Partially hidden Markov models for privacy-preserving modeling of indoor trajectories

Markov models are natural tools for modeling trajectories, following the principle that recent location history is predictive of near-future directions. In this work, we study Markov models for describing and predicting human movement in indoor spaces, with the goal of modeling the movement on a coarse scale to protect the privacy of the individuals. Modern positioning devices, however, provide location information on a much more finer scale. To utilize this additional information we develop a novel family of partially hidden Markov models that couple each observed state with an auxiliary side information vector characterizing the movement within the coarse grid cell. We implement the model as a nonparametric Bayesian model and demonstrate it on real-world trajectory data collected in a hypermarket.