Investigating human activity recognition in a real life setting

Abstract

Human activity recognition from wearable sensor data is an active area of research that promises viable applications such as in fitness, healthcare, lifestyle monitoring and also human-computer interaction. Modern smartphones are a popular medium for such applications, mainly due to their extensive set of readily available embedded sensors, and also due to their increased computational performance. However, such applications can be susceptible to low recognition accuracy, which in turn sabotages the purpose of the entire application and thus rendering it useless. The problem being tackled in this study is with respect to the recognition accuracy a human activity recognition system is able to achieve. This study provides an overview of how a human activity recognition system operates, from the data collection process to obtaining a final activity classification. The entire process is studied in detail with the aim of improving the accuracies that such systems are able to achieve. This will be done by decomposing the main processes of an activity recognition system, and reproducing all of them one by one, which involves the design and implementation of a wearable application to gather user motion data, the construction of an activity dataset through controlled experiments, thorough processing of activity data, and finally the use of different machine learning algorithms to obtain the results of different processing methodologies. Different activities will be studied, where each poses a different predicament. Fitness related activities will also be considered due to the rapid increase in interest in health and fitness related applications. By the end of this study, the reader should have a clear understanding of how different methodologies employed in a human activity recognition system directly affect the results of that system, and how different data needs to be processed in different manners.