概率假设密度高斯混合实现的分量删减  

Component Pruning in Gaussian Mixture Implementation of Probability Hypothesis Density

作　　者：闫小喜[1] 韩崇昭[1] 

机构地区：[1]西安交通大学电子与信息工程学院综合自动化研究所智能网络与网络安全教育部重点实验室机械制造系统工程国家重点实验室,西安710049

出　　处：《自动化学报》2011年第11期1313-1321,共9页Acta Automatica Sinica

基　　金：国家重点基础研究发展计划(973计划)(2007CB311006);国家自然科学基金创新研究群体科学基金(60921003);国家自然科学基金(61074176)资助~~

摘　　要：针对概率假设密度(Probability hypothesis density,PHD)高斯混合实现算法中的分量删减问题,提出了基于Dirichlet分布的分量删减算法以改进概率假设密度高斯混合实现算法的性能.算法采用极大后验准则估计混合参数,采用仅依赖于混合权重的负指数Dirichlet分布作为混合参数的先验分布,利用拉格朗日乘子推导了混合权重的更新公式.算法利用负指数Dirichlet分布的不稳定性,在极大后验迭代过程中驱使与目标强度不相关的分量消亡.该不稳定性还能够解决多个相近分量共同描述一个强度峰值的问题,有利于后续多目标状态的提取.仿真结果表明,基于Dirichlet分布的分量删减算法优于典型高斯混合实现中的删减算法.As far as component pruning in Gaussian mixture （GM） implementation of probability hypothesis density （PHD） is concerned, a component pruning algorithm based on Dirichlet distribution is proposed to improve the performance of Gaussian mixture implementation of probability hypothesis density. The maximum a posterior criterion is adopted for estimation of mixing parameters. Dirichlet distribution with negative exponent parameters, which only depends on mixing weights, is adopted as the prior distribution of mixing parameters. The update formulation of mixing weight is derived by Lagrange multiplier. The instability of Dirichlet distribution with negative exponent parameters is applied to driving the components irrelevant with target intensity to extinction during the maximum a posterior iteration. Besides, the problem that one peak of intensity is presented by several proximate mixing component, can be solved by this instability. It is useful for the following state extraction. Simulation results show that the component pruning algorithm based on Dirichlet distribution is superior to that of typical Gaussian mixture implementation.

关 键 词：概率假设密度 高斯混合实现 分量删减 Dirichlet分布 极大后验 

分 类 号：TN713[电子电信—电路与系统]