【作者】 罗飞腾；

【导师】 王东进；

【作者基本信息】 中国科学技术大学 ， 电磁场与微波技术， 2010， 博士

【摘要】 目标跟踪中存在一些强非线性、非高斯的状态估计问题,粒子滤波是解决这类问题的有效方法。本文以粒子滤波为核心,对粒子滤波的关键技术和几个跟踪问题,在理论和实现技术上展开研究。针对非高斯后验概率分布在重采样时产生的贫化问题,提出了基于分区近似思想的随机矢量产生算法,给出了关键参数的选取方法。该算法通过随机线性组合分区内的样本产生新的不重复的随机矢量,无需后验概率分布的先验信息,因此更适于描述非高斯分布。仿真结果验证了该算法的有效性。利用雷达采样单元的关联性,研究了基于雷达原始数据的目标跟踪性能及其粒子滤波实现方法。该项研究根据雷达信号在空间域、时间域、频率域的扩展特性,建立了雷达采样单元间的关联性模型,推导并分析了基于关联性模型的目标跟踪后验克拉美罗限,给出了相应的粒子滤波实现流程。另外,在幅度未知的情况下,提出了两种避免对幅度维采样的粒子滤波实现方法。仿真结果表明,利用雷达采样单元间关联性的目标跟踪性能优于基于量测的传统目标跟踪性能,且可以通过粒子滤波有效实现。面向机动目标检测前跟踪(TBD)问题,提出了一种分离模型变量的多模型粒子滤波算法。该算法通过目标状态与模型变量的分离,独立地估计模型概率和给定运动模型的目标状态,使得模型概率与粒子数无关,从而可高效地为每个模型分配充足的粒子数,避免了模型切换时发生粒子退化。仿真结果表明,在低信噪比情况下,该算法能显著提高多模型概率的估计性能和目标状态的估计精度,得到良好的TBD检测性能。研究了基于随机集的多目标跟踪理论,改进了其粒子滤波实现：1)通过分析概率假设密度(PHD)滤波器估计目标数的机理、误差来源及影响因素,提出了一种改进的PHD粒子滤波目标数估计、状态估计算法。该算法根据状态可分辨的原则划分状态空间,根据局部目标数判断检测状态,进而采取不同的处理策略,改善了低检测概率下目标数估计和目标状态估计的性能。2)初步探讨了基于PHD的多目标TBD,抽象出一类具有非标准观测模型的PHD滤波问题,指出现有理论在解决这一问题时的约束,并探讨了可能的解决途径。更多还原

【Abstract】 Highly non-linear and non-Gaussian estimation problems are ubiquitous in target tracking, and particle filter (PF) is an effective tool for such problems. In this dissertation, a key technique in PF and several tracking problems are studied in the framework of PF. The main contributions are as follows:In order to handle the impoverishment caused by resampling a non-Gaussian posterior distribution in PF, a random vector generation algorithm is proposed based on the idea of local approximation, in which new non-repeated random vectors are generated by random linear combination of given samples in each region, and the selections of key parameters are discussed. Since no prior information and assumptions about the underlying distribution are required, it is suitable for simulating pseudo-random vectors from a non-Gaussian distribution, and its validity is verified in the simulation.The performance and implementation of radar tracking by exploiting correlativity of radar sampling cells based on raw data are studied. Based on the correlativity model that is built according to the expansion of beam and signal in space and delay-Doppler domain, tracking performance is derived and analyzed, and the PF implementation is introduced. Furthermore, two amplitude-sampling-free implementations for the case of unknown amplitude are also proposed. The simulation results show that radar tracking by exploiting correlativity of sampling cells outperforms the traditional tracking based on abstracted target parameters, and the tracking scheme can be implemented by PF effectively.A multiple model PF with separated model variable is proposed in the research of maneuvering target track-before-detection (TBD). By separating the kinematic state and model variable, the model-conditioned target state and model probability are estimated independently, and thus the constraint between model probability and particle number of individual model is eliminated, which leads to the flexibility of efficiently assigning adequate particles for each model, and prevents degeneracy when model transforms. The simulation results show that the algorithm improves the performance of detection, model estimation, and tracking precision evidently when SNR is low. Multiple target tracking based on random set is studied and its PF implementation is improved:1) According to the analysis to the principle of target number estimation of probability hypothesis density (PHD) filter and related factors, an improved target number estimation and state estimation algorithm for particle PHD filter is proposed. In this algorithm, the space is first divided according to resolvability, and then the detection state is judged according to the local target number estimation, and further processing is taken based on the decision. The algorithm improves the performance of target number estimation and state estimation when detection probability is low.2) A generic of PHD filter with non-standard measurement model is proposed in the preliminary research on multitarget TBD, and the difficulties of this problem and possible solution are discussed.更多还原