Seite - 369 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

healthintegration[24].Theysplit theintegrationindifferent layers,dataacquisition,data processing,dataanalytics,andapplication.Nevertheless, theproposal is toogeneralwith respect tohowhomogeneizedifferent sourcesof data, since itsmajor concern is related tohowhandlingmissingvalueswhile integratingdatabases. Since acquiring labeled time series is a costlyprocedure in termsof resource, time, andpeopleinvolved,wethinkthattheintegrationofexistingdatasetsis therightdirection despite the strongheterogeneityof thedata. In this article we propose a platform that semi-automatically integrates heteroge- neousdataandprovides theminahomogenous form.Themaincontributionof this arti- cle is thedefinitionofanewplatformthat: • harmonizesheterogeneousdatafrominertial sensors,beingthemalreadyexisting datasetsor coming fromonlineacquisitions; • distributes sets of labeled signals according to specific requests, suchas, data re- lated to a specific activity, data from subjectswith specific physical characteris- tics,or,moregenerally,data fromsubjects thatperformedasetofspecifiedactiv- ities; • distributesactivity recognitionmodels; • providesanonline servicebyassociating labels to seriesof signals in real-time. Thearticle isorganizedas follows:Section2providesanoverviewofourplatform; Section3specifieshowheterogeneousdatasets canbeacquired inorder tobe integrated in a unified dataset; Section 4 describes how the heterogeneous datasets can be unified so that theyareorganized inauniformformat;Section5provides insightsabouthowthe UnifiedDataset canbeexploited;finally,Section6sketches futuredirections. 2. ContinuousLearningPlatform Human Activity Recognition (HAR) is a very active research field. Many techniques havebeenproposed,most of thembasedon theanalysis of inertial signals fromsensors embeddedinsmartphones, smartwatches,fitness trackers,andmanyothersad-hocwear- abledevices. See thepaper byCornacchia et al. for a surveyonactivitydetectionusing wearabledevices [10]. Themain aimof theContinuousLearningPlatform (CLP in the sequel) is tomake available (i) a large amount of labeled inertial signals related ofADLs and falls; (ii) a catalogue of downloadable activity recognitionmodels, and (iii) a service that, given a setof rawdata, identifies thecorrespondingADL. Labeled inertial signalscanbeusedbyresearches tobothvalidatenewADLsrecog- nitiontechniques.Activityrecognitionmodelscanbeintegratedintoexistingdomainde- pendentapplications that requireADLsrecognition inorder toprovideapplication func- tionalities (e.g., estimation of energy expenditure [30,15],monitoring the development of theParkinson’s disease [29], and early detectionof dementia [33]). Finally, if an ap- plication that requires to identify theADLs performed by the user does not include an activity recognitionmodel, it can rely on the service the platformprovides each time a newseriesofdata is acquired. To fulfil the aim, CLP collects inertial signals from existing datasets or applica- tions,manages the collected inertial signals, and distributes uniformed labeled inertial A.Ferrari etal. /AFramework forLong-TermDataCollection 369