The Deep Learning-based Human Activity Recognition Using Smart Wearable Sensors : A Tutorial

Abstract

Human activity recognition (HAR) mechanisms that distinguish human behavior utilizing wearable
sensors have advanced significantly over several years. Not only have state-of-the-art techniques ignored
hand-crafted features in favor of end-to-end deep learning approaches, but best practices for
designing experiments, preparing datasets, and assessing activity recognition systems have changed
in lockstep. This tutorial will provide an in-depth, hands-on introduction to the topic of sensor-based
HAR for those who are new to it. We will concentrate on deep learning-based HAR in this tutorial
utilizing data from intelligent wearable sensor devices. This tutorial introduces the SDL-HAR
framework, which provides a general-purpose framework for data preprocessing, data generation,
model development, and evaluation. We describe each aspect of the provided system in-depth, offer
references to relevant research, and explain the community’s best practice methodologies for activity
identification. Two exemplary deep learning approaches, convolutional neural network (CNN) and
long short-term memory neural network (LSTM), are deployed in this lesson using state-of-the-art
public HAR datasets. Additionally, this tutorial highlights the problems and future research directions
of sensor-based HAR.