A dynamic Bayesian network approach for device-free radio vision: Modeling, learning and inference for body motion recognition

In this paper, a time-varying dynamic Bayesian network model is shown to describe human-induced RF fluctuations for the purpose of non-cooperative and device-free radiobased body motion recognition (radio vision). The technology relies on pre-existing wireless communication network infrastructures and processes channel quality information (CQI) for human-scale sensing. Body movements leave a characteristic footprint on the CQI sequences collected during consecutive radio transmissions over multiple co-located links. Body-induced RF footprints are proved to be effectively characterized by temporarily coupled hidden Markov chains: abrupt changes of body postures make CQIs observed over co-located links temporarily coupled while being uncoupled for slow body movements. Learning and classification/inference problems are discussed based on experimental measurements. Device-free radio vision performances are evaluated for arm gesture and fall detection applications.