Layered hidden Markov models for real-time daily activity monitoring using body sensor networks

This paper presents an inferring and training architecture for long-term and continuous daily activity monitoring using a wearable body sensor network. Energy efficiency and system adaptivity to wearers are two of the most important requirements of a body sensor network. This paper discusses a two-layered hidden Markov model (HMM) architecture for in-network data processing to achieve energy efficiency and model individualization. The bottom-layer HMM is used to process sensory data locally at each wireless sensor node to significantly reduce data transmissions. The top-layer HMM is utilized to find the activity sequence from the result of the local processing. This approach is energy efficient in that only the results of the decoding procedure in each node need to be transmitted rather than raw sensing data. Therefore, the volume of data are significantly reduced. When the algorithm is applied in online monitoring systems, the results of local processing are transmitted only upon hidden state changes. The top-layer processing uses “old data” of one sensor node when it does not receive a “new” result sequence of the local processing from that sensor node. The adaption to various wearers is also discussed, and the robustness of this classification system is depicted. Experiments of 19 activity sequences to be classified are taken by 5 subjects to evaluate the performance of this system.