自动化学报2025，Vol.51Issue(7)：1612-1625,14.DOI:10.16383/j.aas.c240497
基于广义最大相关熵准则的几何滤波方法
Geometric Filtering Method Based on Generalized Maximum Correntropy Criterion
摘要
Abstract
Geometric filtering is a method that uses observed data to optimally estimate the geometric state on a manifold,and it plays a significant role in rigid body pose estimation.Aiming at the problem of the performance de-gradation of geometric filtering under non-Gaussian conditions,a geometric filtering method based on the general-ized maximum correntropy criterion(GMCC)is proposed.Firstly,according to the evolution relationship of the geometric state on the manifold,the state prediction is performed using the unscented transformation on the mani-fold.Secondly,in order to suppress the adverse effects of non-Gaussian noise,the GMCC is extended to the mani-fold to correct the predicted state,thereby improving the robustness of the filtering.Then,for the manifold nonlin-ear optimization problem induced by GMCC,a statistical linearization method on the manifold is designed,and the optimization problem is solved by Riemannian manifold optimization and fixed-point iteration method.In particu-lar,an adaptive adjustment strategy for the generalized Gaussian kernel parameters is designed to adjust the hyper-parameters of the generalized correntropy online.Finally,simulation results demonstrate that,compared to existing methods,the proposed method has higher accuracy and robustness.关键词
几何滤波/卡尔曼滤波/广义相关熵/矩阵李群Key words
Geometric filtering/Kalman filtering/generalized correntropy/matrix Lie group引用本文复制引用
杨旭升,夏晓翠,金宇强,顾欣星,张文安..基于广义最大相关熵准则的几何滤波方法[J].自动化学报,2025,51(7):1612-1625,14.基金项目
国家自然科学基金(62473335,62173305),浙江省引进培育领军型创新创业团队(2023R01006),中国博士后科学基金(2024M752864),宁波市公益性研究计划重点项目(2023S018)资助Supported by National Natural Science Foundation of China(62473335,62173305),Leading Innovative and Entrepreneur Team Introduction Program of Zhejiang(2023R01006),China Postdoctoral Science Foundation(2024M752864),and Ningbo So-cial Public Welfare Research Key Project(2023S018) （62473335,62173305）
