面向复杂系统估计的采样型非线性滤波方法研究

With the means of perception increasingly richer, perception ability enhanced and the deeped understanding of perception object as well as the gradually increasing of perception requirements, it is difficult to avoid the problems such as nonlinear, non-Gaussian, noise correlated, distributed and uncertain problems in the estimation of complex system. Combining with the advantages of sampling nonlinear filter, the design and optimization of filters are studied for noise correlated, distributed and uncertain features appeared in complex system. Some important research work will be included in this project as following: First, based on the realization principle of stochastic sampling filters, meanwhile, by virtue of the reconstruction of system state model, a sampling nonlinear filter with correlated multisensor measurement noise and process noise is to be studied. Second, according to the structural features of estimated system and the advantages of deterministic and stochastic sampling, a distributed sampling nonlinear filter based on comprehensive deterministic-stochastic sampling is to be studied. Third, based on the construction of measurement Jump/Switch Markov model, meanwhile, combining with Markov Chain Monte Carlo sampling, re-sampling technique and weight fusion, a sampling nonlinear filter with particle weight optimization in multi-sensor measurement uncertainty is to be studied. Finally, based on the application background of multi-sensor maneuvering multi-target tracking, the specific characteristics of systems are to be analyzed, and the relevant algorithms are to be optimized.

伴随着感知手段日趋丰富、感知能力日益增强，以及对感知对象认识的不断深入和感知要求的日益提高，人们在复杂系统估计中越来越难以回避非线性、非高斯、噪声相关、分布式和不确定等问题。本项目结合采样型非线性滤波器的自身优势，针对复杂系统呈现出的噪声相关、分布式以及不确定等特征开展滤波器设计和优化工作。主要内容包括：第一，拟从随机采样型滤波器实现原理入手，结合系统状态模型重构技术，研究多传感器量测噪声和过程噪声相关的采样型非线性滤波器；第二，拟结合被估计系统的结构特点以及确定性和随机性采样的优点，研究基于确定-随机采样综合的分布式采样型非线性滤波器；第三，拟在量测跳变马尔可夫模型构建基础上，结合马尔科可夫蒙特卡罗采样、重采样和加权融合技术，研究多传感器量测不确定下基于粒子权重优化的采样型非线性滤波器；最后，以多传感器机动多目标跟踪为应用场景，进一步分析系统所呈现的具体特征，优化算法设计。

本项目针对复杂系统估计中日益呈现的非线性、非高斯、多传感器、分布式和不确定等特征，以采样型非线性滤波理论的发展和完善为突破口，以多传感器机动多目标跟踪为应用背景，以仿真验证为平台实现理论和实践相结合、相验证、相促进，开展了采样型非线性滤波器的设计和优化工作。主要贡献如下：.针对Marginalized粒子滤波中随机量测噪声对于非线性状态估计精度的不利影响以及线性状态估计中计算量较大问题，提出了基于权重一致性优化的实时Marginalized粒子滤波算法。针对单传感器量测环境下量测噪声随机性对滤波精度的不利影响，提出了基于量测提升策略的卡尔曼滤波算法，考虑到工程应用中实时性、准确性以及鲁棒性等需求，设计了分布式加权融合和集中式一致性融合的两种实现结构。.在机动目标跟踪中，用于模型辨识和状态估计的非线性滤波器的合理选择和优化是提升滤波精度的关键，融合量测迭代更新集合卡尔曼滤波和交互式多模型方法，提出了基于量测迭代更新集合卡尔曼滤波的机动目标跟踪算法。此外，针对非线性高斯场景下目标数目未知或随时间变化的机动多目标跟踪问题，提出了基于交互式多模型的不敏卡尔曼概率假设密度滤波算法。针对卡尔曼一致滤波的应用受限于被估计系统需满足线性条件的问题，通过容积卡尔曼滤波和一致性策略的动态结合，提出了容积卡尔曼一致滤波算法。针对现有的利用非线性滤波算法对神经网络进行训练中存在易陷入极小值点以及滤波精度受限的缺陷，提出了基于容积卡尔曼滤波的神经网络训练算法。针对量测信息中系统误差对目标状态估计精度造成的不利影响，结合传感器量测的物理特性，提出了基于容积卡尔曼滤波的系统误差与目标状态联合估计算法。针对标准集合卡尔曼滤波实现过程中，量测噪声不确定导致虚拟量测采样出现一致性偏差问题，提出了基于Metropolis-Hastings采样的多传感器集合卡尔曼滤波算法。