Interacting Multiple Model Particle Filter for Indoor Positioning Applications

Abstract

This work utilizes the interacting multiple model particle filter (IMMPF) for positioning purposes within buildings. It allows a combination and mixing of multiple, independently running particle filters. We also add a novel estimation strategy and we use the effective sample size to ensure a certain variance in particle weights. Approaches using standard particle filters for positioning often incorporate many different sensor models by assuming a statistical independence between them. As accuracy, update rate and error-proneness differ per sensor, it is often necessary to coordinate the individual models with each other, which is mainly done via parameterization. Finding the correct set of parameters is time-consuming and often changes depending on the scenario, building or even per path. The IMMPF mixes individual particle filters using a Markov Chain process. In this work we use three sensor models (BLE, Wi-Fi FTM and Movement) in two very different buildings. We then compare a standard particle filter with the IMMPF variant, which uses a separate filter per sensor model. These filters are mixed using a Markov transition matrix, depending on a quality metric of the respective model. The evaluation shows, that the IMMPF allows for higher positioning accuracy, as we are able to lower the impact of a certain sensor model in situations, where the quality metric indicates uncertainties and vice versa.