,andzaxisofthe100Hzsampledaccelerationdata.Thesewindowsweremovedoverthedatain25msincrements,producingthefollowingfeatures,outputat40Hz:

Meanofx-axis

Varianceofx-axis

Acountofthenumberofpeaks（forx,y,z）

Meanamplitudeofthepeaks（forx,y,z）

Finallywegloballystandardisedthefeaturessoastoavoidnumericalcomplicationswiththemodellearningalgorithmsinmatlab.

InpreviousworkweemployedsingleGaussianobservationmodels,butthiswasfoundtobeinadequateforsomeclassesunlessalargenumberofstateswereused.Intuitively,thedescriptivepowerofamixtureofGaussianismuchclosertofirealiythanonlyone,andsofortheseclassesamixturemodelwasused.Thespecicnumberofmixturesandthenumberofhiddenstatesusedwereindividuallytailoredbyhandforeachclass.Theparametersthemselvesweretrained

fromthedata.

Awindowofwlen,inwjmpincrements,wasrunovertheaccelerationfeatures,andthecorrespondingloglikelihoodforeachHMMclassmodelcalculated.

Classicationiscarriedoutforeachwindowbychoosingtheclasswhichproducesthelargestloglikelihoodgiventhestreamoffeaturedatafromthetestset.

2.3Fusionofclassiers

Comparisonoftopchoices（COMP）Thetoprankingsfromeachofthesoundandaccelerationclassiersforagivenjumpingwindowsegmentaretaken,compared,andreturnedasvalidiftheyagree.Thosewherebothclassiersdisagreearethrownout-classiedasnull.

Logisticregression（LR）Themainproblemwithadirectcomparisonoftopclassierrankingsisthatitfailstotakeintoaccountcaseswhereoneclassiermightbemorereliablethananotheratrecognisingparticularclasses.Ifoneclassierreliablydetectsaclass,buttheotherclassierfails

to,perhapsrelegatingtheclasstosecondorthirdran