using non-intrusive environmental sensing for adls ... · itistheminimumactionrequiredfordailylife...
TRANSCRIPT
DOI: 10.4018/IJSI.2018100102
International Journal of Software InnovationVolume 6 • Issue 4 • October-December 2018
Copyright©2018,IGIGlobal.CopyingordistributinginprintorelectronicformswithoutwrittenpermissionofIGIGlobalisprohibited.
16
Using Non-Intrusive Environmental Sensing for ADLS Recognition in One-Person HouseholdLong Niu, Graduate School of System Informatics, Kobe University, Kobe, Japan
Sachio Saiki, Graduate School of System Informatics, Kobe University, Kobe, Japan
Masahide Nakamura, Graduate School of System Informatics, Kobe University, Kobe, Japan & RIKEN AIP, Tokyo, Japan
ABSTRACT
Thisarticledescribeshowpervasivesensingtechnologiesarepromisingforincreasingone-personhousehold(OPH),whereasystemmonitorsandassistsaresidenttomaintainhealthyliferhythm.Automaticrecognitionofactivitiesofdailyliving(ADLS)hasbeenahotresearchtopicinpervasivecomputing.However,mostexistingmethodshavelimitationsindevelopmentcost,privacyexposure,and inconvenience for residents. To cope with the limitations, this article presents a new ADLrecognitionsystemespeciallyforOPH.Tominimizethedevelopmentcostaswellasintrusionstouserandhouse,thesystemexploitsanIoT-basedenvironmentsensingdevice,calledautonomoussensorbox(sensorbox)whichcanautonomouslymeasure7kindsofenvironmentattributes.Thesystemthenappliesmachine-learningtechniquestopredict7kindsofADLS.Finally,thisarticleconductsanexperimentwithinanactualapartmentofasingleuser.TheresultshowsthattheproposedsystemachievestheaverageaccuracyofADLSrecognitionwithabout88%,bycarefullydevelopingthefeaturesofenvironmentattributes.
KEywORDSActivities of Daily Living, ADLS Recognition, Feature Engineering, Machine Learning, Non-Intrusive Environment Sensing, One-Person Household
INTRODUCTION
Thegrowingnumberofunmarriedpeopleandlatemarriagesindevelopedcountriesleadstoasocialissueofone-personhousehold(OPH).InJapan,thenumberofOPHincreasingrapidly.Itisestimatedthat37.4%householdswillbecomeOPHin2030(MinistryofHealth,LabourandWelfare,2010).InsevenstatesofUSA,thepercentageofOPHexceeds30.3%in2015(Statista,2017).InChina,therearemorethan60millionpeoplecurrentlylivingalone(Yeung&Cheung,2015).AccordingtoAsaoka,Fukuda&Yamazaki(2004)andFujinoetal.(2006),peopleinOPHeasilylosecontrolhealthyliferhythm,sincenooneelsecantakecareofthelivinginOPH.Sincethelossofhealthyliferhythmoftenleadstohealthdeterioration,itisessentialtomaintaintheliferhythmespeciallyinthecontextofOPH.Ingeneral,aliferhythmischaracterizedbyactivitiesofdailyliving(ADLs,forshort).TypicalADLsinOPHincludeeating,takingbath,sleeping,etc.IfthecycleofADLs
International Journal of Software InnovationVolume 6 • Issue 4 • October-December 2018
17
becomesverydifferentfromtheoneinahealthyliferhythm,theresidentislosinghis/herliferhythm.Tomaintaintheliferhythm,onehastokeeparegularrecordofADLs.However,keepingmanualrecordingrequiresstrongmindandpatience.
ToautomatetheADLrecordinginOPH,pervasivesensingtechnologiescombinedwithmachinelearningarequitepromising,becausetheycanrecognizeADLsfromautomaticallymeasureddata.TherehavebeenmanystudiesforADLrecognition.Someapproaches(e.g.,Fiore,Bodor,Drenner,Somasundaram&Papanikolopoulos,2008;Ouchi&Doi,2013)trytodirectlycapturethelivingusingcamera,ormicrophone.However,suchsystemsaretoointrusiveoftheuserinthesensethatthedailylivingisexposedasitis.Therearemanystudiesusingwearablesensors,and/orindoorpositioningsystemtorecognizeADLs(e.g.,Kusano,Muro,Hayashi,Harada&Shimakawa,2011;Peietal.,2013).However,thewearablesensorisintrusivetohumanbody,astheuseralwayshastowearthesensordeviceathome.Indeed,thehomeisaplacewheretheuserisfreefromtediousthings.Theindoorpositioningisintrusivetoahouse,inthesensethatsensorsandbeaconsmustbeinstalledintothehouseandobjects.Thisusuallycausesexpensivecostfordeploymentandmaintenance.
Toovercomethelimitations,weproposeanewsystemthatrecognizesADLsofOPHbasedonnon-intrusiveenvironmentalsensingwithmachinelearning.Intheproposedsystem,weexploitanIoT-basedenvironment-sensingdevice,calledautonomoussensorbox(wesimplycallSensorBox,hereinafter).SensorBoxhasbeendeveloped inourpreviouswork (Sakakibara,Saiki,Nakamura&Matsumoto,2016),andisdesignedtominimizetheeffortofdeploymentandoperation.Onceapowercableisconnected,SensorBoxautonomouslymeasuresseventypesofenvironmentattributes(temperature,humidity,light,sound,vibration,gaspressureandmotion)aroundthebox,andthenperiodicallyuploadsthedatatoacloudserver.Thus,alltheoperationsfordeploymentandmaintenanceareperformedwithouthumanintervention,orexpensiveinfrastructure.
AsSensorBoxismeasuring theenvironment inOPH, theproposedsystemalsorequires theinitialtraining,wheretheresidentmanuallyrecordsADLsusingadesignatedlifelogtool.Theinitialtrainingissupposedtobeperformedinseveraldays,toassociatelabelsofADLswiththesensordata.Intheproposedsystem,wedefinesevenbasicADLs(cooking,working,cleaning,takingbath,sleeping,eating,absence),whicharethemosttypicalADLsformaintainingtheliferhythm.Forthelabeleddataset,weapplysupervisedlearningalgorithmstoconstructamodelofADLrecognitionfor thehouse.For this,weperformcareful featureengineering todetermineessentialpredictorsthatwellexplainADLsinOPH.Furthermore,wetryseveraldifferentclassificationalgorithmstocomparetheperformance.
To evaluate the proposed system, we have deployed one SensorBox in an actual apartmentofasingleperson,andconductedanexperimentfortendays.ExperimentalresultsshowthattheaverageaccuracyofallthesevenADLswasaround87%withDecisionForestsupervisedlearning.TheaccuracyofsomespecificADLsachievedover90%.Fromthisresult,weconfirmedthattheproposedsystemachievesnon-intrusiveandpracticalADLrecognitioninOPH,usingSensorBox.
PRELIMINARy
Activities of Daily Living (ADL)ADLisaprofessionalwordoriginallyusedathospital.Itistheminimumactionrequiredfordailylifesuchassleeping,meal,takingbath,etc.itisusedasanindicatoroftheaginganddegreeofdisability.ThediscoveryandrecognitionofADLisanessentialfunctionofthesystemthatprovidesnecessaryassistanttotheresidentsofOPH.Basedontheresultsofthisprocess,theintelligentsystemcandecidewhichactiontotakeinordertosupporttheresidents’well-beingandunderstandresidents’liferhythmbasedontheregularrecordofADLs.
International Journal of Software InnovationVolume 6 • Issue 4 • October-December 2018
18
ADLs RecognitionSincetheneedofADLrecognitionisgreat,researchershavebeenstudyinganddevelopinganumberofmethodologiestotacklethisproblem.TheapproachestotheADLrecognitioncanbedividedroughlyintotwocategories,dependingonthetypeofcontextual informationanalyzed.Thefirstcategoryusesmultimediadatatakenbyvideocamerasormicrophonerecordings,tocapturethedailylivingdirectly.Thesecondcategoryusestime-seriesdatameasuredbyvarioussensors,includingaccelerometer,gyroscope,RFIDandpower-meterssensors.
Multimedia DataBrdiczka,Reignier&Crowley(2007)proposedasmarthomethattakesvideosofresidents,andprocessesthevideotorecognizeactivities.Althoughgeneralpeoplehavebeenresistedtotheat-homevideomonitoring(Hensel,Demiris&Courtney,2006),theacceptanceofthistechnologyinthehomeisincreasing.Ontheotherhand,processingthevideoiscomputationallyexpensive.Itreliesuponthefirsttrackingoftheresidentbeforethecorrectvideodatacanbecapturedandanalyzed.
Sensor DataSincetakingvideoandaudioexposestoomuchinformationofdailyliving,itisconsideredtobeintrusivetothelife.Therefore,itismoreappreciatedtousepassiveinformation.Hence,mostofthecurrentresearchinADLsrecognitionusesensordata.Researchershavefoundthatcombiningdifferenttypesofsensoriseffectiveforclassifyingdifferenttypesofactivities.
Kusanoetal.(2011)proposedasystemthatderivesliferhythmfromtrackingelderlymovementbyusingRFIDpositioningtechnology.TheyinstallmanyRFIDreadersonthefloorofahouse,andaskparticipantstowearslipperswithRFIDtags.Thereaderscaptureindoorlocationofresident.Thesystemreasonstheliferhythmoftheuserfromthetime-serieslocationdata.However, it isdifficult todeterminetheexactactivityusingthemovementhistory.Asaresult, theaccuracyofADLsrecognitionislow.
Tapia,Intille&Larson(2004)focusedoninteractionsofaresidentwithanobjectofinterestsuchasadoor,awindow,arefrigerator,akeyandamedicinecontainer.Theyinstalledstate-changesensorsondailyitemstocollecttheinteractiondata.
Philiposeetal. (2004)attachedanRFIDtagonevery item,andaskedaparticipant toweargloveswithanRFIDtagreader.Whentheparticipantisclosetotheitem,theinteractionisrecorded.
Peietal.(2013)combinedapositioningsystemandmotionsensorofasmartphonetorecognizehumanmovementsinnaturalenvironments.However,whenturningonthemotion-sensor,Wi-FiandGPSsimultaneously,thebatterydrainisveryhigh.Anotherproblemisthatausermaynotwanttocarrysmartphonealltimeathome,whichisthelimitationofcollectingdata.
SensorBoxSensorBoxisanIoTdevicewithmultipleenvironmentsensors,developedbyourresearchgroup(Sakakibaraetal.,2016).Itcanmeasuresevenenvironmentalattributesaroundthebox,whicharetemperature,humidity,lightingintensity,atmospherepressure,soundvolume,humanmotionandvibration.Itwasdesignedtominimizethecost-intensiveinfrastructureandconfigurationlabor.Onceconnectedtopowerandnetwork,SensorBoxautonomouslymeasuresenvironmentattributearoundtheboxanduploadsdetecteddatatocloudserver.
Challenges and Research GoalTheADLrecognitionhasbeenwidelystudiedforfewyears.BykeepingtrackofADLs,asmartpervasivesystemcanprovideremindersforresidents,aswellasreacttohazardoussituations(Wren&Tapia,2006).Mostofthesestudiesapplytoelderlypeople,cancerpatients,andordinaryfamilies.However,therearenotsomanystudiesforOne-PersonHousehold(OPH).Theuniquecharacteristics
International Journal of Software InnovationVolume 6 • Issue 4 • October-December 2018
19
ofOPHaretheresidentislivingalone,andisoftenbusytodoeverythingbyoneself.He/shedoesnotwanttochangetheownwayofliving,orpayforexpensivesystemjustformonitoringADLs.
AsmentionedinSection2,therearemanyexistingsystemsthatusewearablesensors,object-embeddedsensors,orindoorpositioningsystems.However,weconsideritdifferentforpeopleinOPHtoacceptthesetechnologies,becausetheyaretooexaggeratedandintrusivefortheirlife.Wecaneasilyimaginethatmostresidentswillforgetorgiveupwearingthesensor,sincethehomeistheplacewheretheresidentmakeoneselfcomfortable.Althoughlabsorcompaniescanmanagethelarge-scaleequipment,itisstilltooexpensivetodeployinOPH.
Ourresearchgoalistominimizesuchlimitationsoftheconventionalapproaches,andtoachievehigh-qualityADLrecognitionofOPH.
OUTLINE OF PROPOSED SySTEM
In order to achieve the research goal, we propose a new ADL recognition system for OPH. Tominimizetheintrusionsandthecost,theproposedsystemjustreliesontheenvironmentalsensingbytheautonomoussensorbox(SensorBox)(Sakakibaraetal.,2016).Figure1showsthearchitectureoftheproposedsystem.Usingthefigure,weexplaintheproposedsystemfromlefttoright.
First,wesetupthesystemwithinatargetOPH.Wedeployasingle(ormultipleifnecessary)SensorBoxinapositionwhereADLsarewellobservedasenvironmentmeasures.Wealsoinstallasoftware,calledLifeLogger,onuser’sPC.Toapplysupervisedmachine-learningalgorithms,theproposedsystemrequirestrainingdataattheinitialphaseofoperation.Forthis,LifeLoggerisusedtoattachcorrectlabelsofADLs(aslifelog)totheenvironmentsensingdata.
Then,thesystembeginstocollecttime-seriesdata.SensorBoxuploadsthemeasureddatatoMongoDBinacloudserver,whereasLifeLoggerinsertsthelifelogintoMySQLintheclouddata.
Finally,thesystemjoinsthetwotime-seriesdatawiththetimestamptoformthetrainingdata.WeapplymachinelearningtothetrainingdatatoconstructapredictionmodelofADLrecognition.
DATA COLLECTION
Environment SensingTobeabletodetectADLbyanalyzingnon-intrusiveenvironmentattributes,thetargetattributesmustbesensitivetothechangingofresident’sADL.ConsideringthattherangeofsensibleisonlyaroundtheSensorBox,theboxshouldbeputonwhereresident’sADLisfrequentlyconducted.However,thelayoutofeachhouseandlivingstayofeachsinglearedifferentthingsfromOPHtoOPH.HencethemostsuitablepositionofSensorBoxisalsodifferentfromOPHtoOPH.
Figure 1. Architecture of proposed system
International Journal of Software InnovationVolume 6 • Issue 4 • October-December 2018
20
SensorBoxmeasures sevenenvironmentattribute,whichare temperature,humidity, lightingintensity,atmospherepressure,soundvolume,humanmotionandvibration,inevery10seconds.Figure2showsthescreenshotoftherawsensorsdatathatbemodifiedtoJSONformattext.AndFigure3showsthescreenshotofanapplication,whichvisualizescollectedrawsensorsdataoncloudservice.ThefirstgraphofFigure3showsthechangingofbrightnesswithtime.Andwecaneasilyfindthatthosechangingisman-made.
Activity LabelingDuringinitialseveraldays,theresidentneedstoinputcorrectlabelsofADLs,sothatthesystemcanlearntheADLsfromtheenvironmentsensingdata.Forthis,weasktheresidenttouseLifeLogger.Figure 4 shows the user interface of LifeLogger. As shown in the figure, LifeLogger has eightbuttons,eachofwhichcorrespondstoanADL.WhentheresidentstartsanADL,he/shejustpressesthecorrespondingbuttontorecordthecurrentADL.Basedontherelevantstudies(Fujinoetal.,2006;Fioreetal.,2008),wehavechoseneighttypesoftypicalADLs(sleep,eat,cook,workingat
Figure 2. Raw sensors data
Figure 3. Sensor data visualization
International Journal of Software InnovationVolume 6 • Issue 4 • October-December 2018
21
PC,clean,bath,absenceandother),andregisteredtheminLifeLogger.Whentheresidentpressesabutton,thesystemrecordsthelabel,andstoresitinMySQLinacloudserver.
Integrate Environment Sensing and Activity Labeling DataForsupervisedlearning,thesystemneedstotrainingdatawhichhavethecorrespondencebetweentheADLsanddatainadvance.Inordertoestablishtrainingdata,weintegratethetwotime-seriesdatacollectedbySensorBoxandLifeLoggerbyjoiningbasedontimestamp.Sincedatalabelledas‘other’wasbeyondthescopeoftheADLrecognition,thosenoisedatamustbefiltered.Table1showsthetrainingdata.
ESTABLISH MACHINE LEARNING RECOGNITION MODEL
Analysis Activity-sensitive Environment Sensing SensorsForaccurateADLrecognition,itisessentialtoidentifywhatenvironmentalvaluesinthesensingdatawellpredicttheADL.FromthesevenenvironmentalattributesofSensorBox,weonlychoose
Table 1. Training data
DateTime vibration light motion gaspressure temerature Humidity sound activityID
2017/2/193:33:02 495 1 0 98.8 13.33 35.84 50.15 5
2017/2/193:33:12 494 1 0 98.8 13.33 36.04 0 5
2017/2/193:33:22 494 1 0 98.8 13.33 36.04 51.62 5
2017/2/193:33:32 494 1 0 98.8 13.33 36.04 0 5
Figure 4. Lifelogger tool
International Journal of Software InnovationVolume 6 • Issue 4 • October-December 2018
22
temperature,humidity,lightingintensity,soundvolumeandhumanmotion,sincetherestattributes(vibration and atmosphere pressure) seem irrelevant to the target ADLs (i.e., sleep, eat, cook,workingatPC,clean,bath,absenceandother).Forexample,fromthefourgraphsofFigure3showsthechangingoffourenvironmentattributesinoneday,wecanseeonlythegraphofgasPressureisasmoothcurve.AccordingtothisfigurewemakejudgementsthatthegasPressureisalmostnotaffectedbytheresident’sADL.
Feature EngineeringFeaturevalueisdatathatiseffectivetoidentifytheADLs.Inourstudy,wegetthefeaturevaluefromtrainingdata,asthefollowingprocess.
First,wedeterminethesizeoftime-window.Toenhancethefeatureofthetime-seriesdata,weaggregatetherawdatawithinthesametime-windowintoonedata.Forthis,thewindowsizeaffectstheaccuracy.Ifthesizeistoolarge,thewindowislikelytocontaindifferentactivities.Iftoosmall,thewindowwillnotcontainsufficientdatatoreasonandpredictanADL.Hence,wetest3variationsof1,2and3minutes.InordertofacilitatethediscussioninSection4,weusesymbols(‘A’,‘B’,‘C’)topresentdifferentdatasetswithdifferentsizeoftime-window.ThedetailisA:1minute,B:2minutesandC:3minutes.
Finally,foreachofthefiveenvironmentattributeschosen,wedetermineanaggregationfunction.Anaggregationfunctionaggregatesallthedatawithinthesametime-window.TypicalaggregationfunctionincludesMAX,MIN,AVG,STDEV,andsoon.Basedonthenatureofeachenvironmentattribute,wecarefullychooseanappropriatefunction.Weneedtoapplydifferentaggregationfunctiontoeachenvironmentattribute.Byanalyzingall thetest,wewillfindtheoptimalcombinationofaggregationfunctionfromtest.However,ifwewanttotestallsituations,weneedtoconducthundredsroundsoftest,whichwilltakealotoftime.Toeffectivelytestsallcasesoffunctioncombination,weusedatoolcalledPICT(Jacek,2009).PICTgeneratesacompactsetofparametervaluechoicethatrepresentthetestcasesyoushouldusetogetcomprehensivecombinatorialcoverageofyourparameters.Table2showsthe9casesofcombinationgeneratedbyPICT.
Establish Recognition ModelForthedevelopedfeaturesofthetrainingdata,weapplymachine-learningalgorithm,inordertoconstructapredictmodelforADLrecognition.Weusepopularclassificationalgorithms,includingLogistic Regression, Decision Forest and Neural Network. By using these algorithms, we haveconstructedpredictionmodelsthatclassifiesgivenenvironmentsensordataintooneofthesevenADLs.
Table 2. Nine groups of aggregation function
Group light motion temperature humidity sound
G1 MIN MAX AVE AVE AVE
G2 MAX MAX STD STD STD
G3 AVE AVE STD STD MAX
G4 MAX AVE AVE AVE MAX
G5 MIN AVE AVE STD AVE
G6 AVE AVE AVE AVE STD
G7 MAX MAX STD AVE AVE
G8 AVE MAX AVE AVE AVE
G9 MIN AVE STD STD STD
International Journal of Software InnovationVolume 6 • Issue 4 • October-December 2018
23
EVALUATE OF EXPERIMENT
Experiment SetupWedeployedtheproposedsysteminarealapartmentofasingleresident.AsshowninFigure5,theapartmentisanordinaryapartmentinJapan,consistingofabed/livingroom,abathroomandakitchen.WehaveplacedonesingleSensorBoxintheKitchenroom,sothatSensorBoxcanobserveADLsofresidentwell.ThepositionofSensorBoxismarkedbyaredpininFigure5.Atotalof45,693rowsoflabeledsensordata,whichdonotincludedatalabeledwith‘other’,becollectedduring10dayswithintheapartment.
ResultWehaveestablished81recognitionmodelsbasedonthe3sizeoftime-windows,9combinationsofAggregationFunctionsand3machine-learningalgorithms,andtestedthosemodelsbytrainingandlearningthecollecteddata.Inthissubsection,weshowthetestresultofallmodels,theAverageAccuracyofeachtrainedADLsrecognitionmodel.Accuracymeasuresthegoodnessofaclassificationmodelastheproportionoftrueresulttototalcases.Averageaccuracywhichistheaverageofeachaccuracyperclass(sumofaccuracyforeachclasspredicted/numberofclass).
Inordertofacilitatetheobservation,wedivideallthemodelsintothreetablesbythesizeoftime-window,Table3,4and5,anddrawabargraphforeachtable,Figure6,7and8.
Forthetables,eachcolumnstandsforthemethodofdataprocessinginfeatureengineeringandisidentifiedbyanamethatcontains2characters,acapitalletterandanumber.Thecapitallettermeans the size of time-windows, as mentioned in subsection Feature Engineering. The numbermeansthecombinationofaggregationfunctions’groupnumberinTable2.Inoneexample,inTable3,A4 stands for the sizeof time-window is1minutesandutilizedcombinationofAggregationFunctionsisG4onthedataprocessoffeatureengineering.Therowoftableisidentifiedbythenameofmachine-learningalgorithm.Eachcellshowstheaverageaccuracyofeachmodel.Inoneexample,therecognitionmodel’saverageaccuracyexceeded88.10%inthesituationwherethesizeoftime-windowis3minutes,utilizedcombinationofaggregationfunctionisG4andalgorithmisthemulticlassdecisionforest.
Forthethreebargraphs,theverticalaxisisaverageaccuracyandthelateralaxisstandsforthecolumnofrelevanttable.Thecolorofbarstandsforthealgorithms,therowofrelevanttable.
Bycomparingtheaverageaccuracyofmodelsondifferentsizeoftime-window,suchasthebluebarsofA1B1andC1,wecanseethatthesizeoftime-windowhasslightlyinfluenceontheaverageaccuracyofmodel.Bycomparingtheresultofmodelsondifferentmethodsoffeatureengineeringinonegraph,wecanseethatthemodelshavesignificantlydifferentperformancewiththecombination
Figure 5. Testbed apartment
International Journal of Software InnovationVolume 6 • Issue 4 • October-December 2018
24
Figure 6. Visualization of Table 3
Table 3. All results of time-windows on 1 minute
Multiclass Algorithm A1 A2 A3 A4 A5 A6 A7 A8 A9
NeuralNetwork 84.36% 83.98% 84.27% 83.95% 83.96% 84.61% 84.28% 85.27% 84.23%
DecisionForest 85.83% 86.56% 86.24% 83.91% 84.43% 86.78% 87.54% 87.83% 86.41%
LogisticRegression 83.51% 85.23% 85.21% 83.16% 85.73% 83.18% 85.92% 82.05% 85.22%
Table 4. All results of time-windows on 2 minute
Multiclass Algorithm B1 B2 B3 B4 B5 B6 B7 B8 B9
NeuralNetwork 84.10% 84.92% 84.87% 85.14% 83.08% 85.66% 84.45% 85.14% 85.09%
DecisionForest 84.52% 87.68% 86.18% 87.56% 84.62% 84.59% 87.30% 83.72% 86.49%
LogisticRegression 83.57% 85.79% 85.14% 83.05% 85.57% 83.92% 85.94% 82.53% 85.71%
Table 5. All results of time-windows on 3 minute
Multiclass Algorithm C1 C2 C3 C4 C5 C6 C7 C8 C9
NeuralNetwork 84.62% 85.14% 85.07% 84.99% 82.11% 85.44% 84.25% 84.29% 84.70%
DecisionForest 85.92% 86.44% 85.10% 88.10% 85.25% 84.62% 87.10% 84.66% 86.21%
LogisticRegression 83.00% 86.77% 84.99% 82.59% 85.44% 84.29% 85.99% 82.59% 86.44%
International Journal of Software InnovationVolume 6 • Issue 4 • October-December 2018
25
ofaggregationfunctions.Byobservingthethreegraphoftable,Figure6,7,8,wecanseethemodelsutilizedmulticlassdecisionforest,representedbyorangebars,havebetterperformancethanothermodels.
EvaluationInthesubsection,weevaluatetherelationshipbetweenthe3factorsandaccuracyofrecognitionADLsbasedonthepartofrepresentativedatawhichisselectedfromhugevolumesofexperimentaldata.
First,theeffectoftime-windowonaccuracy.Table6showsthreerecognitionmodels’accuracyof some ADLs (cooking, sleeping, and eating). The three recognition models are utilized same
Figure 7. Visualization of Table 4
Figure 8. Visualization of Table 5
International Journal of Software InnovationVolume 6 • Issue 4 • October-December 2018
26
aggregationfunctionandalgorithmexpectthesizeoftime-window.Fromtheresults,wecanseethattheaccuracyofthreeADLsisslightlyaffectedbythesizeoftime-window,andthe2minutesislikelyanappropriatevaluebecauseaccuracyofrecognitionADLsishigherwhenthesizeoftime-windowis2minutes,expectrecognitioneating.Althoughthechangeofaccuracyisalsoslightlycauseofthechangeofsizeoftime-windowsisverysmall,theresultconfirmstheviewoftheabove-mentionedthatthesizeoftime-windowshouldnottoolargeandtoosmall.
Second,theeffectofaggregationfunctiononaccuracy.Table7showsthreerecognitionmodels’accuracy of recognition some ADLs (cleaning, sleeping, and going out). The three recognitionmodelsutilizedthesamesizeoftime-windowandalgorithmexceptthecombinationofaggregationfunctionforthefiveenvironmentattributes.Fromtheresult,wecanseeaggregationfunctionhasagreatinfluenceontheaccuracyofeachADLsrecognition.AsfortheG4,theaccuracyofgoingoutrecognitionisonly17.6%,whichis45%lessthanG8,buttheaccuracyofpredictsleepingisalmostequaltoG8.Forthethreemodels,theaccuracyofsleepingrecognitionachievesthehighestvaluewithusingthecombinationofG7.ButthemodelofG7performsbadlyonrecognitioncleaningandgoingout.ForG8,theaccuracyofpredictsleepingislessthanG7,buttheperformanceofpredictcleaningandgoingoutismuchbetterthanG7.Hence,fromthecompareswecanseethatsystemneedtoapplydifferentcombinationsoffunctiontoeachADLs,inotherwords,thefeaturevalueofADLsisdifferentwitheachother.
Last,theeffectofalgorithmsonaccuracy.Table8showsthreerecognitionmodels’accuracyofrecognitionsomeADLs(cooking,sleeping,andeating).Thosemodelsutilizedthesamesizeoftime-windowandcombinationofaggregationfunctionsexceptalgorithms.Fromtheresults,wecanseethatDecisionForestperformsbetterthanLogisticRegressionontherecognitionthreeADLs.FortheNeuralNetwork,ithasthebestperformanceonpredictsleepingandeating,butithasworstperformanceonpredictcooking.Forthereasonofabnormalresult,wethinkthatthisamountofdata
Table 6. Compare the accuracy of three models on different time-windows
Time windows Cooking Sleeping Eating
1minute(A) 86.00% 89.90% 54.10%
2minutes(B) 89.70% 89.90% 55.70%
3minutes(C) 87.80% 89.70f% 55.90%
Table 7. Compare the accuracy of three models on different aggregate functions
Aggregate Function Cleaning Sleeping Absence
G4 47.10% 73.00% 17.60%
G7 39.40% 95.50% 21.10%
G8 62.60% 72.70% 62.40%
Table 8. Compare the accuracy of three models on different algorithms
Multiclass Algorithm Cooking Sleeping Eating
LogisticRegression(LR) 59.60% 44.30% 41.30%
DecisionForest(DF) 67.50% 72.70% 62.60%
NeuralNetwork(NN) 20.20% 85.30% 93.50%
International Journal of Software InnovationVolume 6 • Issue 4 • October-December 2018
27
isnotsuitableforneuralnetwork.Becausetheamountofcookingandeating’sdataisonly3.8%and5.2%ofthetotaldata,lessthan400elements.Forthesethreemodels,themodelofDecisionForestperformsmorerobustthanotherswiththelimitednumberoftrainingdata.
CONCLUSION
Inthispaper,wehaveproposedanewsystemthatautomaticallyrecognizesADLinOPH.ConsideringthecharacteristicsofOPH,theproposedsystemexploitsonlyenvironmentalsensingbySensorBox.Thisminimizesthecostofdeployment,aswellastheintrusiontotheresidentandthehouse.Toevaluatetheproposesystem,wedeployedthesysteminanactualapartmentofasingleresident,andcollectedsensordataandlifelog(ascorrectlabels)for10days.Throughsupervisedlearningwithcareful featureengineering,wewereable toconstructpractically feasiblemodelsofseven typesofADLs.TheaverageaccuracyofallADLsachievedmore than88%.Forsleepingrecognition,theaccuracyofrecognitionachievedmorethan90%.Andtheinfluenceofsizeoftime-windows,aggregationfunctionandmachine-learningalgorithmsontheaccuracyofrecognitionADLs.
As for the futurework,weevaluate theproposedsystem inmultiplehouses to seehow thelearningprocessesvariesfromonehousetoanother.Moreover,wewanttovalidateiftheproposedeighttypesofADLsareenoughtocapturetheliferhythmsinOPH.Finally,developingservicesthatactuallyassisthealthyliferhythmusingtherecognizedADLsisourlong-termgoal.
ACKNOwLEDGMENT
ThisresearchwaspartiallysupportedbytheJapanMinistryofEducation,Science,Sports,andCulture[Grant-in-AidforScientificResearch(B)(No.16H02908,No.15H02701),ChallengingExploratoryResearch(15K12020)],andTateishiScienceandTechnologyFoundation(C)(No.2177004).
International Journal of Software InnovationVolume 6 • Issue 4 • October-December 2018
28
REFERENCES
Asaoka, S., Fukuda, K., & Yamazaki, K. (2004). Effects of sleep–wake pattern and residential status onpsychologicaldistressinuniversitystudents.Sleep and Biological Rhythms,2(3),192–198.doi:10.1111/j.1479-8425.2004.00138.x
Brdiczka,O.,Reignier,P.,&Crowley, J.L. (2007).Detecting IndividualActivities fromVideo inaSmartHome.InKES 2007 and XVII Italian Workshop on Neural Networks Conference on Knowledge-based Intelligent Information and Engineering Systems: Part I(pp.363-370).Springer-Verlag.doi:10.1007/978-3-540-74819-9_45
Fiore,L.,Bodor,R.,Drenner,A.,Somasundaram,G.,&Papanikolopoulos,N.(2008).Multi-CameraHumanActivityMonitoring.Journal of Intelligent & Robotic Systems,52(1),5–43.doi:10.1007/s10846-007-9201-6
Fujino,Y.,Iso,H.,Tamakoshi,A.,Inaba,Y.,Koizumi,A.,Kubo,T.,&Yoshimura,T.(2006).AprospectivecohortstudyofshiftworkandriskofischemicheartdiseaseinJapanesemaleworkers.American Journal of Epidemiology,164(2),128–135.doi:10.1093/aje/kwj185PMID:16707650
Hensel,B.K.,Demiris,G.,&Courtney,K.L.(2006).DefiningObtrusivenessinHomeTelehealthTechnologies:AConceptualFramework.Journal of the American Medical Informatics Association,13(4),428–431.doi:10.1197/jamia.M2026PMID:16622166
JacekCzerwonka.(2009).PICT3.3User’sGuide.RetrievedFebruary04,2009,fromhttp://amibugshare.com/pict/help.html
Kusano,K.,Muro,H.,Hayashi,T.,Harada,F.,&Shimakawa,H.(2011).DerivationofLifeRhythmfromTracingElderlyMovement.InThe 10th Forum on Information Technology(Vol.10,pp.891-892).IEICE.
MinistryofHealth.LabourandWelfare.(2010)SupportsforIndependenceinOrdertoStabilizeLivingandSociety.Annual Health, Labour and Welfare Report 2008-2009.Retrieved2010,fromhttp://www.mhlw.go.jp/english/wp/wp-hw3/
Ouchi,K.,&Doi,M.(2013).Smartphone-basedMonitoringSystemforActivitiesofDailyLivingforElderlyPeopleandTheirRelativesetc. InPervasive and ubiquitous computing adjunct publication (pp.103–106).ACM.doi:10.1145/2494091.2494120
Pei,L.,Guinness,R.,Chen,R.,Liu,J.,Kuusniemi,H.,Chen,Y.,&Kaistinen,J.et al.(2013).HumanBehaviorCognition Using Smartphone Sensors. Journal of Sensors, 13(2), 1402–1424. doi:10.3390/s130201402PMID:23348030
Philipose,M.,Fishkin,K.P.,Perkowitz,M.,Patterson,D.J.,Fox,D.,Kautz,H.,&Hahnel,D.(2004).Inferringactivities from interactionswithobjects. International Journal of IEEE Pervasive Computing,3(4),50–57.doi:10.1109/MPRV.2004.7
Sakakibara,S.,Saiki,S.,Nakamura,M.,&Matsumoto,S.(2016).IndoorEnvironmentSensingServiceinSmartCityusingAutonomousSensorBox.In15th IEEE/ACIS International Conference on Computer and Information Science(pp.885-890).IEEE.doi:10.1109/ICIS.2016.7550871
Statista.(2017).Percentageofsingle-personhouseholdsintheU.S.2015,bystate.Retrieved2017,fromhttps://www.statista.com/statistics/242284/percentage-of-single-person-households-in-the-us-by-state/
Tapia,E.M.,Intille,S.S.,&Larson,K.(2004).ActivityRecognitionintheHomeUsingSimpleandUbiquitousSensors.InA.Ferscha&F.Mattern(Eds.),Pervasive Computing(pp.158–175).SpringerBerlinHeidelberg.doi:10.1007/978-3-540-24646-6_10
Wren,C.R.,&Tapia,E.M.(2006).TowardScalableActivityRecognitionforSensorNetworks.InM.Hazas,J.Krumm,&T.Strang(Ed.),Location- and Context-Awareness:Second International Workshop(pp.168-185).SpringerBerlinHeidelberg.doi:10.1007/11752967_12
Yeung,W. J.,&Cheung,A. (2015).LivingAlone:One-personhouseholds inAsia, a special collection inDemographicResearch.Demographic Research,32,1099-1112.RetrievedJune03,2015fromhttp://www.demographic-research.org/special/15/
International Journal of Software InnovationVolume 6 • Issue 4 • October-December 2018
29
Long Niu received a B.E. Degree in Information engineering from Zhengzhou University, China, in 2012, a M.E. Degree in Computational Science from Kobe University, Japan, in 2016. Now he has a Ph.D. candidate in the Graduate School of System Informatics at Kobe University. His research interests include Activity Recognition by Using Non-intrusive Environment Sensing Data and Integration Framework for Indoor Location.
Sachio Saiki is an assistant professor in the Graduate School of System Informatics at Kobe University. His research interests include software engineering education and signal processing.
Masahide Nakamura received B.E., M.E., and Ph.D. degrees in Information and Computer Sciences from Osaka University, Japan, in 1994, 1996, 1999, respectively. From 1999 to 2000, he has been a post-doctoral fellow in SITE at University of Ottawa, Canada. He joined Cybermedia Center at Osaka University from 2000 to 2002. From 2002 to 2007, he worked for the Graduate School of Information Science at Nara Institute of Science and Technology, Japan. He is currently an associate professor in the Graduate School of System Informatics at Kobe University. In 2015, he worked for Universite Grenoble Alpes as a visiting professor. In 2018, he joined RIKEN Center for Advanced Intelligence Project (AIP)
as a visiting researcher. His research interests include the service/cloud computing, smart home, smart city, and gerontechnology. He is a member of the IEEE, IEICE and IPSJ.