Activity recognition systems aim to understand what people (and animals) are doing by observing their movement and their environment. The emergence over the past decade of novel sensing, low-power wireless communication, fast processing and statistical algorithms has made recognition practical and useful in several fields. Notable successes include gaming, surveillance, elder care, personal fitness, sports physiology and ecological systems monitoring and protection [Philipose et al. 2004; Pollack 2005; Yang 2009]. In the past, traditional works in activity recognition shared several characteristics. They focused on activities performed one at a time in fixed instrumented areas, by individuals rather than groups, and they mostly relied on specialized sensors using kinematic, location and object-manipulation-based cues, and view data on interactive time scales. The underlying technical machinery was typically fully supervised propositional time-series analysis based on machine learning and data mining.