edu ABSTRACT WiFi signals are typically information carriers between a trans mitter and a receiver In this paper we show that WiFi can also extend our senses enabling us to see moving objects through walls and behind closed doors In particular we can ID: 13118
Download Pdf The PPT/PDF document "See Through Walls with WiFi Fadel Adib a..." 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.
[2]LAN/MANCSMA/CDE(ethernet)accessmethod.IEEEStd.[3]LP0965..EttusInc.[4]NintendoWii.[5]RadarVision..TimeDomain[6]Seeingthroughwalls-MITÕsLincolnLaboratory.[7]UrbanEyes..LawrenceLivermoreNationalLaboratory.[8]USRPN210..EttusInc.[9]X-boxKinect..Microsoft.[10]R.Bohannon.Comfortableandmaximumwalkingspeedofadultsaged20-79years:referencevaluesanddeterminants.Ageandageing,1997.[11]G.Charvat,L.Kempel,E.Rothwell,C.Coleman,andE.Mokole.Athrough-dielectricradarimagingsystem.IEEETrans.AntennasandPropagation,2010.[12]G.Charvat,L.Kempel,E.Rothwell,C.Coleman,andE.Mokole.Anultrawideband(UWB)switched-antenna-arrayradarimagingsystem.IEEEARRAY,2010.[13]K.Chetty,G.Smith,andK.Woodbridge.Through-the-wallsensingofpersonnelusingpassivebistaticwiÞradaratstandoffdistances.IEEETrans.GeoscienceandRemoteSensing,2012.[14]J.Choi,M.Jain,K.Srinivasan,P.Levis,andS.Katti.Achievingsinglechannel,fullduplexwirelesscommunication.InACM,2010.[15]G.Cohn,D.Morris,S.Patel,andD.Tan.Humantenna:usingthebodyasanantennaforreal-timewhole-bodyinteraction.InACM,2012.[16]T.CoverandJ.Thomas.ElementsofinformationtheoryWiley-interscience,2006.[17]S.Gollakota,F.Adib,D.Katabi,andS.Seshan.ClearingtheRFsmog:Making802.11robusttocross-technologyinterference.InACMSIGCOMM,2011.[18]S.Hong,J.Mehlman,andS.Katti.Picasso:fullduplexsignalshapingtoexploitfragmentedspectrum.InACMSIGCOMM,2012.[19]M.Jain,J.Choi,T.Kim,D.Bharadia,S.Seth,K.Srinivasan,P.Levis,S.Katti,andP.Sinha.Practical,real-time,fullduplexwireless.InACMMobiCom,2011.[20]H.Junker,P.Lukowicz,andG.Troster.Continuousrecognitionofarmactivitieswithbody-worninertialsensors.InIEEEISWC,2004.[21]Y.KimandH.Ling.HumanactivityclassiÞcationbasedonmicro-dopplersignaturesusingasupportvectormachine.Trans.GeoscienceandRemoteSensing,2009.[22]K.Lin,S.Gollakota,andD.Katabi.RandomaccessheterogeneousMIMOnetworks.InACMSIGCOMM,2010.[23]B.Lyonnet,C.Ioana,andM.Amin.HumangaitclassiÞcationusingmicrodopplertime-frequencysignalrepresentations.InIEEERadarConference,2010.[24]B.Michoud,E.Guillou,andS.Bouakaz.Real-timeandmarkerless3Dhumanmotioncaptureusingmultipleviews.MotionÐUnderstanding,Modeling,CaptureandAnimation,2007.[25]A.Oppenheim,R.Schafer,J.Buck,etal.Discrete-timesignalprocessing.PrenticehallEnglewoodCliffs,NJ:,1989.[26]H.Rahul,S.Kumar,andD.Katabi.JMB:scalingwirelesscapacitywithuserdemands.InACMSIGCOMM,2012.[27]T.Ralston,G.Charvat,andJ.Peabody.Real-timethrough-wallimagingusinganultrawidebandmultiple-inputmultiple-output(MIMO)phasedarrayradarsystem.InIEEEARRAY,2010.[28]S.Ram,C.Christianson,Y.Kim,andH.Ling.Simulationandanalysisofhumanmicro-dopplersinthrough-wallenvironments.IEEETrans.GeoscienceandRemoteSensing,2010.[29]S.Ram,Y.Li,A.Lin,andH.Ling.Doppler-baseddetectionandtrackingofhumansinindoorenvironments.JournaloftheFranklin,2008.[30]S.RamandH.Ling.Through-walltrackingofhumanmoversusingjointdopplerandarrayprocessing.IEEEGeoscienceandRemoteSensingLetters,2008.[31]T.-J.Shan,M.Wax,andT.Kailath.Onspatialsmoothingfordirection-of-arrivalestimationofcoherentsignals.IEEETrans.onAcoustics,SpeechandSignalProcessing,1985.[32]F.SoldovieriandR.Solimene.Through-wallimagingviaalinearinversescatteringalgorithm.IEEEGeoscienceandRemoteSensingLetters,2007.[33]R.Solimene,F.Soldovieri,G.Prisco,andR.Pierri.Three-dimensionalthrough-wallimagingunderambiguouswallIEEETrans.GeoscienceandRemoteSensing,2009.[34]P.StoicaandR.L.Moses.SpectralAnalysisofSignals.PrenticeHall,[35]W.C.Stone.Nistconstructionautomationprogramreportno.3:Electromagneticsignalattenuationinconstructionmaterials.InConstructionAutomationWorkshop1995[36]K.Tan,H.Liu,J.Fang,W.Wang,J.Zhang,M.Chen,andG.Voelker.SAM:EnablingPracticalSpatialMultipleAccessinWirelessLAN.ACMMobiCom,2009.[37]D.Titman.Applicationsofthermographyinnon-destructivetestingofstructures.NDT&EInternational,2001.[38]H.Wang,R.Narayanan,andZ.Zhou.Through-wallimagingofmovingtargetsusinguwbrandomnoiseradar.IEEEAntennasandWirelessPropagationLetters,2009.[39]J.XiongandK.Jamieson.ArrayTrack:aÞne-grainedindoorlocationsystem.InUsenixNSDI,2013.[40]Y.YangandA.Fathy.See-through-wallimagingusingultrawidebandshort-pulseradarsystem.InIEEEAntennasandPropagationSocietyInternationalSymposium,2005[41]Y.YangandA.Fathy.Designandimplementationofalow-costreal-timeultra-widebandsee-through-wallimagingradarsystem.InIEEE/MTT-SInternationalMicrowaveSymposium,2007.ConvergenceofIterativeNulling.Weprovewhyiterativenullingpro-posedin¤4converges.Wi-Vimodelsthechannelestimateerrorsasad-ditive(inlinewithcommonpracticeofmodelingquantizationerror[25]).Hence,bysubstituting,and,inEq.1,weres h2+!2$#h1 h2!2"!1+!1!2 whichfollowsfromtheÞrstorderTaylorseriesapproximationof IteratingonWeÞrstanalyzehowthealgorithmconvergesifitwereiteratingonlyonStep1.AccordingtoAlgorithm1,isreÞnedtores.Byupdatingtheprecodingvector,thenewreceivedchannelafterresres byapplyingtheÞrstorderTaylorseriesapproximation .Hence,resres.Therefore,afterthe-thiteration,resres IteratingonWenowanalyzehowthealgorithmconvergesifitwereiteratingonlyonStep2.AccordingtoAlgorithm1,isreÞnedtores .Byupdatingtheprecodingvector,thenewreceivedchannelafternullingis:res res res whichfollowsfromtheÞrstorderTaylorseriesapproximationof res.Hence,resres,andresconvergesasabove.IterativenullingonBytheabovearguments,afteriterationson,thenulledchannelbecomes:resres h2$i+j(11) 86 0 0.2 0.4 0.6 0.8 1 0 5 10 15 20 25 30 Fraction of experiments SNR of Gestures (in dB) Bit Õ0Õ Bit Õ1Õ Figure11CDFofthegestureSNRs.TheÞgureshowstheCDFsoftheSNRafterapplyingthematchedÞltertakenoverdifferentdistancesfromWi-Vi. Percentage of correct detection 100 100 100 100 87.5 0 20 40 60 80 100 Free Space Tinted Glass 1.75" Solid Wood Door 6" Hollow Wall 8" Concrete (a)DetectionAccuracy SNR (in dB) 0 5 10 15 20 25 30 35 Free Space Tinted Glass 1.75" Solid Wood Door 6" Hollow Wall 8" Concrete (b)SNRFigure12Gesturedetectionindifferentbuildingstructures.(a)plotsthedetectionaccuracyofWi-Vifordifferenttypesofobstructions.(b)showstheaverageSNRoftheexperimentsdonethroughthesedifferentmaterials,withtheerrorbarsshowingtheminimumandmaximumachievedSNRsacrossthetrials.before,wealsotestWi-Viinasecondbuildinginouruniversitycampus,wherethewallsaredifferent.Inparticular,weexperimentwith4typesofbuildingmaterials:8concretewall,6hollowwallsupportedbysteelframeswithsheetrockontop,1.75solidwooddoor,andtintedglass.Inaddition,weperformexperimentsinfreespacewithnoobstructionbetweenWi-Viandthesubject.Ineachexperiment,thesubjectisaskedtostand3metersawayfromthewall(orWi-Viitselfinthecaseofnoobstruction)andper-formtheÔ0Õbitgesturedescribedabove.Foreachtypeofbuildingmaterial,weperform8experiments.Fig.12showsWi-ViÕsperformanceacrossdifferentbuildingma-terials.SpeciÞcally,Fig.12(a)showsthedetectionrateasthefrac-tionofexperimentsinwhichWi-Vicorrectlydecodedthegesture,whereasFig.12(b)showstheaverageSNRsofthegestures.TheÞguresshowthatWi-Vicandetecthumansandidentifytheirges-turesacrossvariousindoorbuildingmaterials:tintedglass,solidwooddoors,6hollowwalls,andtoalargeextent8walls.Asexpected,thethickeranddensertheobstructingmaterial,theharderitisforWi-Vitocapturereßectionsfrombehindit.DetectinghumansbehinddifferentmaterialsdependsonWi-ViÕspoweraswellasitsabilitytoeliminatetheßasheffect.Fig.13plotstheCDFoftheamountofnulling(i.e.,reductioninSNRs)thatWi-Viachievesinvariousexperiments.TheplotshowsWi-ViÕsnulling 0 0.2 0.4 0.6 0.8 1 0 10 20 30 40 50 60 Fraction of experiments Nulling (in dB) Figure13CDFofachievednulling.TheÞgureplotstheCDFwhichshowstheabilityofnullingtoreducethepowerreceivedalongstaticpaths.reducesthesignalfromstaticobjectsbyamedianof40dB.ThisnumberindicatesthatWi-Vicaneliminatetheßashreßectedoffcommonbuildingmaterialsuchasglass,solidwooddoors,interiorwalls,andconcretewallswithalimitedthickness[1].However,itwouldnotbeabletoseethroughdensermateriallikere-enforcedconcrete.Toimprovethenulling,onemayuseacirculatorattheanalogfrontend[18]orleveragerecentadvancesinfull-duplexradio[14],whichwerereportedtoproduce80dBreductioninin-terferencepower[19].8.CWepresentWi-Vi,awirelesstechnologythatusesWi-Fisig-nalstodetectmovinghumansbehindwallsandinclosedrooms.Incontrasttoprevioussystems,whicharetargetedforthemilitary,Wi-Vienablessmallcheapsee-through-walldevicesthatoperateintheISMband,renderingthemfeasibletothegeneralpublic.Wi-Vialsoestablishesacommunicationchannelbetweenitselfandahu-manbehindawall,allowinghim/hertocommunicatedirectlywithWi-Viwithoutcarryinganytransmittingdevice.WebelievethatWi-Viisaninstanceofabroadersetoffunc-tionalitythatfuturewirelessnetworkswillprovide.FutureWi-Finetworkswilllikelyexpandbeyondcommunicationsanddeliverservicessuchasindoorlocalization,sensing,andcontrol.Wi-VidemonstratesanadvancedformofWi-Fi-basedsensingandlocal-izationbyusingWi-Fitotrackhumansbehindwall,evenwhentheydonotcarryawirelessdevice.ItalsoraisesissuesofimportancetothenetworkingcommunitypertinenttouserprivacyandregulationsconcerningtheuseofWi-Fisignals.Finally,Wi-Vibridgesstate-of-the-artnetworkingtechniqueswithhuman-computerinteraction.ItmotivatesanewformofuserinterfaceswhichrelysolelyonusingthereßectionsofatransmittedRFsignaltoidentifyhumangestures.Weenvisionthatbylever-agingÞnernullingtechniquesandemployingbetterhardware,thesystemcanevolvetoseeinghumansthroughdenserbuildingmate-rialandwithalongerrange.TheseimprovementswillfurtherallowWi-Vitocapturehigherqualityimagesenablingthegesture-basedinterfacetobecomemoreexpressivehencepromisingnewdirec-tionsforvirtualreality.Acknowledgments:WethankOmidAbari,HaithamHassanieh,EzzHamad,andJueWangforparticipatinginourexperiments.WealsothankNabeelAhmed,ArthurBerger,DiegoCifuentes,PeterIannucci,ZackKa-belac,SwarunKumar,NateKushman,HariharanRahul,LixinShi,there-viewers,andourshepherd,VenkatPadmanabhan,fortheirinsightfulcom-ments.ThisresearchissupportedbyNSF.WethankmembersoftheMITCenterforWirelessNetworksandMobileComputing:Amazon.com,Cisco,Google,Intel,Mediatek,Microsoft,STMicroelectronics,andTelefonicafortheirinterestandsupport.9.R[1]HowSignalisaffected..CityofCumberlandReport. 85 0 0.2 0.4 0.6 0.8 1 0 1 2 3 4 5 Fraction of experiments Spatial Variance of the MUSIC image (in tens of millions) No humans One human Two humans Three humans Figure9CDFofspatialvarianceforadifferentnumberofmovingAsthenumberofhumansincreases,thespatialvarianceincreases. !!!!!!!ActualDetected 0 1 2 3 0 100% 0% 0% 0% 1 0% 100% 0% 0% 2 0% 0% 85% 15% 3 0% 0% 10% 90% Table2AccuracyofAutomaticDetectionofHumans.Thetableshowstheaccuracyofdetectingthenumberofmovinghumansbasedonthespatialvariance.seethatthespatialvarianceishigherwithmoremovingbodiesintheroom.Interestingly,theseparationbetweensuccessiveCDFsdecreasesasthenumberofhumansincreases.Inparticular,theseparationislargerbetweentheCDFsofnohumansandonehuman,thanbetweentheCDFsofonehumanandtwohumans.Thesepara-tionistheleastbetweentheCDFsof2humansand3.Tounder-standthisbehavior,recallthatbecausetheroomhasaconÞnedspace,asthenumberofpeopleincreases,thefreedomofmove-mentdecreases.Hence,addingahumantoacongestedspaceisexpectedtoaddlessspatialvariancethanaddinghertoalesscongestedspacewhereshehasmorefreedomtomove.Next,wewouldliketoautomatethethresholdsfordistinguish-ing0,1,2,and3movinghumans.Todoso,wedividethedataintoatrainingsetandatestingset.ToensurethatWi-Vicangeneralizeacrossenvironments,weensurethatthetrainingexamplesareallconductedinoneconferenceroom,whilethetestingexamplesareconductedinanotherconferenceroom(Recallthatthetworoomshavedifferentsizes).Weusethetrainingsettolearnthethresholdstoseparatethespatialvariancescorrespondingto0,1,2,and3hu-mans.Wethenusethesethresholdstoclassifytheexperimentsinthetestingset.Finally,weperformcross-validation,i.e.,werepeatthesameprocedureafterswitchingthetrainingandtestingsets.Table2showstheresultoftheclassiÞcation.ItshowsthatWi-Vicanidentifywhetherthereis0or1personinaroomwith100%accuracy;thisisexpectedbasedontheCDFsinFig.9.Also,row3showsthattwohumansareneverconfusedwith0or1.How-ever,Wi-Viconfused2humanswith3humansin15%ofthetrials,whereasitaccuratelyidentiÞedtheirnumberin85%ofthecases.7.5GestureDecodingNext,weevaluateWi-ViÕsabilitytodecodethebitsassociatedwiththegesturesin¤6.Ineachexperiment,ahumanisaskedtostandataparticulardistancefromthewallthatseparatestheroomfromourdevice,andperformthetwogesturescorrespondingtobitÔ0ÕandbitÔ1Õ.Eachhumantookstepsatalengththeyfoundcomfortable.Typicalstepsizeswere2-3feet.Theexperimentsarerepeatedatvariousdistancesintherange[1m,9m].Allexperi-mentsareconductedinthesameconferenceroomsdescribedaboveandunderthesameexperimentalconditions.Oneofourconference Percentage of correct detection Distance (in m) 75 Bit Õ0Õ 100 100 100 100 100 100 100 75 0 Bit Õ1Õ 0 20 40 60 80 100 1 2 3 4 5 6 7 8 9 Figure10AccuracyofGestureDecodingasaFunctionofDistance.TheÞgureshowsthefractionofexperimentsinwhichWi-Vicorrectlyde-codedthebitassociatedwiththeperformedgestureatdifferentdistancesseparatingthesubjectfromthewall.NotethatWi-VidecodesagestureonlywhenitsSNRisgreaterthan3dB;thisexplainsthesharpcutoffbetween8and9meters.roomsisonly7mwide,whereastheotheris11mwide.Hence,theexperimentswithdistanceslargerthan6metersareconductedinthelargerconferenceroom,whereasforalldistanceslessthanorequal6meters,ourexperimentsincludedtrialsfrombothrooms.TheobtainedtracesareprocessedusingthematchedÞlterandde-codingalgorithmdescribedin¤6.2.Fig.10plotsthefractionoftimethegesturesweredecodedcor-rectlyasafunctionofthedistancefromthewallseparatingWi-Vifromtheclosedroom.Wenotethefollowingobservations:Wi-Vicorrectlydecodedtheperformedgesturesatalldistanceslessthanorequalto5m.ItidentiÞed93.75%ofthegesturesper-formedatdistancesbetween6mand7m.At8m,theperformancestarteddegrading,leadingtocorrectidentiÞcationofonly75%ofthegestures.Finally,Wi-Vicouldnotidentifyanyofthegestureswhenthepersonwasstanding9mawayfromthewall.Itisimportanttonotethat,inourexperiments,Wi-Vinevermis-tookaÔ0ÕbitforaÔ1Õbitortheinverse.Whenitfailedtodecodeabit,itwasbecauseitcouldnotregisterenoughenergytodetectthegesturefromthenoise.ThismeansthatWi-ViÕserrorsareerasureerrorsasopposedtostandardbiterrors.Wemeasuredthetimeittookthedifferentsubjectstoperformaonebitgesture.Averagedoveralltraces,oursubjectstook2.2stoperformagesture,withastandarddeviationof0.4s.TogainfurtherinsightintoWi-ViÕsgesturedecoding,Fig.11plotstheCDFsoftheSNRsoftheÔ0ÕgestureandtheÔ1Õgesture,acrossalltheexperiments.Interestingly,thegestureassociatedwithaÔ0ÕbithasahigherSNRthanthegestureassociatedwithaÔ1Õbit.Thisisduetotworeasons:First,theÔ0Õgestureinvolvesastepfor-wardfollowedbyastepbackward,whereastheÔ1ÕgesturerequiresthehumantoÞrststepbackwardthenforward.Hence,forthesamestartingpoint,thehumanisonaverageclosertoWi-Viwhileper-formingtheÔ0Õgesture,whichresultsinanincreaseinthereceivedpower.Second,takingastepbackwardisnaturallyharderforhu-mans;hence,theytendtotakesmallerstepsintheÔ1Õgesture.ThisobservationisvisuallyevidentinFig.5whereaÔ0Õgesturehasahigherpower(red)thantheÔ1Õgesture.WenotethatthemainfactorlimitinggesturedecodabilitywithincreaseddistanceisthelowtransmitpowerofUSRPs.ThelineartransmitpowerrangeforUSRPsisaround20mW(i.e.,beyondthispowerthesignalstartsbeingclipped),whereasWi-FiÕspowerlimitis100mW.Hence,onewouldexpectthatwithbetterhardware,Wi-Vicanhaveahigherdecodingrange.7.6TheEffectofBuildingMaterialFinally,weevaluateWi-ViÕsperformancewithdifferentbuildingmaterials.Thus,inadditiontothetwoconferenceroomsdescribed 84 0 1 2 3 4 5 6 7 Time ( seconds ) -100 -80 -60 -40 -20 0 20 40 60 80 100 Angle (degrees) 1 1 0 1 2 3 4 5 6 7 Time ( seconds ) -100 -80 -60 -40 -20 0 20 40 60 80 100 Angle (degrees) 1 2 1 0 1 2 3 4 5 6 7 Time ( seconds ) -100 -80 -60 -40 -20 0 20 40 60 80 100 Angle (degrees) 1 2 3 (a1)(b1)(c1) 0 1 2 3 4 5 6 7 Time ( seconds ) -100 -80 -60 -40 -20 0 20 40 60 80 100 Angle (degrees) 0 1 2 3 4 5 6 7 Time ( seconds ) -100 -80 -60 -40 -20 0 20 40 60 80 100 Angle (degrees) 1 2 0 1 2 3 4 5 6 7 Time ( seconds ) -100 -80 -60 -40 -20 0 20 40 60 80 100 Angle (degrees) 1 2 3 (a2)(b2)(c2) 0 1 2 3 4 5 6 7 Time ( seconds ) -100 -80 -60 -40 -20 0 20 40 60 80 100 Angle (degrees) 0 1 2 3 4 5 6 7 Time ( seconds ) -100 -80 -60 -40 -20 0 20 40 60 80 100 Angle (degrees) 1 2 0 1 2 3 4 5 6 7 Time ( seconds ) -100 -80 -60 -40 -20 0 20 40 60 80 100 Angle (degrees) 1 2 3 (a)OneHuman(b)TwoHumans(c)ThreeHumansFigure8TrackinghumanmotionwithWi-Vi.TheÞguresshowoutputtraceswithadifferentnumberofhumansafterprocessingwiththesmoothedMUSICalgorithm.Theyplotplot!,n]where!istheangleinin"90,90isplottedonthey-axisandtimeisonthex-axis.(a)showstracesforonehuman;(b)fortwohumans;and(c)forthreehumansmovingbehindthewallofaclosedroom.conÞnedspace,apersonthatmovestowardsWi-Viwilleventuallyhavetomoveawayorstop.Second,thebrightnessofthelinetyp-icallyindicatesdistance.Notethatforthesamespatialangle,onemaybecloseorfarfromWi-Vi.Hence,somelargeanglesappearbrightordimdependingonthepartofthetracewelookat.Athirdobservationisthatasthenumberofhumansincreases,itbecomeshardertoseparatethem.Theproblemisthatthecurvedlinesarefuzzybothduetoresidualnoiseandthefactthatahumancanmovehisbodypartsdifferentlyashemoves.Forexample,wav-ingwhilemovingmakesthelinessigniÞcantlyfuzzierasin8(a3).Finally,ourexperimentsareconductedinmultipath-richindoorenvironments.Thus,theresultsinFig.8showthatWi-Viworksinthepresenceofmultipatheffects.ThisisbecausethedirectpathfromamovinghumantoWi-Viismuchstrongerthanindirectpathswhichbounceofftheinternalwallsoftheroom.Amovinghumanactslikealargeantenna.Inordertoblockthedirectpath,thehumanbodymustbeobstructedbyapillaroralargepieceoffurniture,andstayobstructedforthedurationofWi-ViÕsmeasurements. Wenotethattheexperimentsinthispaperwereperformedinscenarioswheretheseparatorishomogeneouswall(e.g.,concrete,wooden,glass,etc.).Theremightbescenariosinwhichtheseparatorisnon-homogeneous(e.g.,theÞeldofviewofWi-ViÕsdirectionalantennacapturesasideofawallandaglasswindow),whichmaycausesomeindirectpathstobestrongerthanthedirectpath.Inthiscase,Wi-Viwillstilldetectamoving7.4AutomaticDetectionofMovingHumansWeareinterestedinevaluatingwhetherWi-Vicanusethespa-tialvariancedescribedin¤5.2toautomatethedetectionofmovinghumans.Asintheprevioussection,werunourexperimentsinthesameconferenceroomsdescribedin¤7.3.Again,wepositionWi-Visuchthatitfacesawallthathasneitheradoornorawindow.Foreachofourexperiments,weaskanumberofhumansbetween0and3fromourvolunteerstoentertheroomandmoveatwill.Eachexperimentlastsfor25secondsexcludingthetimerequiredforiterativenulling.Weperformeachexperimentwithadifferentsubsetofsubjects,andconductatotalof80experiments,withequalnumberofexperimentsspanningthecasesof0,1,2,and3movinghumans.Weprocessthecollectedtracesofßineandcomputethespatialvarianceasdescribedin¤5.2.Fig.9showstheCDFs(cumulativedistributionfunctions)ofthespatialvariancefortheexperimentsrunwitheachnumberofmov-inghumans:0,1,2,and3.Weobservethefollowing:Thespatialvarianceprovidesagoodmetricfordistinguishingthenumberofmovinghumans.Inparticular,thevarianceincreasesasthenumberofhumansinvolvedineachexperimentincreases.ThisisalsoevidentfromtheÞguresin8,whereonecanvisually objectbutmayhaveerrorsintrackingtheangleofthemovementorpredict-ingthenumberofmovinghumans. 83 -1500 -1000 -500 0 500 1000 1500 0 2 4 6 8 10 12 14 Matched Output Time (Seconds) (a)OutputofmatchedÞlter. -1.5 -1 -0.5 0 0.5 1 1.5 0 2 4 6 8 10 12 14 Mapped Symbols Time (Seconds) Bit Ô0Õ Bit Ô1Õ (b)Decodedbits.Figure7GestureDecodinginWi-Vi.TheÞgureshowshowWi-Vide-codesthegesturesofFig.5.(a)showstheoutputofthematchedÞlterstep.(b)showstheoutputofthepeakdetector.ThesequencebitÔ0Õ,whereasthesequence1,1representsbitÔ1Õ.backward.BecauseofthestructureofthesignalshowninFig.5,thetwomatchedÞltersaresimplyatriangleabovethezeroline,andaninvertedtrianglebelowthezeroline.Wi-ViappliestheseÞltersseparatelyonthereceivedsignal,thenaddsuptheiroutput.Fig.7showstheresultsofapplyingthematchedÞltersonthereceivedsignalinFig.5.NotethatthesignalafterapplyingthematchedÞlterslooksfairlysimilartoaBPSKsignal,whereapeakabovethezerolinerepresentsaÔ1ÕbitandatroughbelowthezerolinerepresentsaÔ0Õbit.(Though,inWi-Vi,ourencodingissuchthatapeakoratroughaloneonlyrepresentshalfabit.)Next,Wi-Viusesastandardpeakdetectortodetectthepeaks/troughsandmatchthemtothecorrespondingbits.Fig.7showstheidentiÞedpeaksandthedetectedbitsforthetwo-bitmessageinFig.5.7.IMPLEMENTATIONANDVALUATIONInthissection,wedescribeourimplementationandtheresultsofourexperimentalevaluation.7.1ImplementationWebuiltWi-ViusingUSRPN210softwareradios[8]withSBXdaughterboards.ThesystemusesLP0965directionalantennas[3],whichprovideagainof6dBi.ThesystemconsistsofthreeUS-RPsconnectedtoanexternalclocksothattheyactasoneMIMOsystem.TwooftheUSRPsareusedfortransmitting,andoneforreceiving.MIMOnullingisimplementeddirectlyintotheUHDdriver,sothatitisperformedinreal-time.Post-processingusingthesmoothedMUSICalgorithmisperformedontheobtainedtracesofßineinMatlabR2012aunderUbuntu11.10ona64-bitmachinewithInteli7processor.Matlabalreadyhasabuilt-inandhighlyoptimizedsmoothedMUSICimplementation.Processingtracesof25-secondlengthtookonaverage1.0564spertrace,withastandarddeviationof0.2561s.WeimplementstandardWi-FiOFDMmodulationintheUHDcode;eachOFDMsymbolconsistsof64subcarriersincludingtheDC.Thenullingprocedurein¤4isperformedonasubcarrierbasis.Thechannelmeasurementsacrossthedifferentsubcarriersarecom-binedtoimprovetheSNR.SinceUSRPscannotprocesssignalsinreal-timeat20MHz,wereducedthetransmittedsignalbandwidthto5MHzsothatournullingcanstillruninrealtime.Finally,theemulatedantennaarraywastakenover0.32seconds.Thecollectedsamplesduringthisdurationwereaveragedintoanantennaarrayofsize100,whichwasprovidedasaninputtothesmoothedMUSICalgorithm.7.2ExperimentalSetupMostofourexperimentswereruninoneofÞcebuildingusingtwodifferentconferencerooms.Theroomshavestandardfurniture:tables,chairs,boards,etc.Theinteriorwallsofthebuildingare6-inchhollowwallssupportedbysteelframeswithsheetrockontop.TheÞrstconferenceroomis74meters;thesecondis117meters.Wealsoconductedsomeexperimentsinasecondbuildingonourcampuswhichhas8-inchconcretewalls.Theexperimentswereconductedwitheighthumansubjects,threewomenandÞvemen,ofdifferentheightsandbuilds.Forthetrackingexperiments,weaskedthesubjectstoenteraroom,closethedoor,andmoveatwill.Thethrough-wallgestureexperimentswereperformedwithfoursubjects(onewomanandthreemen).Thepersonswereshownthegesturesinadvanceandtriedthemafewtimes.Then,eachofthementeredtheroomseparatelyandperformedthegestures.Theexperimentsarerepeatedindifferentlocationsindifferentrooms,andindifferentlocationsineachroom.7.3MicroBenchmarksFirst,wewouldliketogetabetterunderstandingoftheinforma-tioncapturedbyWi-Vi,andhowitrelatestothemovingobjects.Werunexperimentsintwoconferenceroomsinourbuilding.Bothconferenceroomshave6hollowwallssupportedbysteelframeswithsheetrockontop.Inalloftheseexperiments,wepositionWi-Vionemeterawayfromawallthathasneitheradoornorawindow.Foreachofourexperiments,weaskanumberofhumansbetween1and3toentertheroom,closethedoor,andmoveatwill.Wi-Viperformsnullinginrealtimeandcollectsatraceofthesignals.Weperformeachexperimentwithadifferentsubsetofoursubjects.WeprocessthecollectedtracesusingthesmoothedMUSICalgorithmasdescribedin¤5.2.Fig.8showstheoutputofWi-Viinthepresenceofone,two,orthreehumansmovinginaclosedroom.ConsidertheplotswithonehumaninFigs.8(a).BesidestheDC,thegraphsshowonefuzzycurvedline.Thelinetracksthespatialangleofthemovinghuman.ComparetheseÞgureswiththesetofÞguresin8(b),whichcapturetwomovinghumans.In8(b),wecandiscerntwocurvedlinesthattracktheangularmotionofthesehumanswithrespecttoWi-Vi.Ifwetakeaverticallineatanytime,inanyofthetwo-humanÞgures,weseeatmosttwobrightlines,besidestheDC.Thisisbecause,intheseÞgures,atanypointintime,thereareatmosttwomovingbodiesintheroom.Letuszoominontheintervalal1s,2s]in8(b1).Duringthisinterval,weseeonlyonecurvedline.Thishastwopos-sibleinterpretations:eitheroneofthetwopeoplestoppedmovingorhe/shewastoodeepinsidetheroomthatwecouldnotcapturehis/hersignal.Aswemoveto8(c),theÞguresgetfuzziersincewehavemorepeoplemovinginthesamearea.Howeverthegen-eralobservationscarrytotheseÞgures.SpeciÞcally,wecanidentifythepresenceofthreehumansfromobservingmultipleintervalsinwhichwecandiscernthreecurvedlines.Forexample,considertheintervalal1.8s,2.5in8(c1);itshowstwolineswithpositiveanglesandonewithanegativeangle.TheselinesindicatethattwopeoplearemovingtowardsWi-Vi,whileonepersonismovingaway.Onecanalsomakemultipleobservationsbasedontheshapeofthelines.First,apositiveanglemeansthehumanismovingtowardWi-Vi,whileanegativeanglemeansthatheismovingaway.Thevalueofthatangledependsontheorientationofthehumanandthedirectionofmotion.Eachlinelookslikeawavebecause,givena 82 workswellinpractice.Asexplainedearlier,movinghumansappearascurvedlinesinthe2-Dfunctionfunction!,n].Anyhumancanbeonlyatonelocationatanypointintime.Thus,atanypointintime,thelargerthenumberofhumans,thehigherthespatialvariance.Thespatialvarianceiscomputedasfollows.First,Wi-Vicomputesthespatialcentroidasafunctionoftime:time:n]=90%!="90!á20loglog!,n],(7)whereA![!,n]isgivenbyEq.6.Itthencomputesthespatialvari-anceas:VARRn]=90%!="90!2á20loglog!,n]!C[n]2(8)Thisvarianceisthenaveragedoverthedurationoftheexperimenttoreturnonenumberthatdescribesthespatialvarianceintheroomforthedurationofthemeasurement.Wi-Viusesatrainingsetandatestingsettolearnthethresholdsthatseparatethespatialvariancescorrespondingto0,1,2,or3humans.Thetestingandtrainingexperimentsareconductedindifferentrooms.In¤7.4,weevalu-atethisschemeandmeasureitsabilityatautomaticallycapturethenumberofmovinghumans.6.THROUGHNICATIONForahumantotransmitamessagetoacomputerwirelessly,shetypicallyhastocarryawirelessdevice.Incontrast,Wi-Vicanen-ableahumanwhodoesnotcarryanywirelessdevicetocommu-nicatecommandsorshortmessagestoareceiverusingsimpleges-tures.Wi-VidesignatesapairofgesturesasaÔ0ÕbitandaÔ1Õbit.Ahumancancomposethesegesturestocreatemessagesthathavedif-ferentinterpretations.Additionally,Wi-VicanevolvebyborrowingotherexistingprinciplesandpracticesfromtodayÕscommunicationsystems,suchasaddingasimplecodetoensurereliability,orre-servingacertainpatternofÔ0ÕsandÔ1Õsforpacketpreambles.Atthisstage,Wi-ViÕsinterfaceisstillverybasic,yetwebelievethatfutureadvancesinthrough-walltechnologycanrenderthisinter-facemoreexpressive.Below,wedescribethegesture-basedcommunicationchannelthatweimplementedwithWi-Vi.6.1GestureEncodingAtthetransmitterside,theÔ0ÕandÔ1Õbitsmustbeencodedusingsomemodulationscheme.Wi-Viimplementsthisencodingusinggestures.Onecanenvisionawidevarietyofgesturestorepresentthesebits.However,inchoosingourencodingwehaveimposedthreeconditions:1)thegesturesmustbecomposableÐi.e.attheendofeachbit,whetherÔ0ÕorÔ1Õ,thehumanshouldbebackinthesameinitialstateasthestartofthegesture.Thisenablesthepersontocomposemultiplesuchgesturestosendalongermessage.2)ThegesturesmustbesimplesothatahumanÞndsiteasytoperformthemandcomposethem.3)Thegesturesshouldbeeasytodetectanddecodewithoutrequiringsophisticateddecoders,suchasmachinelearningclassiÞers.Giventheaboveconstraints,wehaveselectedthefollowingges-turestomodulatethebits:aÔ0Õbitisastepforwardfollowedbyastepbackward;aÔ1Õbitisastepbackwardfollowedbyastepfor-ward.ThismodulationissimilartoManchesterencoding,whereaÔ0Õbitisrepresentedbyafallingedgeoftheclock,(i.e.,anin-creaseinthesignalvaluefollowedbyadecrease,)andaÔ1Õbitisrepresentedbyarisingedgeoftheclock,(i.e.,areductioninsig- Figure5GesturesasdetectedbyWi-Vi.TheÞgureshowsasequenceoffoursteps:stepforward,stepbackward,stepbackward,stepforward.Forwardstepsappearastrianglesabovethezeroline;backwardstepsappearasinvertedtrianglesbelowthezeroline.Eachpairofstepsrepresentsagesture/bit:theÞrsttworepresentbitÔ0Õ,thesecondtworepresentbitÔ1Õ. !"#$% !"#$% (a)Forward(b)Backward(c)SlantedFigure6GesturesasAngles.ÕsmagnitudeandsignasdeÞnedin¤5.1.In(a),thesubjecttakesonestepforward;theemulatedantennaarrayÕsnormalformsanangleof90withthelinefromthehumantoWi-Vi.BecausethevectorofthemotionandthevectorfromthehumantoWi-Viareinsamedirection,ispositive;hence,itis+90.In(b),thesubjecttakesastepbackward,and90degrees.In(c),thesubjectdoesnotexactlyknowwheretheWi-Videviceis,soheperformsthestepstowardsthewall,withoutorientinghimselfdirectlytowardWi-Vi.NotethatthevectorofmotionandthevectorfromthehumantoWi-Viareinthesamedirection;ispositive.However,duetotheslantedorientation,itisnow+60(ratherthan+90nalvaluefollowedbyanincrease)[2].Thesegesturesaresimple,composableandeasytodecodeasweshowin¤6.2.Fig.5showsthesignalcapturedbyWi-Vi,attheoutputofthesmoothedMUSICalgorithmforeachofthesetwogestures.TakingastepforwardtowardstheWi-Videviceproducesapositiveangle,whereastakingastepbackwardproducesanegativeangle.Theex-actvaluesoftheproducedanglesdependonwhetherthehumanisexactlyorientedtowardsthedevice.Recallthattheangleisbe-tweenthevectororthogonaltothemotionandthelineconnectingthehumantotheWi-Videvice,anditssignispositivewhenthehu-manismovingtowardWi-ViandnegativewhenthehumanmovesawayfromWi-Vi.AsshowninFig.6,ifthehumanisdirectlyori-entedtowardsthedevice,thetwoanglesare+90and-90.IfthehumandoesnotknowtheexactlocationoftheWi-Videviceandsimplystepsinitsgeneraldirection,theabsolutevalueoftheangleissmaller,buttheshapeofthebitismaintained.6.2GestureDecodingDecodingtheabovegesturesisfairlysimpleandfollowsstan-dardcommunicationtechniques.SpeciÞcally,Wi-ViÕsdecodertakesasinputinput!,n].Similartoastandarddecoder[16],Wi-ViappliesamatchedÞlteronthissignal.However,sinceeachbitisacombinationoftwosteps,forwardandbackward,Wi-ViappliestwomatchedÞlters:oneforthestepforwardandoneforthestep 81 fromthehumantoWi-Viisgettingsmaller.)Around1.8s,thepersoncrossesinfrontoftheWi-Videvice,atwhichtimehisan-glebecomeszero.From1.8sto3s,thepersonismovingawayfromWi-Vi,andhence,hisangleisnegative.Buttheabso-lutevalueoftheangleisdecreasing.At3,thepersonturnsandstartsmovinginward,causingtheangletogobacktowardzero,butthesignalbecomesweakerasheisnowrelativelyfarfromtheWi-Vireceiver.5.2TrackingMultipleHumansInthissection,weshowhowWi-Viextendsitstrackingproce-duretomultiplehumans.Ourpreviousdiscussionaboutusinghu-manmotiontoemulateanantennaarraystillholds.However,eachhumanwillemulateaseparateantennaarray.SinceWi-Vihasasingleantenna,thereceivedsignalwillbeasuperpositionoftheantennaarraysofthemovinghumans.Inparticular,insteadofhav-ingonecurvedlineasinFig.3(b),atanytime,therewillbeasmanycurvedlinesasmovinghumansatthatpointintime.However,withmultiplehumans,thenoiseincreasessigniÞ-cantly.Ononehand,eachhumanisnotjustoneobjectbecauseofdifferentbodypartsmovinginalooselycoupledway.Ontheotherhand,thesignalreßectedoffallofthesehumansiscorrelatedintime,sincetheyallreßectthetransmittedsignal.Thelackofinde-pendencebetweenthereßectedsignalsisimportant.Forexample,thereßectionsoftwohumansmaycombinesystematicallytodimeachotheroversomeperiodoftime.Theproblemofdisentanglingcorrelatedsuper-imposedsignalsiswellstudiedinsignalprocessing.Thebasicapproachforprocess-ingsuchsignalsreliesonthesmoothedMUSICalgorithm[31,39].SimilartothestandardantennaarrayprocessinginEq.4,smoothedMUSICcomputesthepowerreceivedalongaparticulardirection,whichwecallcall!,n]becauseitestimatesthesamefunctioninEq.4butinmannermoreresilienttonoiseandcorrelatedsig-nals[34].Foragivenantennaarrayyn],...,h[n+w])ofsizeMUSICÞrstcomputesthecorrelationmatrixmatrixn]:R[n]=E[hhH],(5)whereHreferstothehermitian(conjugatetranspose)ofthevector.Itthenperformsaneigendecompositionofofn]toremovethenoiseandkeepthestrongesteigenvectors,whichinourcasecorrespondtothefewmovinghumans,aswellastheDCvalue.Forexample,inthepresenceofonlyonehuman,MUSICwouldproduceonemaineigenvector(inadditiontotheDCeigenvector).Ontheotherhand,if2or3humanswerepresent,itwoulddiscover2or3eigenvectorswithlargeeigenvalues(inadditiontotheDCeigenvector).MUSICpartitionstheeigenvectormatrixmatrixn]into2subspaces:thesignalsignaln]andthenoisespacespacen],wherethesignalspaceisthespanofthesignaleigenvectors,andthenoisespaceisthespanofthenoiseeigenvectors.MUSICthenprojectsalldirectionsonthenullspace,thentakestheinverse.ThiscausestheÕscorrespondingtotherealsignals(i.e.,movinghumans)tospike.Moreformally, Interestingly,evenwhenthedirectionofmotionisperpendiculartothelineconnectingthepersontothedevice,Wi-Viregistersthismotion(notehowtheDClineismuchwiderat5thanat0).ThisisbecauseEq.4approximatesWi-Viasamonostaticradar,i.e.,itsimpliÞesthemodelbyassumingallantennasareco-located.Amoredetailedmodelthataccountsforthefactthattheantennasarenotcompletelyco-locatedshowsthatforatrajectorytobeinvisible(i.e.,coincidewiththeDCline)twoconditionshavetohold:(1)thepersonmovesonanellipsewhosefociaretheÞrsttransmitantennaandthereceiveantenna,(2)shemovesonanellipsewhosefociarethesecondtransmitantennaandthereceiveantenna.However,thelocusofsuchmotionisdiscontinuous. Figure4Wi-Vitracksthemotionoftwohumans.TheÞgureshowshowthepresenceoftwohumanstranslatesintotwocurvedlineswhoseanglesvaryintime,andonestraightlinewhichcorrespondstotheDC.MUSICcomputesthepowerdensityalongeachangles &Kk=1||&w=1e"j2! n](i,k)||2.(6)isthetotalnumberofnoiseeigenvectors.IncomparisontotheconventionalMUSICalgorithmdescribedabove,smoothedMUSICperformsanadditionalstepbeforeitcom-putesthecorrelationmatrix.Itpartitionseacharrayofsizeoverlappingsub-arraysofsize.Itthencomputesthecor-relationmatricesforeachofthesesub-arrays.Finally,itcombinesthedifferentcorrelationmatricesbysummingthemupbeforeper-formingtheeigendecomposition.TheadditionalstepperformedbysmoothedMUSICisintendedtode-correlatesignalsarrivingfromspatiallydifferententities.SpeciÞcally,bytakingdifferentshiftsforthesameantennaarray,reßectionsfromdifferentbodiesgetshiftedbydifferentamountsdependingonthedistanceandorientationofthereßector,whichhelpsde-correlatingthem[31].Fig.4showstheresultofapplyingsmoothedMUSIConthesig-nalcapturedfromtwomovinghumans.SimilartoFig.3(b),they-axiscorrespondstotheangle,andthex-axiscorrespondstotime.Asbefore,thezerolinecorrespondstoDC.Atanypointintime,weseesigniÞcantenergyattwoangles(besidestheDC).Forexample,attime0.5s,bothhumanshavenegativeanglesand,hence,aremovingawayfromWi-Vi.Between1sand2s,onlyoneangleispresent.Thismaybebecausetheotherhumanisnotmovingorhe/sheistoofarinsidetheroom.Again,from2sto3s,weseebothhumans,onemovingtowardsthedeviceandtheothermovingaway(sinceonehasapositiveanglewhiletheotherhasanegativeangle).Onepointisworthemphasizing:thesmoothedMUSICalgorithmisconceptuallysimilartothestandardantennaarraybeamformingdiscussedin¤5.1;bothapproachesaimatidentifyingthespatialangleofthesignal.However,byprojectingonthenullspaceandtakingtheinversenorm(asdescribedinEq.6),MUSICachievessharperpeaks,andhenceisoftentermedasuper-resolutiontech-nique[34].BecausesmoothedMUSICissimilartoantennaarraybeamforming,itcanbeusedeventodetectasinglemovingobject,i.e.,thepresenceofasingleperson.Infact,Fig.3(b)wasgeneratedbythesmoothedMUSICalgorithm.Finally,toenableWi-Vitoautomaticallydetectthenumberofhumansinaclosedroom,oneoptionistotrainamachinelearningclassiÞerusingimageslikethoseinFig.3(b)andFig.4.Wedis-covered,however,thatasimpleheuristicbasedonspatialvariance Plottingthemagnitudeofof!,n]asopposedtoto!,n]givesthesameÞgurebutwithmorenoise.Thisisbecause,unlikestandardbeamforming,theMUSICalgorithmdoesnotincursigniÞcantsidelobeswhichwouldotherwisemaskpartofsignalreßectedfromdifferentobjects. 80 Figure2TimesamplesasAntennaArrays.Wi-Vigroupsconsecutivetimesamplesintooverlappingwindowsofsizew,thentreatseachwindowwn]...h[n+w]asanantennaarray.Thisallowsittotrackthedirectionofamovingobjectwithrespecttothereceiver.Wi-Vi,however,avoidsusinganantennaarrayfortworeasons:First,inordertoobtainanarrowbeamandhenceachieveagoodresolution,oneneedsalargeantennaarraywithmanyantennaele-ments.Thiswouldresultinabulkyandexpensivedevice.Second,sinceWi-VieliminatestheßasheffectusingMIMOnulling,addingmultiplereceiveantennaswouldrequirenullingthesignalateachofthem.Thiswouldrequireaddingmoretransmitantennas,thusmakingthedeviceevenbulkierandmoreexpensive.TocapturethebeneÞtsofanantennaarraywhileavoidingitsdrawbacks,Wi-Vileveragesatechniquecalledinversesyntheticapertureradar(ISAR).ISARexploitsthemovementofthetar-gettoemulateanantennaarray.Existingsystemswhichusean-tennaarrayscapturethesignalreßectedoffatargetfromspatiallyspacedantennasandprocessesthisinformationtoidentifythedi-rectionofthetargetwithrespecttothearray.Incontrast,inISAR,thereisonlyonereceiveantenna;hence,atanypointintime,thereceivercapturesasinglemeasurement.However,asthetargetmoves,he/shesamplesthereceivedsignalatsuccessivelocationsinspace,asifwehadareceiveantennaateachofthesepoints.Fur-thermore,becauseofchannelreciprocity,successivetimesamplesreceivedbyWi-Vicorrespondtosuccessivespatiallocationsofthemovingtarget.Hence,Wi-Vieffectivelyreceivesintimewhatanantennaarraywouldreceiveinspace.Bytreatingconsecutivetimesamplesasspatialsamples,Wi-Vicanemulateanantennaarrayanduseittotrackmotionbehindthewall.Inwhatfollows,weformalizetheabovediscussion.LetLetn]bethesignalsamplereceivedbyWi-Viatadiscretetimepoint.De-ÞnethespatialangleastheanglebetweenthelineconnectingthehumantoWi-Viandthenormaltothemotion,asshowninFig.1(b).NotethatthesignofispositivewhenthevectorfromthehumantoWi-Viandthevectorofthemotionareinthesamedirection,andnegativewhenthesetwovectorsareinoppositedirections.Weareinterestedincomputingcomputing!,n],afunctionthatmeasuresthesignalalongthespatialdirectionattime.Tocomputethisvalue,Wi-ViÞrstprocessesthereceivedsamplestoremovetheef-fectofthetransmittedsignal,andobtainthechannelasafunc-tionoftime,i.e.,i.e.,n]=y[n]/x[n].Toemulateanantennaarrayofsize,Wi-Viconsidersconsecutivechannelmeasurementsmeasurementsn]...h[n+w],asshowninFig.2.Wi-Vithencomputescomputes!,n]byapplyingstandardantennaarrayequations[34]asfollows: isthewavelength,andisthespatialseparationbetweensuccessiveantennasinthearray.Atanypointintime,thevaluethatproducesthehighestvalueinin!,n]willcorrespondtothedirectionalongwhichtheobjectismoving.Tocomputee!,n]fromtheaboveequation,weneedtoestimate,theantennaspacingintheemulatedarray.Sincehumanmotionemulatestheantennasinthearray,,whereisWi-ViÕs istwicetheone-wayseparationtoaccountfortheround-triptime. !"# ! " # !"$ ! " $ !"% ! " !"& " & ! " % ExperimentalSetup Wi-ViÕsoutputFigure3Wi-VitracksasinglepersonÕsmotion.(a)showstheexperi-mentalsetupofatrialwhichconsistedofasinglepersonmovingaroundinaconferenceroom.(b)showshowWi-Viisabletotrackthemotionofthepersonbycomputingthevariationoftheinverseangleofarrivalwithtime,time,!,n]for!in["90#,90samplingperiod,andisthevelocityofthemotion.Ofcourse,Wi-Vidoesnotknowtheexactspeedatwhichthehumanismov-ing.However,therangeofspeedsthathumanshaveinaconÞnedroomisfairlynarrow.Hence,wecansubstituteavalueformatchescomfortablewalking(ourdefaultis1m/s[10]).Notethaterrorsinthevalueoftranslatetoanunderestimationoranoverestimationoftheexactdirectionofthehuman.Errorsinve-locity,however,donotpreventWi-Vifromtrackingthatthehumanismovingcloser(i.e.,angleispositive)ormovingawayfromtheWi-Videvice(angleisnegative).Inotherwords,becausewedonotknowtheexact,wecannotpinpointthelocationofthehuman,butwecantrackher/hisrelativemovements.Fig.3showsresultsfromoneofourexperiments.Inparticular,3(a)showsadiagramofthemovement,and3(b)plotsthemag-nitudeofof!,n](indB)asaheatmap.TherearetwolinesinFig.3(b):theÞrstoneisazeroline,whichrepresentstheDC(i.e.,theaverageenergyfromstaticelements).Thislineispresentre-gardlessofthenumberofmovingobjects.Second,thereisacurvedlinewithachangingangle.Thislinetracksthehumanmotion.0seconds,thepersonstartsmovingtowardstheWi-Videvice.Asaresult,thespatialangleispositiveanddecreasing.(Itispositivebecausethevectorofmotionandthelinefromthehu-mantoWi-Viareinthesamedirection,anditisdecreasingbecausetheabsoluteanglebetweenthenormalonthemotionandtheline Forexample,inoneofourexperiments,Wi-ViestimatedthehumanÕsdi-rectionofmotionat30whentheactualdirectionwas40butshewasmovingataspeedaround1.2m/sRecallthatnullingmitigatesthesereßectionssothattheydonotsaturatethereceiverÕsADC,enablingWi-Vitoregistertheminutechannelvariationsduetomovingobjectsbehindthewall.However,minusculeerrorsinchan-nelestimatesduringthenullingphasewouldstillberegisteredasaresidualDCbyWi-Vi. 79 Algorithm1PseudocodeforWi-ViÕsNulling INITIALNULLING:ChannelEstimationTxant.1sends;RxreceivesTxant.2sends;RxreceivesPOWERBOOSTING:TxantennasboostpowerTxant.1transmits,Txant.2transmitsITERATIVENULLING:repeatRxreceivesy;resievenresres TxantennastransmitconcurrentlyConverges Intheidealcase,wheretheestimatesareperfect,thereceivedsignalreswouldbeequaltozero.Hence,bytheendofthisphaseWi-Vihaseliminatedthesig-nalsreßectedoffallstaticobjectsaswellasthedirectsignalfromthetransmitantennastothereceiveantenna.Ifnoobjectmoves,thechannelwillcontinuebeingnulled.However,sinceRFreßec-tionscombinelinearlyoverthemedium,ifsomeobjectmoves,itsreßectionswillstartshowingupinthechannelvalue.PowerBoosting.Simplynullingstaticreßections,however,isnotenoughbecausethesignalsduetomovingobjectsbehindthewallaretooweak.Say,forexample,theßasheffectwas30to40dBabovethepowerofreßectionsoffmovingobjects.Eventhoughweremovedtheßasheffect,wecanhardlydiscernthesignalduetomovingobjectssinceitwillbeimmersedinthereceiverÕshardwarenoise.Thus,wenextboostthetransmittedsignalpower.Notethatbecausethechannelhasalreadybeennulled,i.e.,res0,thisincreaseinpowerdoesnotsaturatethereceiverÕsADC.However,itincreasestheoverallpowerthattraversesthewall,and,hence,improvestheSNRofthesignalduetotheobjectsbehindthewall.IterativeNulling.Afterboostingthetransmitpower,residualreßectionswhichwerebelowtheADCquantizationlevelbecomemeasurable.SuchreßectionsfromstaticobjectscancreatesigniÞ-cantclutterinthetrackingprocessifnotremoved.Toaddressthisissue,Wi-Viperformsaprocedurecallediterativenulling.Atahighlevel,theobjectiveissimple:weneedtonullthesignalagainaf-terboostingthepowertoeliminatetheresidualreßectionsfromstaticobjects.Thechallenge,however,isthatatthisstage,wecan-notseparatelyestimatethechannelsfromeachofthetwotransmitantennassince,afternulling,weonlyreceiveacombinedchannel.Wealsocannotremovethenullingandre-estimatethechannels,becauseafterboostingthepower,withoutnulling,theADCwouldHowever,Wi-Vicanleveragethefactthaterrorsinthechannelestimatesaremuchsmallerthanthechannelestimatesthemselves,andusethisobservationtoreÞneitsestimates.SpeciÞcally,byas-sumingthattheestimateforisaccurate(i.e.,),Eq.1isleftwithonlyoneunknownvariable.Bysolvingforthisun- InourUSRPimplementation,weboostthepowerby12dB.ThisvalueislimitedbytheneedtostaywithinthelinearrangeoftheUSRPtransmitter.Afternulling,wecanalsoboostthereceivegainwithoutsaturatingthere-ceiverÕsADC.Onaverage,wenull42dBofthesignal,whichallowsalargeboostinthereceivegain.knownvariable,weobtainabetterestimateof.Inparticular,thenewestimateresSimilarly,byassumingthattheestimateforisaccurate(i.e.,),wecansolveEq.1foraÞnerestimateforres Therefore,Wi-ViiteratesbetweenthesetwostepstoobtainÞnerestimatesforboth,untilthetwoestimatesverge.Thisiterativenullingalgorithmconvergesexponentiallyfast.Inparticular,intheappendix,weprovethefollowinglemma:Assumethat ,then,afteriiterations,resres AfewpointsareworthnotingaboutWi-ViÕsproceduretoelimi-natetheßasheffect:BesidesremovingthewallÕsreßection,italsoremovesreßec-tionsreceivedfromotherstationaryobjectsbothinfrontofandbehindthewall,suchasthetableonwhichtheradioismounted,theßoor,theradiocaseitself,etc.Inaddition,itremovesthedi-rectsignalfromthetransmittingantennastoourreceiveantenna.NotethatthedirectchannelsbetweenWi-ViÕstransmitanten-nasanditsreceiveantennaaresigniÞcantlyattenuatedbecauseWi-Viusesdirectionaltransmitandreceiveantennasfocusedto-wardsthewall(andawayfromthedirectpath).Wi-ViÕsnullingalgorithmprovidesa42dBmeanreductioninsignalpower,asshownin¤7.6.ThisreductionissufÞcienttore-movetheßasheffectfromawiderangeofwallstructuresinclud-ingsolidwooddoors,6hollowwalls,andmostindoorconcretewalls.Further,sinceWi-Viusesdirectionalantennasfocusedontheimagedwall,thedirectsignalfromthetransmitantennastoWi-ViÕsreceiveantennaisweakerthanintypicalMIMOsys-tems,andbecomesnegligibleafternulling.Nullingcanbeperformedinthepresenceofobjectsmovingbe-hindthewall;itcanalsobeperformedinthepresenceofobjectsmovinginfrontofthewallaslongastheyareoutsidetheÞeldofviewofWi-ViÕsdirectionalantennas.Becausenullingismathe-maticallyequivalenttosubtraction,thepresenceofsuchmovingobjectsleadstoasmalladditiveconstantattheoutputofWi-Viafternulling.Suchadditiveconstantsdonotpreventlatertrack-ingofmovingobjects.5.IDENTIFYINGANDRACKINGNowthatwehaveeliminatedtheimpactofstaticobjectsintheenvironment,wecanfocusontrackingmovingobjects.Wewillrefertomovingobjectsashumanssincetheyaretheprimarysub-jectsofinterestforourapplication;however,oursystemisgeneral,andcancaptureothermovingbodies.Below,weÞrstexplainhowWi-Vitracksthemotionofasinglehuman.Wethenshowhowtoextendourapproachtotrackmultiplemovinghumans.5.1TrackingaSingleHumanMostpriorthrough-wallsystemstrackhumanmotionusinganantennaarray.TheysteerthearrayÕsbeamtodeterminethedirec-tionofmaximumenergy.ThisdirectioncorrespondstothesignalÕsspatialangleofarrival.Bytrackingthatangleintime,theyinferhowtheobjectmovesinspace. Forexample,wehavesuccessfullyexperimentedwithtrackinganiRobotCreaterobot. 78 limitstheirdetectioncapabilities.Hence,mostofthesesystemsaredemonstratedeitherinsimulation[28],orinfreespacewithnoob-struction[21,23].Theonesdemonstratedwithanobstructionusealow-attenuationstandingwall,anddonotworkacrosshigherat-tenuationmaterialssuchassolidwoodorconcrete[29,30].Wi-Visharestheobjectivesofthesedevices;however,itintroducesanewapproachforeliminatingtheßasheffectwithoutwidebandtrans-mission.Thisenablesittoworkwithconcretewallsandsolidwooddoors,aswellasfullyclosedrooms.TheonlyattemptwhichweareawareofthatusesWi-Fisignalsinordertoseethroughwallswasmadein2012[13].Thissystemrequiredboththetransmitterandareferencereceivertobeinsidetheimagedroom.Furthermore,thereferencereceiverintheroomhastobeconnectedtothesameclockasthereceiveroutsidetheroom.Incontrast,Wi-Vicanperformthrough-wallimagingwithoutaccesstoanydeviceontheothersideofthewall.Gesture-basedinterfaces.Today,commercialgesture-recognitionsystemsÐsuchastheXboxKinect[9],NintendoWii[4],etc.Ðcanidentifyawidevarietyofgestures.Theacademiccommunityhasalsodevelopedsystemscapableofidentifyinghumangesturesei-therbyemployingcameras[24]orbyplacingsensorsonthehumanbody[15,20].Recentworkhasalsoleveragednarrowbandsignalsinthe2.4GHzrangetoidentifyhumanactivitiesinline-of-sightusingmicro-Dopplersignatures[21].Wi-Vi,however,presentstheÞrstgesture-basedinterfacethatworksinnon-line-of-sightscenar-ios,andeventhroughawall,yetdoesnotrequirethehumantocarryawirelessdeviceorwearasetofsensors.Infraredandthermalimaging.SimilartoWi-Vi,thesetechnolo-giesextendhumanvisionbeyondthevisibleelectromagneticrange,allowingustodetectobjectsinthedarkorinsmoke.TheyoperatebycapturinginfraredorthermalenergyreßectedofftheÞrstob-stacleinline-of-sightoftheirsensors.However,camerasbasedonthesetechnologiescannotseethroughwallsbecausetheyhaveveryshortwavelengths(fewmtosub-mm)[37],unlikeWi-Viwhichemployssignalswhosewavelengthsare12.5cm.3.WVERVIEWWi-Viisawirelessdevicethatcapturesmovingobjectsbehindawall.ItleveragestheubiquityofWi-Fichipsetstomakethrough-wallimagingrelativelylow-power,low-cost,low-bandwidth,andaccessibletoaverageusers.Tothisend,Wi-ViusesWi-FiOFDMsignalsintheISMband(at2.4GHz)andtypicalWi-Fihardware.Wi-Viisessentiallya3-antennaMIMOdevice:twooftheanten-nasareusedfortransmittingandoneisusedforreceiving.Italsoemploysdirectionalantennastofocustheenergytowardthewallorroomofinterest.Itsdesignincorporatestwomaincomponents:1)theÞrstcomponenteliminatestheßashreßectedoffthewallbyperformingMIMOnulling;2)thesecondcomponenttracksamov-ingobjectbytreatingtheobjectitselfasanantennaarrayusingatechniquecalledinverseSAR.Wi-Vicanbeusedinoneoftwomodes,dependingontheuserÕschoice.Inmode1,itcanbeusedtoimagemovingobjectsbehindawallandtrackthem.Inmode2,ontheotherhand,Wi-Vifunctionsasagesture-basedinterfacefrombehindawallthatenableshumanstocomposemessagesandsendthemtotheWi-Vireceiver.Insections4-6,wedescribeWi-ViÕsoperationindetail.4.ELIMINATINGTHE Thelongerthewavelengthofanelectromagneticwaveis,theloweritsattenuationis[35].Infraredandthermalimagingdevicesemploysignalswhosewavelengthsareveryclosetovisiblelight;hence,theydonotpene-tratebuildingmaterialssuchaswoodorconcrete.Directionalantennashaveaformfactorontheorderofthewavelength.AtWi-Fifrequencies,thiscorrespondstoapproximately12cm. BuildingMaterials 2.4GHz Glass 3dB SolidWoodDoor1.75inches 6dB InteriorHollowWall6inches 9dB ConcreteWall18inches 18dB ReinforcedConcrete 40dB Table1ÑOne-WayRFAttenuationinCommonBuildingMaterialsat2.4GHz[1].Inanythrough-wallsystem,thesignalreßectedoffthewall,i.e.,theßash,ismuchstrongerthananysignalreßectedfromobjectsbehindthewall.ThisisduetothesigniÞcantattenuationwhichelectromagneticsignalssufferwhenpenetratingdenseobstacles.Table1showsafewexamplesoftheone-wayattenuationexpe-riencedbyWi-Fisignalsincommonconstructionmaterials(basedon[1]).Forexample,aone-waytraversalofastandardhollowwalloraconcretewallcanreduceWi-Fisignalpowerby9dBand18dBrespectively.Sincethrough-wallsystemsrequiretraversingtheob-stacletwice,theone-wayattenuationdoubles,leadingtoan18-36dBßasheffectintypicalindoorscenarios.Thisproblemisexacerbatedbytwootherparameters:First,theactualreßectedsignalissigniÞcantlyweakersinceitdependsbothonthereßectioncoefÞcientaswellasthecross-sectionoftheob-ject.Thewallistypicallymuchlargerthantheobjectsofinterest,andhasahigherreßectioncoefÞcient[11].Second,inadditiontothedirectßashcausedbyreßectionsoffthewall,through-wallsys-temshavetoeliminatethedirectsignalfromthetransmittothereceiveantenna,whichissigniÞcantlylargerthanthereßectionsofinterest.Wi-Viusesinterferencenullingtocancelboththewallreßectionsandthedirectsignalfromthetransmittothereceivean-tenna,henceincreasingitssensitivitytothereßectionsofinterest.4.1NullingtoRemovetheFlashRecentadvancesshowthatMIMOsystemscanpre-codetheirtransmissionssuchthatthesignalreceivedataparticularantennaiscancelled[36,17].PastworkonMIMOhasusedthispropertytoenableconcurrenttransmissionsandnullinterference[26,22].Weobservethatthesametechniquecanbetailoredtoeliminatetheßasheffectaswellasthedirectsignalfromthetransmittothereceiveantenna,therebyenablingWi-Vitocapturethereßectionsfromobjectsofinterestwithminimalinterference.Atahighlevel,Wi-ViÕsnullingprocedurecanbedividedintothreephases:initialnulling,powerboosting,anditerativenulling,asshowninAlg.1.InitialNulling.Inthisphase,Wi-ViperformsstandardMIMOnulling.RecallthatWi-Vihastwotransmitantennasandonere-ceiveantenna.First,thedevicetransmitsaknownpreambleonitsÞrsttransmitantenna.Thispreambleisreceivedatthereceiveantennaas,whereisthechannelbetweentheÞrsttrans-mitantennaandthereceiveantenna.Thereceiverusesthissignalinordertocomputeanestimateofthechannel.Second,thedevicetransmitsthesamepreamble,thistimeonlyonitssecondan-tenna,andusesthereceivedsignaltoestimatechannelthesecondtransmitantennaandthereceiveantenna.Third,Wi-Viusesthesechannelestimatestocomputetheratio.Fi-nally,thetwotransmitantennastransmitconcurrently,wheretheÞrstantennatransmitsandthesecondtransmits.Therefore,theperceivedchannelatthereceiveris:res öh2""0(1) 77 stage2.Reßectionsoffstaticobjects,includingthewall,arenulledinthisstage.In¤4,wereÞnethisbasicideabyintroducingiterativenulling,whichallowsustoeliminateresidualßashandtheweakerreßectionsfromstaticobjectsbehindthewall.Second,howdoesWi-Vitrackmovingobjectswithoutanan-tennaarray?Toaddressthischallenge,weborrowatechniquecalledinversesyntheticapertureradar(ISAR),whichhasbeenusedformappingthesurfacesoftheEarthandotherplanets.ISARusesthemovementofthetargettoemulateanantennaarray.AsshowninFig.1,adeviceusinganantennaarraywouldcaptureatargetfromspatiallyspacedantennasandprocessthisinformationtoidentifythedirectionofthetargetwithrespecttothearray(i.e.,).Incon-trast,inISAR,thereisonlyonereceiveantenna;hence,atanypointintime,wecaptureasinglemeasurement.Nevertheless,sincethetargetismoving,consecutivemeasurementsintimeemulateanin-verseantennaarrayÐi.e.,itisasifthemovinghumanisimagingtheWi-Videvice.Byprocessingsuchconsecutivemeasurementsusingstandardantennaarraybeamsteering,Wi-Vicanidentifythespatialdirectionofthehuman.In¤5.2,weextendthismethodtomultiplemovingtargets.Additionally,Wi-Vileveragesitsabilitytotrackmotiontoen-ableathrough-wallgesture-basedcommunicationchannel.Specif-ically,ahumancancommunicatemessagestoaWi-Vireceiverviagestureswithoutcarryinganywirelessdevice.WehavepickedtwosimplebodygesturestorefertoÒ0ÓandÒ1Óbits.Ahumanbehindawallmayuseashortsequenceofthesegesturestosendames-sagetoWi-Vi.AfterapplyingamatchedÞlter,themessagesignallookssimilartostandardBPSKencoding(apositivesignalforaÒ1Óbit,andanegativesignalforaÒ0Óbit)andcanbedecodedbyconsideringthesignofthesignal.Thesystemenableslawenforce-mentpersonneltocommunicatewiththeirteamacrossawall,eveniftheircommunicationdevicesareconÞscated.WebuiltaprototypeofWi-ViusingUSRPN210radiosandeval-uateditintwoofÞcebuildings.Ourresultsareasfollows:Wi-Vicandetectobjectsandhumansmovingbehindopaquestructuralobstructions.Thisappliesto8concretewalls,6lowwalls,and1.75solidwoodendoors.AWi-Videvicepointedataclosedroomwith6hollowwallssupportedbysteelframescandistinguishbetween0,1,2,and3movinghumansintheroom.Computedover80trialswith8hu-mansubjects,Wi-Viachievesanaccuracyof100%,100%,85%,and90%respectivelyineachofthesecases.Inthesameroom,andgivenasinglepersonsendinggesture-basedmessages,Wi-Vicorrectlydecodesallmessagesper-formedatdistancesequaltoorsmallerthan5meters.Thede-codingaccuracydecreasesto75%atdistancesof8meters,andthedevicestopsdetectinggesturesbeyond9meters.For8vol-unteerswhoparticipatedintheexperiment,onaverage,ittookaperson8.8secondstosendamessageof4gestures.Incomparisontothestate-of-the-artultra-widebandsee-through-wallradar[27],Wi-Viislimitedintwoways.First,replacingtheantennaarraybyISARmeansthattheangularresolutioninWi-Vidependsontheamountofmovement.Toachieveanarrowbeamthehumanneedstomovebyabout4wavelengths(i.e.,about50cm).Second,incontrastto[27],wecannotdetecthu-mansbehindconcretewallsthickerthan8.ThisisduetoboththemuchlowertransmitpowerfromourUSRPsandtheresidualßashpowerfromimperfectnulling.Ontheotherhand,nullingtheßashremovestheneedforGHzbandwidth.Italsoremovesclutterfromallstaticreßectors,ratherthanjustonewall.Thisin-cludesotherwallsintheenvironmentsaswellasfurnitureinsideandoutsidetheimagedroom.Toreduceclutter,theempiricalre-sultsinpastworkaretypicallycollectedusingaperson-heightstandingwall,positionedeitheroutdoorsorinlargeemptyin- (a)AntennaArray(b)ISARFigure1AMovingObjectasanAntennaArray.In(a),anantennaarrayisabletolocateanobjectbysteeringitsbeamspatially.In(b),themovingobjectitselfemulatesanantennaarray;hence,itactsasaninversesyntheticaperture.Wi-Vileveragesthisprincipleinordertobeamformthereceivedsignalintime(ratherthaninspace)andlocatethemovingobject.doorspaces[27,41].Incontrast,ourexperimentsareinstan-dardofÞcebuildingswiththeimagedhumansinsideclosedfully-furnishedrooms.Contributions:Incontrasttopastworkwhichtargetsthemilitary,Wi-Viintroducesnovelsolutionstothesee-through-wallproblemthatenablenon-militaryentitiestousethistechnology.SpeciÞcally,Wi-ViistheÞrsttointroduceinterferencenullingasamechanismforeliminatingtheßasheffectwithoutrequiringwidebandspec-trum.ItisalsotheÞrsttoreplacetheantennaarrayatthereceiverwithanemulatedarraybasedonhumanmotion.ThecombinationofthosetechniquesenablessmallcheapdevicesthatoperateintheISMband,andcanbemadeaccessibletothegeneralpublic.Fur-ther,Wi-ViistheÞrsttodemonstrateagesture-basedcommunica-tionchannelthatoperatesthroughwallsanddoesnotrequirethehumantocarryanywirelessdevice.2.RELATEDWi-Viisrelatedtopastworkinthreemajorareas:Through-wallradar.Interestinthrough-wallimaginghasbeensurgingforaboutadecade[5].Earlierworkinthisdomainfocusedonsimulations[38,28]andmodeling[32,33].Recently,therehavebeensomeimplementationstestedwithmovinghumans[27,41,13].Thesepastsystemseliminatetheßasheffectbyisolatingthesignalreßectedoffthewallfromsignalsreßectedoffobjectsbe-hindthewall.Thisisolationcanbeachievedinthetimedomain,byusingveryshortpulses(lessthan1ns)[40,5]wherebythepulsereßectedoffthewallarrivesearlierintimethanthatreßectedoffmovingobjectsbehindit.Alternatively,itmaybeachievedinthefrequencydomainbyusingalinearfrequencychirp[11,27].Inthiscase,reßectionsoffobjectsatdifferentdistancesarrivewithdifferenttones.ByanalogÞlteringthetonethatcorrespondstothewall,onemayremovetheßasheffect.Thesetechniquesrequireultra-widebandwidths(UWB)oftheorderof2GHz[11,40].Sim-ilarly,through-wallimagingproductsdevelopedbytheindustry[5,7]hingeonthesameradarprinciples,requiringmultipleGHzofbandwidthandhencearetargetedsolelyatthemilitary.Asathrough-wallimagingtechnology,Wi-VidiffersfromalltheabovesystemsinthatitrequiresonlyfewMHzofbandwidthandoperatesinthesamerangeasWi-Fi.ItovercomestheneedforUWBbyleveragingMIMOnullingtoremovetheßasheffect.ResearchershaverecognizedthelimitationsofUWBsystemsandexploredthepotentialofusingnarrowbandradarsforthrough-walltechnologies[29,30].Thesesystemsignoretheßasheffectandtrytooperateinpresenceofhighinterferencecausedbyreßec-tionsoffthewall.TheytypicallyrelyondetectingtheDopplershiftcausedbymovingobjectsbehindthewall.However,theßasheffect 76 SeeThroughWallswithWi-Fi!FadelAdibandDinaKatabiMassachusettsInstituteofTechnology{fadel,dk}@mit.eduABSTRACTWi-Fisignalsaretypicallyinformationcarriersbetweenatrans-mitterandareceiver.Inthispaper,weshowthatWi-Ficanalsoextendoursenses,enablingustoseemovingobjectsthroughwallsandbehindcloseddoors.Inparticular,wecanusesuchsignalstoidentifythenumberofpeopleinaclosedroomandtheirrelativelocations.Wecanalsoidentifysimplegesturesmadebehindawall,andcombineasequenceofgesturestocommunicatemessagestoawirelessreceiverwithoutcarryinganytransmittingdevice.Thepaperintroducestwomaininnovations.First,itshowshowonecanuseMIMOinterferencenullingtoeliminatereßectionsoffstaticobjectsandfocusthereceiveronamovingtarget.Second,itshowshowonecantrackahumanbytreatingthemotionofahumanbodyasanantennaarrayandtrackingtheresultingRFbeam.Wedemon-stratethevalidityofourdesignbybuildingitintoUSRPsoftwareradiosandtestingitinofÞcebuildings.CategoriesandSubjectDescriptorsC.2.2[SystemsOrganization]:Computer-CommunicationsNetworks.H.5.2[InformationInterfacesandPresentation]:UserInter-faces-Inputdevicesandstrategies.KeywordsSeeingThroughWalls,Wireless,MIMO,Gesture-BasedUserInterface1.INTRODUCTIONCanWi-Fisignalsenableustoseethroughwalls?FormanyyearshumanshavefantasizedaboutX-rayvisionandplayedwiththeconceptincomicbooksandsci-Þmovies.ThispaperexploresthepotentialofusingWi-FisignalsandrecentadvancesinMIMOcommunicationstobuildadevicethatcancapturethemotionofhumansbehindawallandinclosedrooms.Lawenforcementper-sonnelcanusethedevicetoavoidwalkingintoanambush,andminimizecasualtiesinstandoffsandhostagesituations.Emergencyresponderscanuseittoseethroughrubbleandcollapsedstructures.Ordinaryuserscanleveragethedeviceforgaming,intrusiondetec-tion,privacy-enhancedmonitoringofchildrenandelderly,orper-sonalsecuritywhensteppingintodarkalleysandunknownplaces.Theconceptunderlyingseeingthroughopaqueobstaclesissim-ilartoradarandsonarimaging.SpeciÞcally,whenfacedwithanon-metallicwall,afractionoftheRFsignalwouldtraversethewall,reßectoffobjectsandhumans,andcomebackimprintedwithasignatureofwhatisinsideaclosedroom.Bycapturingthesere-ßections,wecanimageobjectsbehindawall.Buildingadevicethatcancapturesuchreßections,however,isdifÞcultbecausethePermissiontomakedigitalorhardcopiesofallorpartofthisworkforpersonalorclassroomuseisgrantedwithoutfeeprovidedthatcopiesarenotmadeordistributedforproÞtorcommercialadvantageandthatcopiesbearthisnoticeandthefullcitationontheÞrstpage.Copyrightsforcomponentsofthisworkownedbyothersthantheauthor(s)mustbehonored.Abstractingwithcreditispermitted.Tocopyotherwise,orrepublish,topostonserversortoredistributetolists,requirespriorspeciÞcpermissionand/orafee.Requestpermissionsfrompermissions@acm.org.August12Ð16,2013,HongKong,China.Copyrightisheldbytheowner/author(s).PublicationrightslicensedtoACM.Copyright2013ACM978-1-4503-2056-6/13/08...$15.00.signalpoweraftertraversingthewalltwice(inandoutoftheroom)isreducedbythreetoÞveordersofmagnitude[11].Evenmorechallengingarethereßectionsfromthewallitself,whicharemuchstrongerthanthereßectionsfromobjectsinsidetheroom[11,27].ReßectionsoffthewalloverwhelmthereceiverÕsanalogtodigitalconverter(ADC),preventingitfromregisteringtheminutevaria-tionsduetoreßectionsfromobjectsbehindthewall.ThisbehavioriscalledtheÒFlashEffect"sinceitisanalogoustohowamirrorinfrontofacamerareßectsthecameraÕsßashandpreventsitfromcapturingobjectsinthescene.SohowcanoneovercomethesedifÞculties?Theradarcom-munityhasbeeninvestigatingtheseissues,andhasrecentlyin-troducedafewultra-widebandsystemsthatcandetecthumansmovingbehindawall,andshowthemasblobsmovinginadimbackground[27,41](seethevideoat[6]forareference).TodayÕsstate-of-the-artsystemrequires2GHzofbandwidth,alargepowersource,andan8-footlongantennaarray(2.4meters)[12,27].Apartfromthebulkinessofthedevice,blastingpowerinsuchawidespectrumisinfeasibleforentitiesotherthanthemilitary.Therequirementformulti-GHztransmissionisattheheartofhowthesesystemswork:theyseparatereßectionsoffthewallfromreßec-tionsfromtheobjectsbehindthewallbasedontheirarrivaltime,andhenceneedtoidentifysub-nanoseconddelays(i.e.,multi-GHzbandwidth)toÞltertheßasheffect.Toaddresstheselimitations,aninitialattemptwasmadein2012touseWi-Fitoseethroughawall[13].However,tomitigatetheßasheffect,thispastproposalneedstoinstallanadditionalreceiverbehindthewall,andconnectthereceiversbehindandinfrontofthewalltoajointclockviawires[13].Theobjectiveofthispaperistoenableasee-through-walltech-nologythatislow-bandwidth,low-power,compact,andaccessibletonon-militaryentities.Tothisend,thepaperintroducesWi-Vi,see-through-walldevicethatemploysWi-Fisignalsinthe2.4GHzISMband.Wi-Vilimitsitselftoa20MHz-wideWi-Fichannel,andavoidsultra-widebandsolutionsusedtodaytoaddresstheßasheffect.Italsodisposesofthelargeantennaarray,typicalinpastsystems,andusesinsteadasmaller3-antennaMIMOradio.So,howdoesWi-VieliminatetheßasheffectwithoutusingGHzofbandwidth?WeobservethatwecanadaptrecentadvancesinMIMOcommunicationstothrough-wallimaging.InMIMO,mul-tipleantennasystemscanencodetheirtransmissionssothatthesig-nalisnulled(i.e.,sumsuptozero)ataparticularreceiveantenna.MIMOsystemsusethiscapabilitytoeliminateinterferencetoun-wantedreceivers.Incontrast,weusenullingtoeliminatereßectionsfromstaticobjects,includingthewall.SpeciÞcally,aWi-Videvicehastwotransmitantennasandasinglereceiveantenna.Wi-Viop-eratesintwostages.IntheÞrststage,itmeasuresthechannelsfromeachofitstwotransmitantennastoitsreceiveantenna.Instage2,thetwotransmitantennasusethechannelmeasurementsfromstage1tonullthesignalatthereceiveantenna.Sincewirelesssignals(in-cludingreßections)combinelinearlyoverthemedium,onlyreßec-tionsoffobjectsthatmovebetweenthetwostagesarecapturedin FilteringisdoneintheanalogdomainbeforethesignalreachestheADC.Wi-VistandsforWi-FiVision. 75