A Possibilistic Approach for Activity Recognition

Chapter2APossibilisticApproachforActivityRecognitioninSmartHomesforCognitiveAssistancetoAlzheimer’sPatientsPatriceC.Roy1,SylvainGiroux1,BrunoBouchard2,AbdenourBouzouane2,CliftonPhua3,AndreiTolstikov4,andJitBiswas41DOMUSLab.,UniversitédeSherbrooke,J1K2R1,Canada2LIARALab.,UniversitéduQuébecàChicoutimi,G7H2B1,Canada3DataMiningDep.,I2R,138632,Singapore4NetworkingProtocolsDep.,I2R,138632,SingaporeCorrespondingauthor:[e-mail]tiveassistancetoAlzheimer’spatientsinsmarthomesisafieldofresearchthatreceivesalotofattentionlately.Therecognitionofthepatient’sbehaviorwhenhecar-riesoutsomeactivitiesinasmarthomeisprimordialinordertogiveadequateassistanceattheopportunemoment.Toaddressthischallengingissue,wepresentaformalactivityrecognitionframeworkbasedonpossibilitytheoryanddescriptionlogics.Wepresentini-tialresultsfromanimplementationofthisrecognitionapproachinasmarthomelaboratory.2.1IntroductionAmajordevelopmentinrecentyearsistheimportancegiventoresearchonambientintel-ligenceinthecontextofrecognitionofactivitiesofdailyliving.Ambientintelligence,inoppositiontotraditionalcomputingwherethedesktopcomputeristhearchetype,consistsofanewapproachbasedonthecapacitiesofmobilityandintegrationofdigitalsystemsinthephysicalenvironment,inaccordancewithubiquitouscomputing.Thismobilityandthisfusionaremadepossiblebytheminiaturizationandreducedpowerconsumptionofelectroniccomponents,theomnipresenceofwirelessnetworksandthefallofproductioncosts.Thisallowsustoglimpsetheopportunecompositionofdevicesandservicesofall3334ActivityRecognitioninPervasiveIntelligentEnvironmentskindsonaninfrastructurecharacterizedbyagranularityandvariablegeometry,endowedwithfacultiesofcapture,action,treatment,communicationandinteraction[1,2].Oneoftheseemerginginfrastructuresistheconceptofsmarthome.Tobeconsideredasintelligent,theproposedhomemustinevitablyincludetechniquesofactivityrecogni-tion,whichcanbeconsideredasbeingthekeytoexploitambientintelligence.Combiningambientassistedlivingwithtechniquesfromactivityrecognitiongreatlyincreasesitsac-ceptanceandmakesitmorecapableofprovidingabetterqualityoflifeinanon-intrusiveway.Elderlypeople,withorwithoutdisabilities,couldclearlybenefitfromthisnewtech-nology[3].Activityrecognition,oftenreferredasplanrecognition,aimstorecognizetheactionsandgoalsofoneormoreagentsfromobservationsontheenvironmentalconditions.Theplanrecognitionproblemhasbeenanactiveresearchtopicinartificialintelligence[4]foralongtimeandstillremainsverychallenging.Thekeyhole,adversarialorintendedplanrecogni-tionproblem[5]isusuallybasedonalogicorprobabilisticreasoningfortheconstructionofhypothesesaboutthepossibleplans,andonamatchingprocesslinkingtheobservationswithsomeactivitymodels(plans)relatedtotheapplicationdomain.Priorworkhasbeendonetousesensors,likeradiofrequencyidentification(RFID)tagsattachedtohouseholdobjects[6],torecognizetheexecutionstatusofparticulartypesofactivities,suchashandwashing[7],inordertoprovideassistivetaskslike,forinstance,remindersabouttheactiv-itiesofdailyliving(ADL).However,mostofthisresearchhasfocusedonprobabilisticmodelssuchasMarkovianmodelsandBayesiannetworks.Butthereissomelimitationswithprobabilitytheory.Firstly,thebeliefdegreeconcerninganeventisdeterminedbythebeliefdegreeinthecontraryevent(additivityaxiom).Secondly,theclassicalprobabilitytheory(singledistri-bution)cannotmodel(fullorpartial)ignoranceinanaturalway[8].Anuniformproba-bilitydistributiononaneventsetbetterexpressrandomnessthanignorance,i.e.theequalchanceofoccurrenceofevents.Ignorancerepresentsthefactthat,foranagent,eachpos-sibleevent’soccurrenceisequallyplausible,sincethereisnoevidencethatisavailabletosupportanyofthembythelackofinformation.Hence,oneofthesolutionstothiskindofproblemispossibilitytheory[9],anuncertaintytheorydevotedtothehandlingofincompleteinformation.Contraryasinprobabilitytheory,thebeliefdegreeofaneventisonlyweaklylinkedtothebeliefdegreeofthecontraryevent.Also,thepossibilisticencodingofknowledgecanbepurelyqualitative,whereastheprobabilisticencodingisnumericalandreliesontheaddi-PossibilisticActivityRecognition35tivityassumption.Thus,possibilisticreasoning,whichisbasedonmaxandminoperations,iscomputationallylessdifficultthanprobabilisticreasoning.Unlikeprobabilitytheory,theestimationofanagent’sbeliefabouttheoccurrenceofeventisbasedontwoset-functions.Thesefunctionsarethepossibilityandnecessity(orcertainty)measures.Finally,insteadofusinghistoricalsensorsdatalikeothersprobabilisticapproaches,weusepartialbeliefsfromhumanexpertsconcerningthesmarthomeoccupant’sbehaviors,accordingtoplausi-bleenvironment’scontexts.Sincetherecognitionsystemcanbedeployedindifferentsmarthomesandthattheirenvironmentsaredynamics,itismoreeasiertoobtainanapproxima-tionoftheoccupant’sbehaviorsfromhumanexperts’beliefsthanlearningthebehaviorsindifferentsmarthome’ssettings.Byusinggeneraldescriptionsofenvironmentalcontexts,itispossibletokeepthesamepossibilitydistributionsfordifferentsmarthome’ssettings.Consequently,anotheradvantageofpossibilitytheoryisthatitiseasiertoc