Seite - 370 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

signals, trained activity recognitionmodels, and labels assigned to series of inertial sig- nals.ThosefunctionalitiesarerespectivelyreifiedbytheDataCollection,DataManage- ment, andDataDistribution components. Figure1 sketches theoverall architecture and its interactionwith theexternal actors. Data Collection Data Management Data Distribution .CSV .MAT X,Y,Z UniMiB Acquisition Push Pull DATASET X,Y,Z SAMPLE RUN LABEL UniMiB Tracker COMPONENT UniMiB Acquisition Push Pull Figure1. Overviewof theplatform. Thefollowingsectionsprovidedetailedinformationabouthowthecomponentshave beenarchitecturallydefinedandaboutourpreliminarydevelopment results. 3. DataCollection TheDataCollection is thecomponent thatacquiresexistinginertialsignalsanduniforms their organization to populate the repository of labeled inertial signals called Unified Dataset. This is achievedby integrating i) existingdatasets (e.g.,UniMiBSHAR[22]); ii) labeled inertial signals coming fromad-hoc applications that have been designed to acquire new labeled inertial signals fromvolunteers (e.g., UniMiBAAL [13]); and iii) signals fromvolunteersperformingADLsthat theplatformenrichesbyassigningthema proper labeland then integrates in therepository(e.g.,SensorDataLogger [1]).Figure2 sketches themoduleswe identified for theDataCollectioncomponent. Oneof themain issues in handlingmultiple datasets and exploiting them to train a single classifier, is the lack of consistency in terms of how the data is stored in the file systemandwhatare the informationprovided.Forexample, in theUCIHARdataset [2] dataarestoredintwoseparateddirectories(trainandtest)whichcontain .txtdatafiles. In contrast, inMobiAct[35]dataaresubdividedin20directories,eachofthemrepresenting an activity or a scenario. Each directory includes .csvfiles each belonging to a specific subject and to a specific trail. Thus, the different structures used influence howdata is storedand thegeneralorganizationof the information. InContinuousLearningPlatformweenforceasinglestoragetechnologyforalldata anda single structure for thedata. Inorder to standardize the formatof the information, CLP offers awell-defined format of the data: from the signals to the supporting infor- mation that enriches the data. Thus, in order to be integrated, a new datasetmust also be coupledwith a software component (called driver) that is able to interpret the data of the sourcedataset andconvert it into the format defined inCLP. Thedevelopment of thedriver is responsibilityof the researcherwhowhishes to loadhis/herdataset into the A.Ferrari etal. /AFramework forLong-TermDataCollection370