ABSTRACT
Many different approaches exist for human activity recognition (HAR) in a private home setting. A combination of ambient sensors and Hidden Markov Models (HMM) enables HAR without undue invasion into a persons privacy. HMM enable capturing the structure of the activities and relating them to sensor readings. Virtual stochastic sensors (VSS) stem from HMM and include arbitrary stochastic distributions for duration modeling, resulting in more realistic activity and behavior representation. In this paper we show how VSS can be successfully applied to activity recognition, using the CASAS Aruba 2010 data-set, and producing results with accuracy of 85%, which is competitive with currently used approaches.
