Partially Observable Markov Decision Process Model for Dynamic Human Activity Recognition Using Radio Frequency Signals

Human activity recognition (HAR) using radio frequency signals and machine learning is a novel approach for sensing and detection of human activities in a privacy preserving scheme. This is particularly important in healthcare systems, since privacy of patients and staff is a priority. Another advantage of this technology is the non-contact activity recognition with no need of line-of-sight (LoS). However, limited attention has been paid to human activity dynamics and their impact on HAR. In this paper, we propose a framework for dynamic human activity recognition (DHAR) based on a partially observable Markov Decision Process (POMDP) model, motivated by potential inaccuracy of current HAR models in real-world environments and human activity dynamics. The POMDP model dynamically uses historical data of observed human activities and model selections to select an HAR model from a set of given models. An approximate method is proposed to solve the POMDP model. The simulation experiments show that the DHAR method can deliver better performance over each single HAR model.