| Summary: | The research presented in this thesis investigates how deep learning and feature learning can address challenges that arise for activity recognition systems in naturalistic, ecologically valid surroundings such as the private home. One of the main aims of ubiquitous computing is the development of automated recognition systems for human activities and behaviour that are sufficiently robust to be deployed in realistic, in-the-wild environments. In most cases, the targeted application scenarios are people’s daily lives, where systems have to abide by practical usability and privacy constraints. We discuss how these constraints impact data collection and analysis and demonstrate how common approaches to the analysis of movement data effectively limit the practical use of activity recognition systems in every-day surroundings. In light of these issues we develop a novel approach to the representation and modelling of movement data based on a data-driven methodology that has applications in activity recognition, behaviour imaging, and skill assessment in ubiquitous computing. A number of case studies illustrate the suitability of the proposed methods and outline how study design can be adapted to maximise the benefit of these techniques, which show promising performance for clinical applications in particular.
|
|---|
