activity-recognition-from-user-annotated-accelerat

ActivityRecognitionfromUser-AnnotatedAccelerationDataLingBaoandStephenS.IntilleMassachusettsInstituteofTechnology1CambridgeCenter,4FLCambridge,[e-mail],algorithmsaredevelopedandevaluatedtode-tectphysicalactivitiesfromdataacquiredusingfivesmallbiaxialac-celerometerswornsimultaneouslyondifferentpartsofthebody.Ac-celerationdatawascollectedfrom20subjectswithoutresearchersu-pervisionorobservation.Subjectswereaskedtoperformasequenceofeverydaytasksbutnottoldspecificallywhereorhowtodothem.Mean,energy,frequency-domainentropy,andcorrelationofaccelerationdatawascalculatedandseveralclassifiersusingthesefeaturesweretested.De-cisiontreeclassifiersshowedthebestperformancerecognizingeverydayactivitieswithanoverallaccuracyrateof84%.Theresultsshowthatalthoughsomeactivitiesarerecognizedwellwithsubject-independenttrainingdata,othersappeartorequiresubject-specifictrainingdata.Theresultssuggestthatmultipleaccelerometersaidinrecognitionbecauseconjunctionsinaccelerationfeaturevaluescaneffectivelydiscriminatemanyactivities.Withjusttwobiaxialaccelerometers–thighandwrist–therecognitionperformancedroppedonlyslightly.Thisisthefirstworktoinvestigateperformanceofrecognitionalgorithmswithmultiple,wire-freeaccelerometerson20activitiesusingdatasetsannotatedbythesubjectsthemselves.1IntroductionOneofthekeydifficultiesincreatingusefulandrobustubiquitous,context-awarecomputerapplicationsisdevelopingthealgorithmsthatcandetectcontextfromnoisyandoftenambiguoussensordata.Onefacetoftheuser’scontextishisphys-icalactivity.Althoughpriorworkdiscussesphysicalactivityrecognitionusingacceleration(e.g.[17,5,23])orafusionofaccelerationandotherdatamodalities(e.g.[18]),itisunclearhowmostpriorsystemswillperformunderreal-worldconditions.Mostoftheseworkscomputerecognitionresultswithdatacollectedfromsubjectsunderartificiallyconstrainedlaboratorysettings.Somealsoevalu-aterecognitionperformanceondatacollectedinnatural,out-of-labsettingsbutonlyuselimiteddatasetscollectedfromoneindividual(e.g.[22]).Anumberofworksusenaturalisticdatabutdonotquantifyrecognitionaccuracy.Lastly,researchusingnaturalisticdatacollectedfrommultiplesubjectshasfocusedonA.FerschaandF.Mattern(Eds.):PERVASIVE2004,LNCS3001,pp.1–17,2004.c
Springer-VerlagBerlinHeidelberg20042L.BaoandS.S.Intillerecognitionofalimitedsubsetofnineorfewereverydayactivitiesconsistinglargelyofambulatorymotionsandbasicposturessuchassittingandstand-ing(e.g.[10,5]).Itisuncertainhowpriorsystemswillperforminrecognizingavarietyofeverydayactivitiesforadiversesamplepopulationunderreal-worldconditions.Inthiswork,theperformanceofactivityrecognitionalgorithmsundercondi-tionsakintothosefoundinreal-worldsettingsisassessed.Activityrecognitionresultsarebasedonaccelerationdatacollectedfromfivebiaxialaccelerometersplacedon20subjectsunderlaboratoryandsemi-naturalisticconditions.Super-visedlearningclassifiersaretrainedonlabeleddatathatisacquiredwithoutresearchersupervisionfromsubjectsthemselves.Algorithmstrainedusingonlyuser-labeleddatamightdramaticallyincreasetheamountoftrainingdatathatcanbecollectedandpermituserstotrainalgorithmstorecognizetheirownindividualbehaviors.2BackgroundResearchershavealreadyprototypedwearablecomputersystemsthatuseac-celeration,audio,video,andothersensorstorecognizeuseractivity(e.g.[7]).Advancesinminiaturizationwillpermitaccelerometerstobeembeddedwithinwristbands,bracelets,adhesivepatches,andbeltsandtowirelesslysenddatatoamobilecomputingdevicethatcanusethesignalstorecognizeuseractivities.Fortheseapplications,itisimportanttotrainandtestactivityrecognitionsystemsondatacollectedundernaturalisticcircumstances,becauselaboratoryenvironmentsmayartificiallyconstrict,simplify,orinfluencesubjectactivitypatterns.Forinstance,laboratoryaccelerationdataofwalkingdisplaysdistinctphasesofaconsistentgaitcyclewhichcanaiderecognitionofpaceandincline[2].However,accelerationdatafromthesamesubjectoutsideofthelaboratorymaydisplaymarkedfluctuationintherelationofgaitphasesandtotalgaitlengthduetodecreasedself-awarenessandfluctuationsintraffic.Consequently,ahighlyaccurateactivityrecognitionalgorithmtrainedondatawheresubjectsaretoldexactlywhereorhowtowalk(orwherethesubjectsaretheresearchersthemselves)mayrelytooheavilyondistinctphasesandperiodicityofaccelerom-etersignalsfoundonlyinthelab.Theaccuracyofsuchasystemmaysufferwhentestedonnaturalisticdata,wherethereisgreatervariationingaitpattern.Manypastworkshavedemonstrated85%to95%recognitionratesforambu-lation,posture,andotheractivitiesusingaccelerationdata.Somearesumma-rizedinFigure1(see[3]forasummaryofotherwork).Activityrecognitionhasbeenperformedonaccelerationdatacollectedfromthehip(e.g.[17,19])andfrommultiplelocationsonthebody(e.g.[5,14]).Relatedworkusingactivitycountsandcomputervisionalsosupportsthepotentialforactivityrecognitionusingacceleration.Theenergyofasubject’saccelerationcandiscriminateseden-taryactivitiessuchassittingorsleepingfrommoderateintensityactivitiessuchaswalkingortypingandvigorousactivitiessuchasrunning[25].RecentworkActivityRecognitionfromUser-AnnotatedAccelera