概率假设密度滤波的谱聚类目标状态提取方法  

A Multi-Target State Extraction Method for the Probability Hypothesis Density Filter Based on Spectral Clustering

作　　者：张慧[1,2,3] 韩崇昭[1,2,3] 闫小喜[1,2,3] 

机构地区：[1]西安交通大学智能网络与网络安全教育部重点实验室,西安710049 [2]西安交通大学机械制造系统工程国家重点实验室,西安710049 [3]西安交通大学电子与信息工程学院,西安710049

出　　处：《西安交通大学学报》2012年第2期1-5,118,共6页Journal of Xi'an Jiaotong University

基　　金：国家重点基础研究发展计划资助项目(2007CB311006);国家自然科学基金创新研究群体科学基金资助项目(60921003);国家自然科学基金资助项目(61074176)

摘　　要：提出了一种谱聚类目标状态提取方法来实现概率假设密度(PHD)滤波中序贯蒙特卡罗(SMC)实现方式的多目标状态估计.该方法利用PHD滤波SMC实现方式输出的大量的加权粒子点间的相似度关系建立相似矩阵,通过变换得到拉普拉斯矩阵,进而对拉普拉斯矩阵进行特征分解,以实现粒子点的聚类,再在每类中寻找粒子的聚类点作为多目标状态的估计值,同时为了减小计算量,利用Nystrm逼近方法求解特征向量.仿真实验表明,PHD滤波的谱聚类目标状态提取方法的估计精度比k均值目标状态提取方法提高了60%以上.A spectral clustering target state extraction method is proposed to obtain multi-target state estimate in the sequential Monte Carlo（SMC） implementation of probability hypothesis density（PHD） filter.An affinity matrix is established based on the similarities among a large number of particles that are output of SMC implementation of PHD filter,and the Laplace matrix of the affinity matrix is obtained through Laplace transform.Then spectral clustering is performed through eigen-decomposition of the Laplace matrix,and the clustering center of each cluster is searched and is regarded as the estimation of multi-target states.Moreover,the Nystrm approximation method is used to generate eigenvectors so that the computation complexity of the algorithm is reduced.Simulation results show that the estimation accuracy of the proposed method is 60% higher than that of the k-means target state extraction method.

关 键 词：多目标状态估计 概率假设密度 状态提取 谱聚类 Nystrm逼近 

分 类 号：TP274[自动化与计算机技术—检测技术与自动化装置]