A geodesic flow particle filter for non-thresholded measurements
Chris Kreucher, Kristine L. Bell · 2017
This paper derives a particle filter with flow induced by a log-homotopy for non-thresholded measurements (i.e., a track-before-detect log-homotopy particle filter). We have elected to use the Geodesic approach to particle flow and study both scaled-identity and Gaussian diffusion. We show the performance of the new filter provides order-of-magnitude tracking improvement over SIR filtering on a 2D tracking problem. While the homotopy approach requires significantly more computations per particle as it numerically solves a stochastic differential to flow each particle, we still show a significant improvement in performance per flop. We expect this performance improvement to widen, perhaps significantly, with higher dimensional state space. The numerics are discussed further in the paper.