随机有限集多目标跟踪技术研究

发布时间：2017-12-28 03:19

  本文关键词：随机有限集多目标跟踪技术研究　出处：《国防科学技术大学》2016年博士论文　论文类型：学位论文

  更多相关文章： 随机有限集 概率假设密度滤波器 序贯蒙特卡罗 多传感器偏差校准 联合目标跟踪与识别 多群目标跟踪 图像目标跟踪 后验克拉美罗下界

【摘要】：多目标跟踪的主要任务是依据传感器获取的含噪数据来联合估计多目标个数及其运动状态或者运动航迹,性能稳定且高效的多目标跟踪算法是多目标跟踪技术和多目标跟踪系统研究的核心,也是本课题研究的出发点和追求目标。目前,多目标跟踪技术正朝着能处理目标数目未知可变、检测不确定、观测源不确定、数据关联不确定等复杂多目标跟踪问题的方向蓬勃发展,其中尤以Ronald Mahler提出的基于随机有限集(Random Finite Set,RFS)的一类多目标跟踪方法对这些复杂的场景具有天然的适应性,不需要进行复杂的数据关联处理即可对目标个数未知且时变的多个目标进行联合检测与跟踪是此类跟踪算法的最大优势。基于RFS的多目标跟踪算法为目标监视与防御、无人驾驶与机器人、遥感、计算机视觉、生物医学、现代通信等领域包含的复杂多目标跟踪问题提供了新的解决途径,代表着多目标跟踪技术发展的新方向。本课题重点挑选随机有限集框架下联合多传感器偏差与多目标状态估计技术、随机集滤波器的航迹提取技术、基于随机集滤波器的多群目标跟踪技术、随机集框架下图像多弱目标检测前跟踪(Track-Before-Detect,TBD)技术及随机集滤波器的后验克拉美罗下界(Posterior Cramer-Rao Lower Bound,PCRLB)性能评估技术这五项研究内容进行深入研究,取得的主要研究成果如下:第二章提出了一种基于分层点过程及多群多目标概率假设密度(Multi-group Multi-target Probability Hypothesis Density,MGMT-PHD)滤波器的联合多传感器偏差与多目标状态估计算法。该算法将多传感器偏差集建模为父过程,多目标状态集则是与多传感器偏差相关联的子过程,通过分开对待两个相互交互的点过程,可以避免对高维增广状态联合估计产生的巨大计算量。在利用MGMT-PHD滤波器解决多传感器偏差和多目标状态的联合估计问题时,由于多传感器偏差的个数即为传感器的个数,即父过程的元素个数已知,且多个传感器独立收集观测,即观测集分割情况是明确的,提出了MGMT-PHD滤波器的粒子实现形式,实现了非线性条件下的联合多传感器偏差与多目标状态估计。仿真实验考虑了一个目标出现、目标消失、目标轨迹交叉事件出现的典型复杂多传感器多目标场景,验证了所提算法的有效性。第三章在载机与诱饵纵向可分辨的情况下,解决了末制导主动雷达导引头拦截战机对抗背景下对波束内载机与诱饵的联合快速检测、识别与稳定跟踪问题。主要贡献为:第一,对现有的加标签粒子PHD(Labeled Particle PHD,L-P-PHD)滤波器存在的一些局限进行改进,提出改进的L-P-PHD(Improved L-P-PHD,IL-P-PHD)滤波器;第二,结合现有的多模型技术,提出能同时对多个机动目标进行跟踪与航迹维持处理的多模il-p-phd(multiplemodelil-p-phd,mm-il-p-phd)滤波器;最后,基于mm-il-p-phd滤波器,结合基于回波幅度特征的干扰存在性检测方法以及对抗场景的特征信息,建立了纵向距离维可分的载机与诱饵的联合快速检测、稳定跟踪与识别处理框架。仿真实验表明,所提方法可以有效地实现对纵向距离维可分的载机与诱饵的快速检测、稳定跟踪与识别处理。第四章将多群目标建模为分层点过程,提出了一种基于随机有限集的新算法,该算法能联合估计群目标个数、估计群中心和群内组件的运动状态、提取群中心航迹。其基本思想及涉及到的主要工作与贡献为:第一,对不可分目标phd(unresolvedtargetphd,ut-phd)滤波器的观测更新过程进行了具体化,给出了ut-phd滤波器观测更新方程的具体计算方法,对ut-phd滤波器进行加标签处理,利用序贯蒙特卡罗技术实现了ut-phd滤波器,提出加标签的粒子ut-phd(labeledparticleut-phd,l-p-ut-phd)滤波器,l-p-ut-phd滤波器能在估计多群目标个数、多群目标中心状态的同时获取多群目标中心的运动轨迹,实现了多群目标中心的联合检测与跟踪;第二,基于群中心状态估计结果提出了更为精确的观测集分割算法,完成观测集分割,将观测集分割结果分配给每个群目标对应的单群粒子phd(single-groupparticlephd,sg-p-phd)滤波器,完成群内组件状态跟踪与个数估计,将群组件个数估计结果反馈至l-p-ut-phd滤波器。仿真实验表明,所提方法可以有效地检测群目标的出现与消失、估计群中心的运动状态、获取群中心的航迹及估计群内组件的运动状态与组件个数。第五章分别研究了影响区域不重叠和影响区域重叠的图像多弱目标tbd技术。针对标准phd-tbd算法存在对新生目标发现延迟较久、对目标个数估计不准且存在起伏的问题,提出了能解决这些问题的广义phd-tbd算法及其粒子实现。对于目标影响区域重叠的图像多弱目标tbd,包含的主要贡献与创新体现在:第一,建立了影响区域重叠的图像目标的叠加传感器观测模型,导出了对应的多目标观测似然函数;第二,基于建立的模型,将mahler提出的近似叠加phd(approximationsuperpositionalphd,as-phd)滤波器引入图像目标跟踪框架,对as-phd滤波器的状态空间进行加标签处理,提出了加标签as-phd滤波器,利用smc技术,提出了加标签as-phd滤波器的粒子实现,解决低信噪比下影响区域重叠的图像多弱目标跟踪问题。仿真实验验证了所提算法的有效性。第六章对基于随机有限集的滤波器处理复杂多目标跟踪问题时所能达到的性能下界及其计算实现问题开展研究。主要贡献为:第一,推导出了随机集框架下能适应目标数目未知可变、检测不确定、观测源不确定、数据关联不确定出现的复杂多目标跟踪问题的多目标pcrlb(multi-targetpcrlb,mt-pcrlb),及其递推计算表达式,用以获取多目标跟踪算法处理此类问题的性能下界;第二,基于IL-P-PHD滤波器获取的多目标航迹,提出了一种高精度的获取多目标航迹和观测集间关联关系的数据关联新方法;第三,基于获取的多目标航迹和数据关联新方法,导出了评估典型雷达多目标跟踪问题性能下界的MT-PCRLB的具体表达式。此外,该性能下界可以与目前流行的加标签随机集滤波器配套使用,基于加标签随机集滤波器获取的多目标航迹及观测集与航迹的关联关系,可以实现MT-PCRLB的递推计算。仿真实验表明,提出的MT-PCRLB确能定量地衡量处理复杂多目标跟踪问题的多目标跟踪算法所能达到的性能下界。第七章总结全文,并指出了下一步可能的研究方向。
[Abstract]:The main task of multi target tracking is based on noisy data from sensor to joint estimation of target number and its state of motion or motion tracking, multiple target tracking algorithm is stable and efficient is the core system of multi target tracking and multiple target tracking technology, is also the starting point of the research and the pursuit of the goal. At present, multiple target tracking technology is moving can deal with unknown target number variable, detection uncertainty, observation source uncertainty, data association uncertainty complex multi-target tracking problem of vigorous development, which is based on the random finite set especially Ronald proposed by Mahler (Random Finite Set, RFS) for a class of multi target tracking the method has the natural adaptability to the complex scenes, without the need for complex data processing can be carried out joint detection and tracking of multiple targets in a number of unknown and time is the biggest advantage of this tracking algorithm. Provides a new way to solve complex problem of multi target tracking multiple targets RFS tracking algorithm for targets containing surveillance and defense, unmanned robot, remote sensing, computer vision, biomedicine, modern communication based on the field, is a new direction for the development of multi target tracking technology. This research mainly focuses on the selection of random finite set under the framework of joint multi sensor multi target state estimation bias and track technology, random set filter extraction technology, random set filter multi target tracking based on image multi weak targets technology, random set framework of tracking before detection (Track-Before-Detect, TBD) and random set filter posterior Clarke (Posterior Cramer-Rao Lower Bound Rao lower bound, PCRLB) performance evaluation technology of the five studies in-depth study, the main results are as follows: chapter second presents a hierarchical point process and multi group multi-objective probability hypothesis density (Multi-group Multi-target Probability Hypothesis Density, MGMT-PHD) combined with multi sensor and multi bias filter an algorithm of target state estimation. The algorithm of multi-sensor deviation modeling process of multiple target state in the father, is associated with multiple sensor bias associated sub process, through the separate two interacting point processes, can avoid the high dimensional augmented state estimation of the large amount of calculation. To solve the problem of multi sensor joint estimation bias and multi target state in the use of MGMT-PHD filter, due to a number of multi sensor error is the number of sensors, the parent process the number of elements known, and a plurality of sensors that collect independent observation, observation set segmentation is clear, put forward the realization form of MGMT-PHD filter the particles for the combined multi sensor deviation under the condition of nonlinear and multi target state estimation. The simulation experiments consider a typical complex multi-sensor and multi-target scene with the appearance of the target, the disappearance of the target and the crossover of the target track. The validity of the algorithm is verified. The third chapter solves the problem of fast detection, recognition and stabilization of the target and the decoys in the terminal guided Active Radar Seeker under the condition of longitudinal resolution. The main contributions are as follows: first, the particle labelling of existing PHD (Labeled Particle PHD, L-P-PHD) some limitations of the existence of the filter is improved, an improved L-P-PHD (Improved L-P-PHD IL-P-PHD) filter; second, combined with the existing technology of multi model, which can simultaneously on multiple maneuvering target tracking and track maintenance of multimode il-p-phd the (multiplemodelil-p-phd, mm-il-p-phd) filter; finally, based on the mm-il-p-phd filter, combined with the characteristics of the existence of interference echo amplitude detection method and the confrontation scene feature information based on the establishment of a joint rapid detection, carrier and decoy stable tracking and recognition processing frame longitudinal distance separable. Simulation results show that the proposed method can effectively realize the rapid detection, carrier aircraft and decoy longitudinal distance separable stable tracking and recognition. In the fourth chapter, multigroup target is modeled as a hierarchical point process. A new algorithm based on stochastic finite set is proposed. It can jointly estimate the number of group targets, estimate the motion state of group centers and components within clusters, and extract group center tracks. The basic idea and relates to the main work and contribution are as follows: first, the target can be divided into PhD (unresolvedtargetphd, ut-phd) filter observation update process in detail, gives the calculation method of ut-phd filter observation update equation, the ut-phd filter with label processing, ut-phd filter is realized by sequential Monte Carlo technology, proposed tagged particle ut-phd (labeledparticleut-phd, l-p-ut-phd) filter, l-p-ut-phd filter can estimate the number of target multi group and multi group target center state by trajectory take multi group target center, implementation of joint detection and tracking of multiple group target center; second, group center state estimation results is presented. A more accurate segmentation algorithm based on the observation set, complete observation set segmentation, the segmentation results will be the observation set, single particle distribution corresponding to each pH group target The D (single-groupparticlephd, sg-p-phd) filter completes the state tracking and number estimation of components in the cluster, and returns the estimation results of group components to l-p-ut-phd filters. The simulation results show that the proposed method can effectively detect the appearance and disappearance of group targets, estimate the motion state of group centers, get the track of group center, and estimate the motion state and component number of components in the cluster. In the fifth chapter, the image multi weak target TBD technology, which affects the region does not overlap and affects the overlap of the region, is studied. The standard phd-tbd algorithm has a long delay in finding the new target, and the problem of the number of target numbers and the fluctuation of the target number.
【学位授予单位】：国防科学技术大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TN713
，

本文编号：1344359