Automated Activity Recognition.

One of the goals of the BESiDE project is to identify enabling and disabling elements of the built environment. In order to understand, for example, how the built environment impacts on residents' physical activity we need methods and tools to analyze resident movement and behavior in the context of the building they live in.

Traditionally, behavior analysis is conducted through observations. These require researchers to spend a significant amount of time carefully observing and transcribing people’s activities. Being observed often makes people feel uncomfortable and may in some situations be perceived as an invasion of privacy. As part of the BESiDE project, we have developed tools to capture and recognise residents' behavior using motion sensors (accelerometers), and to map their activities onto building floorplans automatically. This approach reduces the required amount of researcher’s time and resolves the issues of being observed by a person to some extent. These tools could then be used to identify promising changes to the built environment and to assess the effectiveness of building interventions.

While humans are usually very good at cognitive tasks such as recognizing other people’s actions through visual observation, mimicking this skill in software and recognizing actions automatically is an open research problem. Recognizing human actions from sensors such as accelerometers is oftentimes even harder as they capture very limited information (for example, the motion of a person’s left hand). Sebastian Stein of the BESiDE team has been working on developing methods for activity recognition and physical activity assessment. He has a strong background in pattern recognition and machine learning, and he worked on automatic activity recognition from video and accelerometers for his PhD. Towards his degree, he investigated methods for recognizing food preparation activities, which could be used as part of a situational support system for people with cognitive impairments.