Seite - 373 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

Data Storage Data Aligner Data Access Pull Data Management H T T P Push Data Classifier Label Consolidator Dataset Composer H T T P Pull CLP Formatted Dataset Control and Data Flow CLP Formatted Dataset with Harmonized signals CLP Formatted Dataset with Harmonized signals and Labels Usage Module Figure3. TheDataManagement component. of the delicate nature of this procedure, thismodule is intended to be semi-automatic: it provides suggestions on the assignment of labels, but ultimately it is down to the end user todecidewhetherornot toaccept the suggestions. TheDataClassifiermodule is in charge of assigning labels to inertial signals that have not been labeled. These types of signalsmay result fromapplications that acquire signals only without providing any classification. This module can exploit an activity recognitionmodel already trained. In termsofdatadistribution, theDataComposermodule simply intercepts requests for labeledsignals, processes them,and returns the setof corresponding labeledsignals. Forexample, a request canbe: “all signals labeled running”. Froman implementationpoint of view, theDataManagement component is aweb serviceexposing thestoreDataset, and thegetDataset functions. 4.1. PreliminaryValidationof theDataManagementComponent Tostartvalidating theDataManagementcomponent,weimplemented theDataAligner, the Label Consolidator, and theDataComposermodules. All themodules have been developed inMatlab.Themodulesmanageaccelerationsignalsonly. Our implementation of theDataAlignermodule unifies themeasurement units to g, removes gravity form the signals, and resamples to 50Hz the frequencies (since this is the frequency usually used for ADLs recognition [26]). Removing the gravitational acceleration is not an exact process, however it is common practice and considered in literature [34]. Theimplementationof theLabelConsolidator includesanautomaticsyntacticanal- ysis of the labels based on the Levenshtein distance. Then the module relies on a k- Nearest Neighbor classifier in order to compute a confusionmatrix that helps the user in decidingwhich activities are similar and then can bemerged. This confusionmatrix and,whennecessary, somevisualplotsofactivities, areexploitedby theuser inorder to determine theconcreteassociationsof thevariousactivities and labels. Finally theDataComposermoduleallowstorequestspecificsetsof labeledsignals. The requests areparametrized.Forexample, the requestgetDataset(’activities’, [1 2 3], ’gender’, [’F’, ’age’, 24) returnsall thelabeledsignalsforactivities 1,2, and3performedbyfemales24yearsold. A.Ferrari etal. /AFramework forLong-TermDataCollection 373