基于PHD滤波器的多目标跟踪方法研究

发布时间：2018-01-18 00:05

本文关键词：基于PHD滤波器的多目标跟踪方法研究　出处：《西北工业大学》2015年硕士论文　论文类型：学位论文

【摘要】：多目标跟踪技术已经广泛应用于军事和民用领域,但它仍是多学科、多领域所共同关心的重点和难点问题。近年来,在多目标跟踪问题中,随机有限集(Random finite set,RFS)方法颇受关注,而概率假设密度(Probability Hypothesis Density,PHD)滤波器作为随机有限集框架下的多目标完全概率密度函数一阶统计矩近似产物,解决了随机有限集的实际可执行度问题,且避免了数据关联。本文正是基于PHD滤波器,对多目标跟踪问题进行了深入研究,主要研究成果如下:1、基于PHD滤波器的全局航迹提取方法研究:针对PHD滤波器不能提供目标连续航迹信息的问题,提出了基于PHD滤波器的全局航迹提取算法。该算法考虑目标的全局信息,即考虑相邻两个时刻的全部目标状态估计点的关联性,提出同一时刻预测峰值和估计峰值之间的一致性度量及一致性置信度的概念,同时基于专家知识提出全局航迹提取策略,最后基于一致性置信度及构建的航迹提取决策规则一一提取航迹,实现PHD的全局航迹提取。仿真结果表明,该算法可以稳定跟踪目标,正确起始、维持及终结航迹,在航迹提取精度上有明显优势,且计算量相比是相当的。2、不均匀杂波环境及低检测概率下的改进自适应PHD滤波器设计:针对传统PHD滤波器在不均匀杂波环境及低检测概率下跟踪性能急剧下降的问题,提出了一种改进的自适应PHD滤波器,通过自适应确定杂波区,自适应选择量测及对权值较大的高斯项进行保护来保证算法的快速性和高精度。该滤波器首先利用AP聚类算法对监视区域内满足一定条件的多帧累积的所有回波进行聚类,用凸包确定杂波区,然后再逐观测时刻进行PHD预测和更新。在PHD预测时不用凸包里的回波,但在PHD更新时,需先自适应选择量测,而后进行PHD更新。同时,在该滤波器中,保护权值高的高斯项,保证其权值的稳定性。仿真结果表明,该滤波器可以很好的实现在不均匀杂波环境下和低检测概率情况下的目标跟踪,相比传统的PHD滤波器,改进了目标状态估计精度,提高了计算效率。3、用于地面机动目标跟踪的约束多模型PHD滤波方法研究:考虑地面目标运动受到地形环境等限制,将地理信息用于地面目标跟踪可有效提高跟踪精度。在地面目标跟踪中,将地理信息表示成等式约束的形式来修正目标状态,并采用多模型处理地面目标机动时运动模式的不确定,提出了一种用于地面机动目标的约束多模型PHD滤波器方法。该算法利用模型条件分布和模型的概率,使用多模型方法对GM-PHD滤波器中的每一个高斯分量进行预测和更新,然后将得到的估计值融合得到对应的目标状态,并将道路信息表示成等式约束形式,将约束加入修正目标估计状态,完成地面目标跟踪。仿真结果表明,本文算法可以在杂波环境下有效的估计地面机动目标的状态,相比于未用地理信息的MM-GMPHD方法及传统的GM-PHD滤波器,有效提高了目标状态估计精度。
[Abstract]:Multi target tracking technology has been widely used in military and civilian areas, but it is still much discipline, focus and difficult issues of common concern in many fields. In recent years, the problem of multiple target tracking, random finite sets (Random finite set, RFS) method is popular, while the probability hypothesis density (Probability Hypothesis Density. PHD) filter as a multi-objective stochastic finite set under the framework of the full probability density function of the first-order statistical moment approximation product solves the random finite set of actual execution problem, and avoid the data association. This paper is based on the PHD filter, the tracking problem of multiple targets is studied, the main research results are as follows: 1, research on global path extraction method based on PHD filter: provide continuous track information to solve the problem of PHD filter, the global path extraction algorithm is proposed based on PHD filter. This is By considering the global information of the target, which is considered the two adjacent time all target state estimation related point, put forward the same time prediction measure of consistency between the peak and the estimated peak and the same concept of confidence, and put forward the global strategy of expert knowledge extraction based on the track, the track based on consistency and build confidence the decision rules extraction one track, realize the global path extraction of PHD. The simulation results show that the algorithm can correct the tracking stabilization, initiation, maintenance and termination of the track, in the track extraction accuracy has significant advantages, and the amount of calculation is compared to.2, improved adaptive PHD filter design and heterogeneous clutter under low detection probability: Based on the traditional PHD filter tracking performance fell sharply in inhomogeneous clutter environment and low detection probability under the problem, this paper presents an improved adaptive P HD filter, determined by adaptive clutter region, adaptive choice of measurement and protection of the larger weight Gauss to ensure fast and high accuracy algorithm. By clustering all echo multi frame accumulation of the filter using AP clustering algorithm to monitor the area to meet certain conditions, determine the clutter region with a convex hull. Then by observing time PHD prediction and update. In the prediction of PHD without convex hull in echo, but in the PHD update, first choice adaptive measurement, and then update the PHD. At the same time, in the filter, the protection of Gauss high weights, ensure the stability of the weights. The simulation results show that the the filter can achieve good tracking in inhomogeneous clutter environment and low detection probability under the condition of the target, compared with the traditional PHD filter, improved the accuracy of target state estimation, improve the computational efficiency of.3 for ground Research on constraint multi model PHD filtering method for maneuvering target tracking: the ground moving target by considering the terrain environment, geographic information for ground target tracking can effectively improve the tracking accuracy. In ground target tracking, geographic information is expressed as the form of correction constrained target state, uncertain and multi model ground targets maneuvering motion model, a method is presented for ground maneuvering target constrained multiple model PHD filter. The algorithm uses the probability model and conditional distribution model, using the method of multi model prediction and update of each Gauss component in GM-PHD filter, and then the estimated value corresponding to the target state fusion, and road information will be expressed as equality constraints, the constraint is added to the modification of target state estimation, the completion of the ground target tracking. The simulation results show that the The algorithm can effectively estimate the state of ground maneuvering target in clutters, and effectively improve the accuracy of target state estimation compared with the MM-GMPHD method without traditional geographic information and the traditional GM-PHD filter.

【学位授予单位】：西北工业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN713

【相似文献】