Neural Augmented Particle Filtering with Learning Flock of Particles

Particle filters (PFs) are a popular family of algorithms for state estimation in dynamic systems. The usage of PFs is often limited due to complex or approximated modelling and the need for low latency tracking. In this work, we propose to enhance PFs by augmenting their operation with a dedicated deep neural network (DNN), that learns to correct the output set of particles, which we coin flock, based on the relationships between the particles in the set itself. Our proposed neural augmentation, which can be readily incorporated into different PFs, is designed to facilitate rapid operation by maintaining accuracy with a reduced number of particles. Our method leverages data to utilize a cross-particle correction while supporting different number of particles and preserving the filter operation. We experimentally show the improvements in performance, robustness, and latency on a benchmark radar target tracking PF Algorithm, and even the mitigation of the effect of a mismatched observation modelling.