点目标跟踪的非线性滤波算法研究

【作者】 庄泽森；

【导师】 张建秋；

【作者基本信息】 复旦大学 ， 电路与系统， 2009， 硕士

【摘要】 目标跟踪技术已经在包括军事和民用的不同的领域中得到应用,它是当今国际上十分活跃的热门领域之一,而目标跟踪的核心为滤波算法。如何提出性能更好的非线性滤波算法,来对付实际系统的非线性、非高斯问题,并高效的应到到目标跟踪系统中,是本领域研究的热点和难点所在。首先,在单目标跟踪系统,针对混合线性/非线性目标动态模型,提出一种新的混合滤波算法,算法采用Rao-Blackwellized思想,将线性状态与非线性状态进行分离,对非线性状态运用准高斯粒子滤波(quasi-Gaussian particle filtering,Q-GPF)算法进行估计,并将其后验分布近似为单个高斯分布,再利用非线性状态的估计值对线性状态进行卡尔曼滤波(Kalman filter,KF)估计。仿真结果表明新算法在精度不下降的前提下,计算复杂度大大下降。其次,针对多目标跟踪应用,提出一种新的基于随机集的滤波算法,算法运用Rao-Blackwellized思想,通过挖掘分析“混合线性/非线性模型”的结构,采用序列蒙特卡洛(sequential Monte Carlo,SMC)方法预测与估计概率假设密度滤波器(probability hypothesis density filtering,PHD)迭代式中各个目标的非线性状态,并利用非线性状态粒子中包含的线性状态的信息,使用KF对线性状态进行预测与估计。以更好地估计PHD进而提高各目标状态估计精度。分析与仿真的结果表明,新算法在减少计算量的同时,提升了估计精度。进一步,同样针对多目标跟踪提出一种新的概率假设密度滤波算法,算法在PHD滤波器迭代式计算之后,运用结合了mean-shift算法的核密度估计(kerneldensity estimation,KDE)理论进行PHD分布的二次估计、提取PHD峰值位置作为目标状态估计值。分析与仿真的结果表明新算法的估计精度有大幅的提高。更多还原

【Abstract】 The target tracking technique has been applied in different kinds of martial and civil areas.It is one of the scientific topics which draw lots of research interests nowadays.The key of target tracking is filtering algorithm.It is a hotspot and nodus in target tracking technique research to propose more reliable non linear filtering algorithms to cope with the non linear and non Gaussian problem,and applied in practical tracking system efficiently.Firstly,for single target tracking system,a novel mixed filtering algorithm is proposed for mixed linear/nonlinear state space models.The algorithm utilizes the idea of Rao-Blackwellized to separate the linear and nonlinear states.For the nonlinear states,the posterior distributions of the estimates,which are achieved by the quasi-Gaussian particle filter,are approximated as Gaussian distributions.Also,the linear states are estimated by the Kalman filter with the estimated nonlinear states. The simulation results show that the proposed method saves much computing time with no declined tracking accuracy performance.Next,we proposed a novel random set based filtering algorithm for multi-target tracking(MTT) application.The algorithm,while utilizing the idea of Rao-Blackwell to enhance the estimating performance of the probability hypothesis density(PHD), adopts the sequential Monte Carlo(SMC) method to predict and estimate the nonlinear states of the multiple targets.In addition,the linear states are estimated by the Kalman filter(KF) with the information embedded in the estimated nonlinear states.Simulation results of the proposed method show that,in addition to reducing particle dimensions and computation complexity,the proposed method significantly enhances tracking accuracyMoreover,another novel probability hypothesis density filtering algorithm is proposed for multi-target tracking applications.The algorithm utilizes the kernel density estimation theory and the mean-shift algorithm to further estimate the probability hypothesis density and then to extract target state estimates after the computation of the PHD recursive formula.The simulation results of the proposed method show that,the tracking accuracy of the proposed method is increased significantly.更多还原