UiS4ADL Edge HAR: Design of Human Activity Recognition Model For Edge Devices Using STM32 and Thingy 53

Abstract

Human Activity Recognition (HAR) is crucial in various fields such as healthcare, fitness monitoring, and home automation. There are many Activities of Daily Living (ADLs) recognition approaches that focuses on broad-scale activities such as walking, sitting and standing. There are however few that focuses on fine-scale activities such as writing, washing hands and brushing hair. This thesis therefore focuses on recognizing a set of 24 ADLs were many of the activities are hand-based ADLs. The dataset used were collected from 31 subjects using a wrist-worn inertial sensor. Analysis of the dataset revealed a challenge in the variations of activity duration. Based on this, there were developed two classical HAR pipelines where different Deep Learning (DL) models were evaluated using multiple evaluation metrics. The capabilities of employing a HAR model on an edge device were investigated.