Figure1Systemoverview21AcousticsSeminalworkinacousticshasshownthatpeopleinaspacesigni cantlyimpactreverberationandthatreverberationisfrequency2aswellasroomgeometrydependent4Overthelast120yearst ID: 327152
Download Pdf The PPT/PDF document "humans)senseoccupancy.Reverberationisbot..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
humans)senseoccupancy.Reverberationisbothfrequencydependentandchangesbasedontheroomgeometry,wallmaterialsandfurniturematerial.Makingaccurateandgeneralizablemodelsofre-verberationisquitechallenging.Forthisreason,weproposeanapproachwherethereverberationistrainedonaper-roombasisusingamachinelearningapproach.Insteadofmeasuringandclassifyingthereverberationatdiscretefre-quencieslikewhatisdoneforconcerthalls,weuseultrasonicchirpsthatsweepacrossafrequencyrangetorapidlymea-surethespacesincewearenotconcernedwithexactlyquan-tifyingreverberation.Chirpscanalsobeconstructedusingfade-inandfade-outperiodsthatpreventaudibleartifactsinlow-costspeakersthatcouldbedetectedbyhumans[3].Sincethere ectionscomingbackfromthesesignalsareroomspecic,weapplyasemi-supervisedmachinelearningap-proachthatisabletomodelthecharacteristicsoftheroomundermultipleloadsinordertoestimatehowreverberationchangeswithrespecttonumberofpeople.Typicallythisrequirestakingsampleswhentheroomisemptyaswellaswhentheroomhasenoughpeopletomakeasignicantdif-ferenceinreverberationtimes.Alternativesignalcharacter-isticslikeDopplershiftorsimplytimeofdayschedulescanbeusedtodeterminewhentheroomisemptyforperiodicre-calibrationofthezeropoint.Figure1showsanoverviewofourproposedsystemwhereatweetertransmitsanultrasonicchirpintoaroomandaco-locatedmicrophoneisusedtoreceivethere ectedsig-nal.Anelectronicspackageisresponsibleforgeneratingthesignalandthenprocessingthere ectedsignal.Ourproto-typesystemusesacomputerforthispurpose,butweshowthattheactualrun-timecomputationofthesystemissim-pleenoughtoexecuteentirelyfromaplatformbasedonamicro-controller.Therearefourmainresearchchallengesassociatedwithourproposedsystem.First,weneedtodesignanappropri-ateexcitationsignalthatisbothinaudibletohumansandalsoexcitestheroominamannerthatcanclearlydistin-guishchangesasthenumberofpeopleincrease.Second,weneedatechniquethatcansamplequicklyandecientlysothatoccupancycanbeestimatedbeforethedynamicsoftheroomchange.Thisapproachalsorequiresatransducerthatisabletouniformlydistributetheultrasonicsignal.Third,weneedalgorithmsthatcanclassifyreceivedsignalsinor-dertoestimateload.Finally,weneedanapproachthatcanperiodicallyretraininordertoadapttoslightchangesintheenvironmentovertime.2.RELATEDWORKInthissection,wediscussthebackgroundrelatedtoacous-ticsfollowedbysimilarapproachesthathavebeenusedtomeasurebothpresenceandoccupancy.Commoncommer-ciallyavailableoccupancysensorslikePIRmotiondetectors;ultrasonicmotiondetectorsandmicrowavesensorsusuallyonlydetectpresence(ifoneormorepeopleareinaroom).Camerasandmoreadvancedinfraredsystemsattempttoes-timatetheactualnumberofpeopleinaspace,butaretyp-icallyexpensive,diculttotrainandsu erfromocclusion.Ourproposedapproachiscomparativelylow-cost,relativelyeasytotrainandhastheadvantageofllinganentirespacewithsoundmakingitmoreimmunetoobstacles. Figure1:Systemoverview2.1AcousticsSeminalworkinacousticshasshownthatpeopleinaspacesignicantlyimpactreverberationandthatreverberationisfrequency[2]aswellasroomgeometrydependent[4].Overthelast120yearstherehavebeencountlesse ortsproposedtomodeltheseacousticpropertiesinordertoimprovecon-certhallperformance.Recentworkinthisspacehasusedcomputersimulations[5{8].Itisclearfromthislargebodyofresearchthatcreatingsimplegeneralizablemodelsofre-verberationisquitechallenging.Forthisreason,wepro-poseusingmachinelearningtechniquestolearnandclassifythereverberationresponseonaper-installationbasis.Invariousrecentprolesofreverberation[9],itisclearthatgivenaparticularroomgeometry,audienceabsorptionfol-lowsrelativelydistinctcurvesthatmakeitanidealfeatureforoccupancydetection.Activeacousticapproacheshaveshowngreatpotentialinmultipleformsofsensing.In[10],theauthorsuseasinglespeakerwithmultiplemicrophonestodeterminetheshapeofaroombasedonechoes.In[11],theauthorsshowhowre ectedDopplersignalscanbeusedtoclassifyanythingfromspeech,towalkingmotionandevengestures.Tothebestofourknowledge,thisisoneoftherstsystemwhereultrasoundhasbeenusedtodirectlyestimateoccupancy.2.2OccupancyAsidefromtheconventionalsolutionofusingPIRsensorstodetectthepresenceofpeople,mostotherrelatedworkhasbeencarriedoutonusingcamerasormultiplesensorstomeasureoccupancylevel.Alloftheseapproachesgener-allyfallintotwocategoriesbasedonslightlydi erentgoals.Onegroupfocusesononlydetectingthepresenceofpeo-ple[12][13][14][15],whichoftencomeswithanalysisofmoredetaileduserbehaviorandactions.Theothercate-goriesfocusesonpeoplecounting[16][17][18][19],usuallyinvolvingmoresophisticatedalgorithmsforlearning.PresenceDetectionInthecategoryofpresencedetection,manyapproachesfusedatareadingsfromdi erentsensortypes.Forexamplein Figure6:Rawfeaturesforempty,half-full,andfullroomscenariosnamelythereverberationtime(RT),followsarectangularhyperbolacurveagainstthetotalabsorbingmaterial.Herec20isthespeedofsoundat20degreeCelsius,Visthevol-ume(m3)oftheroom,Sisthetotalsurfacearea(m2)ofaroom,andaistheaverageabsorptioncoecientofroomsurface.RT60=24ln10 c20V Sa'0:1611V Sa(2)SincetheRTisdenedbythetimeforasignaltodecaybyacertaindecibel(dB),weget(3)RT/log(A0 Am)(3)whereA0istheconstantinitialamplitudeofthesoundsourceandAmisthemeasuredamplitudeafterabsorption.Combiningequation(2)and(3),weobtaintherelationshipbetweentheobservedfrequencyamplitudeandnumberofpeopleas(4)Am/e�C0V Sa(4)Asplottedin7,wecanseethatwhenthevolumeoftheroomissmall,thecurvetendstobesimilartoanexponentialregression.However,asthevolumeoftheroomincreases,thecurvebecomessmootherandmorelinearinregression.Thesizeoftheroomcanbeestimatedtohelpchoosethebeststartingmodel.Tocalculatetheamplitudedi erence,werstre-calibratethemeanoftheemptyroomdataastheneworiginoftheprojectedspace,andforeveryclusterswecalculatehowfartheyarefromtheorigin.WetestedwithmultipledistancemetricsanddecidedthatChebyshevdistanceprovidedthebestttoregressionmodelshownacrossouroveralldata.WeusetheChebyshevdistancedenedas, Figure7:Theoreticalregressiontrendswithdi erentroomvolumesbasedonequation(4)Dchebyshev(a;b)=max1in(jai�bij)(5)wherea;baretwoarbitraryn-dimensionaldatapoints.Theunitdistanceisfurthercalculatebasedontheaverageofthepairwise-distancebetweenthetwotrainingdatasets,wheretheunitdistanceisnamelythereferencedistancebe-tweenNand(N+1)peopleinstance.Next,weestimateeachclusterbyttingitsdistancetotheorigintotheregres-sionmodel.Byndingthevariablethatchangesthemostamongallthedata,whichnotedhereisderivedfromalin-earcombinationofallthevariablesintheoriginalspace,wecapturethefeaturethatdi erentiatesthedatathemostanduseditasameasurementtoestimatetheoccupancylevel.Forroomswithasmallvolume,anexponentialregression Figure12:Estimationmadebyouralgorithmcomparedtogroundtruthinsmallandmediumsizerooms Figure13:Estimationcomparedwithgroundtruthaspeo-pleenteranauditoriumsignicantlyimpacttheresult.Weperformtestsincludingopeningthedoortotheroom,openingwindowsintheroom,changingthevolumeofthetransmitter,andthentestinginthesameroomoneweeklater.AsshowninTable1,theerrorwasmoste ectedbychangesinvolumeandslightlybyopeningthewindows.Errorduetochangesinvolumearenotsurprisingsincetheregressionmodelisbuiltaroundmagnitudechangesindi erentfrequencies.Totestthesystem'sabilitytoautomaticallyretrainit-self,anexperimentiscarriedoutinthesameroomaweeklaterwithslightlydi erentpositionandvolume.Withoutself-retrainonthenewenvironment,theerrorincreasesby1:2%.Thiscouldaccumulateandpotentiallygrowworseovertime.However,ifthebaselineandtheunitdistanceiscorrectlycalibrated,whichcanbedoneifanemptyroomcanbedetected,thechangeinerrorisnegligible.TheresultagainshowsinTable1,lessthan1%di erencebetweenthecalibratedunitdistanceandtheidealone.Finally,theoverallperformanceofthesystemissumma-rizedinTable2,andthecomparisonwithrelatedapproachesinpeoplecountingisshowninTable3.Thecomparisonval-ueswereextractedfromeachpaper.Thenumberofpeopleestimatedbythesystemisnomorethan3peopledi erentfromtheactualnumberonaverage,andtheaverageerrorinpercentagetothemaximumcapacityoftheroomisaround5%. InterferenceType ErrorInc.(%) Dooropened 1.63 Windowsopened 2.38 Changevolume 5.38 Changepositionofthedevice 2.12 Datacollectedaweeklater(noretrain) 1.18 Datacollectedaweeklater(auto-retrained) 0.08 Table1:Systemperformancewitherrorsourcesinsmallroom SizesParm MaxCap. Avg.Error Error/MaxCap.(%) Smallroom 8 0.61 7.6 Mediumroom 30 1.6 5.3 Largeroom 150 2.6 1.7 Table2:Systemperformancewithdi erentroomsizes5.LIMITATIONSOurproposedtechniquehasafewdrawbacksassociatedwiththefactthatitisanactivesensingsystem.Ifmultipleofourtransducersareplacedinthesameroom,thereneedstobeamechanismtocoordinatetransmissionssothattheydonotexperiencecross-talk.Forlargespaces,thereneedstobeaproportionallypowerfultransmitterthatwilleventu-allyrequirealargeramplierandtransducer.Asthespaceincreasesinsize,theabilitytonelydistinguishtheexactnumberofpeoplediminishes.Forlargerspaces,thesystemalsorequiresacalibrationpointwithenoughpeopletoreg-isterasapproximately5-10%oftheroomloadforthebestresults.Thiscanalsobehardtocoordinateincertainen-vironments.Weimagineinthefuturethatthisapproachcouldbecoupledwithotherformsofpeoplecountingtohelpaidinautomaticcalibration.Finally,ultrasoundinourparticularfrequencyisstilldetectablebyanimals.Beyondtransducercost(whichbenetsfrombeingcompatiblewithcommodityaudioequipment)thereisnoreasonwhythisapproachcannotoperateathigherfrequencies.Athigherfrequenciessoundbecomesmoredirectional,sofurtherin-vestigationwouldberequiredtodetermineifreverberationisstillassensitivetopersoncount.6.CONCLUSIONInconclusion,thispaperintroducedanultrasonicap-proachforestimatingtheoccupancylevelofaroomus-ingreverberationacrossmultiplefrequencies.Thesystemconsistsofanomni-directionalultrasonictweeterwithaco-locatedmicrophonethatrsttransmitsanultrasonicchirp MethodProposed[16][17][19] Max.Counts5012355 Avg.Error1.60.41.30.7 Environ.indoorindooroutdoorindoor Complexitylowmediumhighmedium Costlowhighmediumlow Table3:Overallsystemperformancecomparisonofmultiplepeoplecountingapproaches