Skip to Main Content
Activity recognition approaches have shown to enable good performance for a wide variety of applications. Most approaches rely on machine learning techniques requiring significant amounts of training data for each application. Consequently they have to be retrained for each new application limiting the real-world applicability of today's activity recognition methods. This paper explores the possibility to transfer learned knowledge from one application to others thereby significantly reducing the required training data for new applications. To achieve this transferability the paper proposes a new layered activity recognition approach that lends itself to transfer knowledge across applications. Besides allowing to transfer knowledge across applications this layered approach also shows improved recognition performance both of composite activities as well as of activity events.