A Study of Machine Learning Techniques based on Human Daily Living Activities via Inertial Sensors

Abstract

The recognition of human daily living activities has attained significant attention in recent times. As a result, many methods have been used in the literature to detect and monitor human lifelogs. Despite the plethora of applications, the classification and detection of human behaviours, which may result in inappropriate reactions and responses, still need to be more accurate. For instance, conventional machine learning techniques employ hand-crafted features based on a classification strategy. However, deep learning methods have shown improved recognition rates with better performance. Thus, this study was aimed at presenting a detailed overview of recent and state-of-the-art supervised and unsupervised human daily lifelog classification techniques. In addition, a comprehensive analysis will be presented of how various design parameters, such as the volume of features and other data fusion techniques from different sensor locations, can affect the overall recognition performance. Furthermore, with the rapid advancement of body-worn sensing technology and modelling approaches, the widespread usage of wearable sensors is anticipated to provide countless opportunities for precise and reliable inferences across a broad range of human activities.