Nonlinear Noise Adaptive Kalman Filtering By Stochastic Search Variational Inference

In this paper, we develop a recursive solution of joint nonlinear state estimation and model parameters identification based on approximation Bayesian inference principle. The stochastic search variational inference, combining natural gradients and stochastic gradients, is adopted to offer a flexible, accurate, and effective approximation of the required posterior distributions. We make two contributions compared with the existing variational inference based noise adaptive filtering. First, we introduce an auxiliary latent variable to decouple the latent variables of state and model parameters, which improves the posterior flexibility and the tightness of the variational lower bound. Second, we split the evidence lower bound optimization into intractable and tractable parts, where stochastic gradients and natural gradients are derived, respectively. The performance of the proposed method is verified on radar target tracking applications by simulated data.