Benefit from unobtrusive contexts

Context-aware systems can be used in a multitude of scenarios and applications. Examples include a pedestrian safety system which can help to detect a potential upcoming accident based on the knowledge of the pedestrian’s activity. In the field of human activity recognition, a majority of research focuses on long-lasting and continuous activities, such as walking or attending a dinner. Yet, the recognition of short and non-periodic activities, such as stepping onto the road, is still poorly understood. Due to the non-periodic nature, missing or incorrect recognitions of such activities imply that the entire activity is not recognized and remains unknown. This thesis focuses on short and non-periodic activities, gives a definition and investigates three challenges in their recognition: inconsistent sensor data, overrepresentation of periodic activities and evaluation of the recognition. For each challenge a possible solution is proposed. It is shown that applying these solutions addresses the challenges and improves the recognition rates of short and non-periodic activities up to 20%.