Newcombe, L (2020) Investigation of Low-Cost Wearable Internet of Things Enabled Technology for Physical Activity Recognition in the Elderly. Doctoral thesis, Liverpool John Moores University.

Preview

Text 2020NewcombePhD.pdf - Published Version Download (28MB) | Preview

Abstract

Technological advances in mobile sensing technologies has produced new opportunities for the monitoring of the elderly in uncontrolled environments by researchers. Sensors have become smaller, cheaper and can be worn on the body, potentially creating a network of sensors. Smart phones are also more common in the average household and can also provide some behavioural analysis due to the built-in sensors. As a result of this, researchers are able to monitor behaviours in a more naturalistic setting, which can lead to more contextually meaningful data. For those suffering with a mental illness, non-invasive and continuous monitoring can be achieved. Applying sensors to real world environments can aid in improving the quality of life of an elderly person with a mental illness and monitor their condition through behavioural analysis. In order to achieve this, selected classifiers must be able to accurately detect when an activity has taken place. In this thesis we aim to provide a framework for the investigation of activity recognition in the elderly using low-cost wearable sensors, which has resulted in the following contributions: 1. Classification of eighteen activities which were broken down into three disparate categories typical in a home setting: dynamic, sedentary and transitional. These were detected using two Shimmer3 IMU devices that we have located on the participants’ wrist and waist to create a low-cost, contextually deployable solution for elderly care monitoring. 2. Through the categorisation of performed Extracted time-domain and frequency-domain features from the Shimmer devices accelerometer and gyroscope were used as inputs, we achieved a high accuracy classification from a Convolutional Neural Network (CNN) model applied to the data set gained from participants recruited to the study through Join Dementia Research. The model was evaluated by variable adjustments to the model, tracking changes in its performance. Performance statistics were generated by the model for comparison and evaluation. Our results indicate that a low epoch of 200 using the ReLu activation function can display a high accuracy of 86% on the wrist data set and 85% on the waist data set, using only two low-cost wearable devices.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Activity Recognition; Internet of Things; Machine Learning
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Divisions:	Computer Science & Mathematics
Date Deposited:	08 Sep 2020 11:50
Last Modified:	19 Dec 2022 14:58
DOI or ID number:	10.24377/LJMU.t.00013558
Supervisors:	Carter, C, Yang, P and Hanneghan, M
URI:	https://researchonline.ljmu.ac.uk/id/eprint/13558

View Item