Multimodal Human Activity Recognition From Wearable Inertial Sensors Using Machine Learning

This research addresses the problem of human activity recognition using wearable sensors. Due to its diverse applications, such as health monitoring, it has been attracting a considerable research attention. In order to benefit from machine learning, as an efficient tool for pattern discovery, in enhancing the recognition accuracy, the literature has witnessed the construction of several datasets with varying number of sensors and activities, among other variations. One of the state-of-the-art (SOTA) datasets and the most comprehensive to date is the publicly-available Human Gait Database (HuGaDB). We adopt diverse machine learning techniques for human activity recognition based on a comprehensive dataset HuGaDB, for which we show that particular classifiers, sensor positions and placements yield higher recognition accuracies. In particular, we study six classifiers with fourteen features for twelve different activities measured using two wearable senors. Random Forest (RF) is shown to consistently outperform other classifiers. We also present further investigations for combining multimodal data and for varying the parameters of the RF classifier in order to further improve the accuracy. A recognition accuracy of 98.6% was attained. This research is reproducible and attains SOTA results on a novel and comprehensive dataset.