|
Abstract:
|
The development of body-worn devices, such as smartphones and smartwatches, has remarkably deepened our understanding on how physical activity (PA) impacts human health. However, the hitherto findings are possibly just a foretaste of what data from wearable devices may reveal as many questions on their processing and analysis remain unanswered. One such question regards activity recognition. In this talk, we introduce a novel method for recognition of walking. In our approach, we utilize temporal dynamics of body motions measured by wearables’ accelerometers. Our focus is given to salient walking features, such as intensity, periodicity, duration, and speed. We investigate their reflection in acceleration signals collected at various body locations typical to wearables, and we create a classification scheme that allows for interpretable estimation of walking bouts. We show that our method demonstrates very high classification accuracy on over a dozen of publicly available PA datasets. Thanks to the transparent methodology, wide validation, and open source code, we present that our method carries high potential for application in various health studies.
|
