Page - 368 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

performedbyhumans represents themainchallengeofHAR. Indeed, in an idealworld, different activities present different signals so that the space of the signals is perfectly separatedandeasy toclassify. In realworldsuchacondition isnot respectedandsignals oftenoverlapeachother. Theseparationofdata is adifficult taskdue todifferent aspectsof theclassification procedure, from the data acquisition to the definition of the activity. For instance, the position of the device influences the results of the performance [9]. Other aspects are related to the intraclass variability and the interclass similarity [8]. The formermeans thatdifferentpeopleperformthesameactivity indifferentwayssoabijectiveassociation between signal and activity performed does not exist; the lattermeans that classes that are fundamentallydifferent showverysimilar characteristics in the sensordata. PreliminaryapplicationsofHARexploitedsupervisedmachinelearningtechniques. These techniques present several challenges [18,8]. Firstly, the performances achieved in laboratoryaredifficult tobe transferred to a real contextwithout losingquality in the classification [3].Secondly, itwasnoticedastrongdependencybetween theselectionof the features and theperformanceof the classification,whichcorresponds to an inability of the algorithms to extract and organize discriminative information from the data [6]. Moreover, issuesabovementioned, suchasposition, intraclassvariability, and interclass similarity still strongly influence theclassificationperformance [21,17,16]. In recentyears,deep learninghasbeensuccessfullyapplied in imageanalysis,natu- ral languageprocessing, sound recognition, andmore recently it hasbeenexploitedalso for1Dsignals [19,28]. Thewidespread use of deep learning techniques is justified by their capability to overcomemost of the presented issues, thanks to their properties of local dependency and scale invariance [36]. Furthermore, deep learningmethodologies permit automated discoveryofabstractionwhichovercomes the featuresextraction issue [6]. While deep learning techniques are powerful and achieve high performance, they relyonverycomplexmodels that strictlydependon theestimationof a largenumberof parameters,which requiresaconsiderableamountofavailabledata [7]. In recent years, some researchers havepublished their owndataset related toHAR. However, these datasets are heterogeneous and their standardization in a single unified dataset requiresconsiderableeffort.Forexample, signalsareexpressed indifferentunits ofmeasure, theymay include gravity or not, and signals have different acquisition fre- quencies. Furthermore, labels arenot alignedwith a commondictionaryand sometimes havedifferentmeaningsamongdifferentdatasets (’sitting’mayrefer to thestateofbeing seated in a chair or the transition fromstanding to sitting). Thus, datasets are heteroge- neousandcannotbeused togetherwithout a significant effort toharmonize them. The availability of a dataset containing a large number of samples, also obtained thanks to the integration of existing datasets, is a well known issue both in the field of ADLs recognition from inertial sensors and in other domains, such as that related to image processing [25]. Indeed,merging labels fromdifferent databases that resulted equals at semanticor syntactic level,may likely result inan inconsistent setofdata, that doesnot improve the trainingofaclassifier [12]. In the context of ADLs recognition from inertial signals, Bartlett et al. propose labels aggregation at semantic level [5]. Recently, Siirtola et al. propose aMatlab tool calledOpenHAR [31] that aggregates labels at a syntactic level but does not consider the semantics of the original signals. Obinikpo et al. propose a system for big data-d- A.Ferrari etal. /AFramework forLong-TermDataCollection368