Daily Activity Recognition using Wearable Sensors via Machine Learning and Feature Selection

Abstract

Human activity recognition has been the focus of significant research due to its various applications. Bio-signals acquired by wearable inertial sensors is one type of data that can be used to accomplish that task. Also, machine learning techniques have become a standard pattern discovery tool in such a problem. This has stimulated the construction of many publicly available datasets to learn from, with variations in the number of sensors and activities, among others. The Human Gait Database (HuGaDB) is a state- of-the-art (SOTA) example of such datasets, and is considered the most comprehensive to date.In this paper, we incorporate four feature selection techniques along with four different classifiers to attain the highest recognition accuracy. Extensive analysis is first applied to determine the optimal number of features, which is then fed to four different techniques of sequential feature selection. We demonstrate that higher recognition accuracies are achievable with significant reduction in the number of features. We also show that sequential forward floating feature selection with the random forest classifier yields the highest recognition accuracies.