State Estimation for AMVs: A GPTVM Distribution With Adaptive Parameter Selection-Based VBUKF Approach

Accurate state estimation for autonomous marine vehicles is critical to mitigate navigation errors caused by non-stationary heavy-tailed measurement (HTM) noise from sensors. This paper develops a novel variational Bayesian unscented Kalman filter (VBUKF) approach based on Gaussian Pearson type VII mixture (GPTVM) distribution. Specifically, a Beta-Bernoulli distributed judgment factor is introduced to distinguish normal measurements and outliers dynamically. Two different Gamma distributed degree-of-freedom parameters are embedded in the GPTVM to improve the adaptive regulating ability over different degrees of HTM noise. An inverse Wishart distribution is selected as the prior of the measurement noise covariance to accommodate time-varying noise characteristics. Simulation results on a 76.2m autonomous surface vehicle demonstrate that, in comparison with the unscented Kalman filter and its variants, the proposed method achieves a reduction in the average root mean square error of position by over 60% and 25%. Experimental results further validate the effectiveness of the algorithm.