Authors:
Gabriela Ciortuz
;
Hawzhin Hozhabr Pour
and
Sebastian Fudickar
Affiliation:
Institute of Medical Informatics, University of Lübeck, Ratzeburger Allee 160, Lübeck, Germany
Keyword(s):
Human Activity Recognition, Wearables, Deep Learning, Neural Networks, Time-Series, Behaviour, Functional Assessment, Physical Condition.
Abstract:
This article provides a comprehensive look at human activity recognition via three consumer devices with different body placements and a deep hybrid model containing CNN and LSTM layers. The used dataset consists of 53 activities recorded from the motion sensors (IMUs) of the three devices. Compared to the available human activity recognition datasets, this dataset holds the biggest number of classes, enabling us to provide an in-depth analysis of activity recognition for health-related assessments, as well as a comparison with other benchmark models such as a CNN and LSTM model. In addition, we categorize the activities into six movement groups and discuss their relevance for health-related assessments. Our results show that the hybrid model outperforms the benchmark models for all devices individually and all together. Furthermore, we show that the smartwatch could as a standalone consumer device classify activities in the six movement groups very well and for most of the use cases
using a smartwatch would be practical.
(More)