Human Activity Recognition Using Wearables

This thesis covers the development and practical implementation of a Human Activity Recognition (HAR) system, created for a Finnish wearable company Oura Health Oy. Oura produces smart rings with sensors collecting real-time biosignals, such as skin temperature, heart rate, and accelerometer values. Physical activity is one of the principal factors that influence health. Knowing what types of activities users perform could allow us to better understand our users' well-being and provide more helpful insights. We decided to build the system to encourage our users to label their daily physical activities by creating a smooth user experience with the HAR system. During our research, we developed a solution capable of producing good results, even having low-quality data as input. We sequentially developed our pipeline, from Activity Detection to Activity Classification, and Results Presentation. For the Activity Detection part, we adopted a time-series segmentation approach which we implemented with the PELT algorithm. The benefit of using segmentation is the ability to separate activities performed on a sequential basis. We also developed an approach for time-series feature extraction, which allowed us to map a varied-length time-series into a fixed dimensionality feature sets. For the Activity Classification, we implemented the XGBoost model as the most appropriate, given the limitations and practical considerations we had. We launched the final product, and it is currently available to 5% of Oura's users. The implemented HAR system improves user engagement and enables an additional stream of data collection for Oura. The accuracy of suggestions is high enough for users to enjoy the feature without constantly modifying suggested activities. When this product becomes available to the general Oura audience, we will be able to further increase the depth of analysis that we do in the company