Adaptive Sequential Monte Carlo Filter for Indoor Positioning and Tracking With Bluetooth Low Energy Beacons

Abstract

We model the tracking of Bluetooth low-energy (BLE) transmitters as a three layer hidden Markov model with joint state and parameter estimation. We are after a filtering distribution by Bayesian approximation using Monte Carlo sampling techniques. In a test environment decorated with multiple BLE sensors, the tracking relies only on the naturally unreliable received signal strength indicator (RSSI) of the captured signals. We assume that the tracked BLE transmitter does not provide any other motion or position related information. Hence, the transition density is designed to be merely a diffusion where the probability measures are diffused into the neighboring space. This makes the diagonal error covariance factor of the prediction density, namely the diffusion factor, the most important parameter to be tuned on the fly. We first show an experimental proof of concept using synthetic data on real trajectories by comparing three parameter estimation approaches: static, decaying and adaptive diffusion factors. We then obtain the results on real data which show that online parameter sampling adapts to the observed data and yields lower error means and medians, but more importantly steady error distributions with respect to a large range of parameters.