基于自然梯度的噪声自适应变分贝叶斯滤波算法

考虑到运动目标跟踪系统机动、隐身等人为对抗特征以及非视距、干扰、遮挡等环境因素, 其系统建模、估计与辨识过程中越来越无法回避非线性、非高斯以及参数未知等复杂系统特征的影响. 针对过程噪声先验信息不准确以及量测噪声非高斯环境下运动目标的非线性状态估计问题, 提出一种基于自然梯度的噪声自适应变分贝叶斯(Variational Bayes, VB)滤波算法. 首先, 利用指数族分布具有统一表达形式的优势, 构建参数化逆威沙特(Inverse-Wishart, IW)分布作为状态一步预测误差协方差的共轭先验分布, 同时选取学生t分布重构因量测随机缺失导致的具有非高斯特点的似然函数; 其次, 在变分贝叶斯优化框架下采用平均场理论将状态变量联合后验分布近似分解为独立的变分分布, 在此基础上, 结合坐标上升方法更新各变量的变分分布参数; 进而, 结合 Fisher 信息矩阵推导置信下界最大化关于状态估计及其估计误差协方差的自然梯度, 使非线性状态后验分布的近似分布沿梯度下降, 以实现对状态后验概率密度函数(Probability density function, PDF)的“紧密”逼近. 理论分析和仿真实验表明: 相对传统的非线性滤波方法, 本文算法对噪声不确定问题具有较好的自适应能力, 并且能够获得较高的状态估计精度.

A Novel Noise Adaptive Variational Bayesian Filter Using Natural Gradient

Considering the increasing complexity and changeability of characteristics such as maneuvering and stealth in moving target tracking system and the influence of adverse factors such as non-line-of-sight, interference and occlusion in measurement environment. State estimation is likely to be confronted with complex system characteristics such as nonlinearity, non-Gaussian noise and unknown parameters. Aiming at nonlinear adaptive state estimation of moving target in a system with unknown process noise and non-Gaussian measurement noise, a novel noise adaptive variational Bayesian (VB) filter using natural gradient is proposed. Firstly, a parameterized inverse-Wishart (IW) distribution and a student＇s t distribution are constructed as the conjugate prior distribution of predicted state error covariance and measurement likelihood respectively. Then, in the framework of variational Bayesian optimization, the joint a posteriori distribution of estimation variables is approximately decomposed into independent variational distributions by using mean-field theory. On this basis, the variational distribution parameters of each variable are updated by combining coordinate ascend method and the characteristics of exponential distributions. Furthermore, under the condition of maximizing evidence lower bound, the natural gradients with respect to state estimation and its error covariance are derived by combining with Fisher information matrix. So that the variational distribution of nonlinear state gradually approaches the posteriori probability density function (PDF) of state along the natural gradient direction. Finally, simulation results show that the proposed algorithm has better adaptive ability to the problem of noise uncertainty and can obtain higher estimation accuracy compared to traditional algorithms.