Machine Learning Techniques for Context-Aware Human Activity Recognition: A Feasibility Study

Falls are a prominent issue in society and the second leading cause of unintentional death globally. Gait analysis is a typical and useful approach to identifying fall risk factors such as gait abnormalities, but this process has been shown to lack accuracy and reproducibility. This process can be made remote, however, the patient will be unobserved during this process. Human Activity Recognition (HAR) is an established field of research which tackles this issue, with many studies proving that walking activities such as level ground walking, navigating stairs, sitting, etc. can be accurately determined using the same wearable sensors that collect the required data for remote gait analysis. However, a person's gait is also dependent on the terrain underfoot and without this contextual information, gait health cannot be accurately assessed. The Context-Aware Human Activity Recognition (CAHAR) dataset is the first terrain and context-labeled dataset, which can be used to build classification models capable of labelling gait data with the full contextual information needed for remote gait analysis. In this study, we achieve an accuracy and precision of 94% using a single-model implementation of Support Vector Machines (SVMs), whilst an investigation into multiple models for classifying activity-terrain combinations outside the training set exhibits an accuracy of 50% and precision of 52%.