Page - 371 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

Data Collection Importer H T T P Dataset Loader H T T P Driver Loader H T T P Local Storage .PY .PY .PY Push .PY Push Pull .PY Push .PY .CSV .MAT Original Dataset CLP Formatted Dataset Control and Data Flow .PY Driver Module integrateDataset Figure2. TheDataCollectioncomponent. UnifiedDataset. Given the uniqueness of each datasets, it is unlikely that two datasets could share the samedriver, thuseachonewill likely requireanad-hoc implementation. Toease theburdenofwriting suchdrivers,weprovidea template interface fordevelop- ingnewdrivers,whichallowsusers to easilybuildnewdrivers that are compatiblewith CLP. TheDataCollectioncomponent includes thefollowingmodules: theDriverLoader, theDatasetLoader,andthe Importerwhichrespectivelyallowtoloadcustomdriversde- veloped to support specificdatasets, to loaddatasets tobe integrated inUnifiedDataset, and toaskfor the integrationof thenewdatasets into theUnifiedDataset.Separating the Dataset Loader from theDriverLoader, allows the dynamic on-boardingof the driver, which may require a reboot of theDataset Loader service in order to be visible and exploitable fromtheservice itself. Froman implementation point of view, all themodules areweb services exposing theloadDriver, theloadDataset, and theintegrateDataset functions. 3.1. PreliminaryValidationof theDataCollectionComponent In order to start validating theData Collection component, we developed the drivers for the following datasets: Motion Sense [20], MobiAct [35], Real Word HAR [32], UmaFall [9], andUniMiBSHAR[22].Weselected thesedatasets for the following rea- sons. First, wewere focused on datasets recorded by smartphone and smartwatch, be- cause thosekindofacquiringdevicesarenot invasivedevicesandarewidespreadamong the population. Second,we considered only datasets that have been acquired forHAR purposes. Indeed, suchdatasetsmay share the set of activities recorded.Third,we con- sidered only datasets that are enrichedwith additional information related to the sub- jects’ characteristics, suchas sex, age,height,weight.This allows theDataDistribution component toprovide sets of signals related to subjectswith specifiedcharacteristics or tomake available trained classifiers onlywith subsets of signals acquired fromsubjects with characteristics similar to the those required. The application of personalization, in fact, seems to provide better results in terms of accuracy [17,11]. Fourth, we selected data sets collected from2016until today forhavingcomparable technologyaccuracy. A.Ferrari etal. /AFramework forLong-TermDataCollection 371