Geometry-Consistent Diffusion Particle Filter for Robust Multistatic Radar Tracking

Abstract

Target tracking in multistatic radar systems is fundamental to wide-area surveillance, yet it faces two major challenges. First, the measurement geometry is highly nonlinear and varies with target range, which may limit the performance of conventional trackers when operating beyond their design regime. Second, individual receivers are subject to time-varying, non-Gaussian outliers, which can severely degrade fusion performance. This letter proposes a Geometry-Consistent Diffusion Particle Filter (GC-DPF) integrating a learned diffusion proposal, a gated geometric correction, and a per-station robust likelihood. Comparative experiments with Kalman-family filters, variational Bayesian methods, conventional particle filters, and a learned-proposal filter show that the same GC-DPF model maintains accurate tracking in the evaluated near- and far-range scenarios and under heavy outlier contamination.