复杂条件下多目标跟踪关键技术研究

发布时间：2018-08-18 17:56

【摘要】：鉴于多目标跟踪技术在信息感知领域的重要地位,大量研究者多年来一直持续对多目标跟踪技术进行研究。目前针对协作式目标的跟踪技术已经比较成熟,针对一般非协作目标的跟踪技术也正在完善之中,但是针对典型对抗性非协作军事目标的跟踪技术仍面临诸多困难。这些困难或源自目标和环境特性,或源自传感器本身。本文以典型多目标跟踪系统面临的复杂目标、环境和传感器观测条件下多目标跟踪需求为牵引,对多目标跟踪方法进行了系统深入的学习、研究和探索,论文主要工作如下:第二章简要介绍了传统多目标跟踪方法、基于随机有限集(Random Finite Set,RFS)的多目标跟踪方法和多目标跟踪性能评估方法的理论基础,为后续章节论述做好铺垫。首先介绍传统多目标跟踪方法,给出了单目标贝叶斯滤波的具体推导过程,阐述了Kalman滤波算法与单目标贝叶斯滤波的关系,并解释了传统多目标跟踪方法如何通过数据关联技术,将多目标跟踪问题分解为若干并行单目标贝叶斯滤波问题。其次介绍有限集统计学(Finite Set Statistics,FISST)和多目标贝叶斯滤波,并给出多目标贝叶斯滤波矩近似的推导方法及迭代逻辑。最后对多目标跟踪性能评估的目的和原理进行阐述,介绍了传统类评估方法和基于综合度量的评估方法,并分析了各种评估方法的优缺点。第三章针对经典联合概率数据关联(Joint Probabilistic Data Association,JPDA)算法在目标密集时存在的航迹合并问题,提出了基于状态偏差估计和去除的方法,并研究了使用目标属性信息辅助的方法。基于偏差估计和去除方法仅使用目标状态信息,在构建目标-目标关联假设的基础上给出JPDA算法目标状态估计偏差的计算逻辑,进而去除偏差得到无偏JPDA算法;其与现有尝试解决航迹合并问题算法的仿真结果对比,表明了该算法的有效性。基于属性信息辅助的JPDA算法要求传感器能够提供目标属性信息,且仅在密集目标间属性不一致时才能实现航迹合并的有效抑制。本章在属性辅助JPDA算法方面的研究主要侧重于属性关联度量及门限的设计方面,提出了一种基于奈曼-皮尔逊(Neyman Pearson,NP)准则的属性关联度量及门限确定方法,用以克服传统固定门限所存在的关联性能不稳定问题。该方法确定的门限是航迹属性后验概率矢量和传感器目标属性区分性能的函数,可使漏检概率达到或尽可能接近预设值,对属性辅助数据关联中属性门技术研究具有相当的参考价值。第四章针对经典势概率假设密度(Cardinalized Probability Hypothesis Density,CPHD)滤波器不能处理标准多目标马尔科夫模型中的衍生目标模型问题,基于FISST推导出了考虑衍生目标模型的一般CPHD滤波器迭代公式,并与现有尝试解决该问题的若干方法进行了对比和分析,证明了现有方法仅是所提出方法的特例。推导过程使用了Faàdi bruno’s行列式规则,并提出了高阶Faàdi bruno’s行列式的可行迭代求解方法,使得所提出的一般CPHD滤波器迭代公式能够方便工程实现。仿真结果表明了所提出方法的有效性。第五章提出了一种适用于非线性观测条件的二项分裂高斯混合无迹Kalman概率假设密度(Binomial Splitting Gaussian Mixture Unscented Kalman Probability Hypothesis Density,BSGM-UKPHD)滤波器,使得高斯混合概率假设密度(Gaussian Mixture Probability Hypothesis Density,GM-PHD)滤波器的优异性能在非线性观测条件下依然能够得到保持。该算法对预测概率假设密度(Probabilistic Hypothesis Density,PHD)的每一高斯分量的非线性度进行计算和评估,当非线性度大于某一预设门限时,对高斯分量进行二项分解,于是得到一族非线性度较小的高斯分量,从而使得非线性观测引起的状态更新误差得到有效抑制,也就使得PHD算法优异性能在非线性观测条件下得到保持。仿真结果表明,提出方法性能显著优于两种传统方法。第六章提出了一种适用于集中式异步异类观测条件的真实更新时间高斯混合概率假设密度(Real Update Time GM-PHD,RUT-GM-PHD)算法。首先分析了集中式异步异类观测条件下多目标跟踪算法难以实施的本质原因,发现问题关键在于一般的目标运动模型和观测模型难以准确描述异步异类这种复杂的观测条件,进而对PHD高斯分量引入更新时间标记,从而提出RUT-GM-PHD算法。两个较为简单的异步异类观测条件下的仿真结果表明了所提出方法的优良性能。最后,阐述了一般异步异类观测条件下实施RUT-GM-PHD算法需要注意的若干问题,并指出了潜在可行解决途径。
[Abstract]:In view of the importance of multi-target tracking technology in the field of information perception, a large number of researchers have been studying multi-target tracking technology for many years. At present, the tracking technology for cooperative targets is relatively mature, and the tracking technology for general non-cooperative targets is also being improved, but for typical antagonistic non-cooperative targets. Military target tracking technology is still facing many difficulties. These difficulties arise either from the target and environment characteristics or from the sensor itself. In this paper, the multi-target tracking method is studied systematically and thoroughly based on the complex target that typical multi-target tracking system faces, and the multi-target tracking requirements under the environment and sensor observation conditions. The main work of this paper is as follows: Chapter 2 briefly introduces the traditional multi-target tracking methods, the theoretical basis of the multi-target tracking method based on Random Finite Set (RFS) and the multi-target tracking performance evaluation method, paves the way for the following chapters. The derivation process of standard Bayesian filtering is described. The relationship between Kalman filtering algorithm and single-target Bayesian filtering is expounded. The traditional multi-target tracking method is explained how to decompose the multi-target tracking problem into several parallel single-target Bayesian filtering problems by data association technique. Secondly, finite set statistics is introduced. Ics, FISST) and multi-target Bayesian filtering are presented, and the derivation method and iterative logic of multi-target Bayesian filtering moment approximation are given. Finally, the purpose and principle of multi-target tracking performance evaluation are described, the traditional class evaluation method and the evaluation method based on comprehensive measurement are introduced, and the advantages and disadvantages of various evaluation methods are analyzed. Aiming at the track merging problem of classical Joint Probabilistic Data Association (JPDA) algorithm when targets are dense, a method based on state bias estimation and removal is proposed, and a method aided by target attribute information is studied. Based on the hypothesis of target-target association, the calculation logic of target state estimation bias of JPDA algorithm is given, and then unbiased JPDA algorithm is obtained by eliminating the bias. Compared with the simulation results of existing algorithms attempting to solve track merging problem, the effectiveness of the algorithm is demonstrated. In this chapter, the research of attribute-assisted JPDA algorithm mainly focuses on the design of attribute association measures and thresholds, and proposes an attribute association measure based on Neyman Pearson (NP) criterion. Threshold determination method is used to overcome the instability of correlation performance in traditional fixed threshold. The threshold determined by this method is a function of the posterior probability vector of track attributes and the distinguishing performance of sensor target attributes. It can make the probability of missed detection reach or approach the preset value as far as possible. In the fourth chapter, the iterative formula of CPHD filter considering the derivative target model is deduced based on FISST for the problem that the classical potential probability Hypothesis Density (CPHD) filter can not deal with the derivative target model in the standard multi-objective Markov model. Several methods for solving this problem are compared and analyzed, and it is proved that the existing methods are only special cases of the proposed methods. The Fa di bruno's determinant rule is used in the derivation process, and a feasible iterative method for solving the high-order Fa di bruno's determinant is proposed, which makes the iteration formula of the proposed general CPHD filter convenient for engineering. Simulation results show the effectiveness of the proposed method. In Chapter 5, a binomial splitting Gaussian Mixture Unscented Kalman Probability Hypothesis Density (BSGM-UKPHD) filter is proposed to make the Gaussian mixture approximate. The excellent performance of the Gaussian Mixture Probability Hypothesis Density (GM-PHD) filter can still be maintained under nonlinear observation conditions. The algorithm calculates and evaluates the nonlinearity of each Gaussian component of the predicted probability hypothesis density (PHD) when the nonlinearity is large. The binomial decomposition of the Gaussian component at a preset threshold results in a family of Gaussian components with less nonlinearity, which effectively suppresses the state update error caused by nonlinear observations, and consequently maintains the excellent performance of the PHD algorithm under nonlinear observation conditions. In Chapter 6, a Real Update Time GM-PHD (RUT-GM-PHD) algorithm is proposed for centralized asynchronous observation. First, the essential reason why the multi-target tracking algorithm is difficult to implement under centralized asynchronous observation is analyzed. The key of the problem is that it is difficult to describe the asynchronous and asynchronous observational conditions accurately in the general target motion model and observation model, and then the RUT-GM-PHD algorithm is proposed by introducing the update time marker to the PHD Gaussian component. After that, some problems needing attention in implementing RUT-GM-PHD algorithm under general asynchronous and asynchronous observation conditions are expounded, and the potential feasible solutions are pointed out.
【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

【相似文献】