Scalable indoor pedestrian localisation using inertial sensing and parallel particle filters

Location-aware computing is a fast emerging area in mobile computing. A plethora of approaches to indoor localisation have been demonstrated, but almost all rely on extensive infrastructure. A popular alternative is to use dead reckoning to track inertial sensors. However, sensor drift must be addressed by incorporating external constraints such as the building layout. This dictates the use of computationally expensive particle filters that hinder scalability, especially during localisation phases where the system does not have any estimate of where the user is within a building. In this paper, we address the scalability problem by exploiting the latent parallelism in the algorithm and adapting it for execution on commodity Graphical Processing Units (GPUs). We describe how to parallelise the particle filter and evaluate different filter architectures. We find that our GPU implementation can iterate 8.8 times faster than the fastest CPU variant. We also show how to handle multiple filters using a novel memory paging scheme and an adaptable particle number. We find that between 17 and 101 users can be localised in real-time using only a mid-range GPU installed in a standard desktop machine, compared with at most one using a previous sequential approach.