基于随机有限集的多目标跟踪及航迹维持算法研究

发布时间：2018-01-16 13:35

本文关键词：基于随机有限集的多目标跟踪及航迹维持算法研究　出处：《江南大学》2017年博士论文　论文类型：学位论文

【摘要】：多目标跟踪技术作为信息融合理论与先进滤波方法中最活跃的研究领域之一,被广泛应用于以航空、航天为代表的军事与民用领域。由于无需传统跟踪方法中所需的复杂的数据关联技术,基于随机有限集理论的多目标跟踪方法备受国、内外相关研究领域学者及工程技术人员的广泛关注。本文以随机有限集理论为支撑,采用概率假设密度滤波器为主要工具,重点围绕复杂跟踪场景中多目标跟踪及航迹维持问题开展了较为深入、系统的研究工作,主要包括以下几个方面:1.针对紧邻目标跟踪场景中目标状态及数目估计问题,线性高斯假设下提出一种紧邻多目标GM-PHD跟踪算法。在标准GM-PHD滤波器各离散时刻目标预测强度量测更新结束后,采用目标权值再分配方法检测并重新分配目标后验强度中目标分量不合理的权值;在目标后验强度的分量删减阶段,提出一种融合了分量标记法和分量权值度量法的目标分量剪枝与融合方法,一定程度上能够避免重要目标分量的融合错误问题。与现有相关紧邻多目标PHD滤波器相比,提出算法具有较高的目标状态估计精度和准确的目标数目估计。2.针对多目标跟踪场景中新生目标先验强度未知时,标准PHD滤波器难以正确估计场景中目标状态及数目问题,提出一种基于新生目标强度自适应估计的多目标GM-PHD滤波算法。新生目标强度自适应估计方法分别利用PHD预滤波技术和目标速度特征方案,从各离散时刻量测集中获取最大可能的源于真实新生目标的量测集,然后利用这些量测建模未知的新生目标先验强度。此外,在标准GM-PHD滤波器的量测更新步引入量测驱动更新方案,将各离散时刻量测集划分存活目标量测集、新生目标量测集和杂波集。更新步中不同类型的目标预测强度采用其对应的量测集分别更新,且禁止杂波集更新目标预测强度。仿真实验结果表明,提出算法不仅具有较高的目标状态估计精度和较低的目标数目估计误差,而且具有稳定且较低的计算代价。3.为解决不精确检测概率环境下的多目标跟踪问题,提出一种融合指数衰减函数的目标权值更新策略和多帧目标状态抽取策略的多目标GM-PHD滤波算法。目标权值更新策略利用指数衰减函数及目标前一时刻权值,对由目标量测在状态空间中不合理分布导致的伪漏检目标的权值进行惯性衰减,以确保目标后验强度中各目标分量具有一个合理、有效的权值;多帧目标状态抽取策略利用各个目标若干个历史权值为参考,从各离散时刻目标后验强度中抽取由较低检测概率环境中因目标量测丢失导致的真漏检目标的状态估计。仿真实验表明,不精确检测概率环境下的多目标跟踪场景中提出算法具有较高的目标状态和数目估计精度,且滤波性能相对稳定。4.为了实现紧邻目标跟踪场景中多目标航迹维持,提出一种基于高斯混合概率假设密度滤波的多目标航迹维持算法。与经典GM-PHD跟踪器相比,所提多目标航迹维持算法滤波迭代中融合了目标状态关联与更新策略和不规则窗口多目标航迹管理方案。基于目标预测强度中目标分量的多个历史状态估计和当前量测集,目标状态关联与更新策略构建一个用于目标预测强度更新的关联更新因子矩阵。量测更新步中利用该关联更新因子矩阵实现了目标强度更新及目标与量测最优关联的同步。不规则窗口多目标航迹管理方法通过充分利用一段时刻内目标航迹的状态估计,不仅有效地维持了真实目标航迹的连续性,而且有效地解决了滤波过程中由虚警或杂波产生的虚假目标航迹。多目标交叉与平行跟踪场景仿真实验表明,提出的多目标GM-PHD航迹维持算法不仅能够改善紧邻目标状态及数目的估计精度,且其计算代价相对较小及具有优良的目标航迹维持性能。
[Abstract]:As one of the multiple target tracking information fusion theory and advanced filtering method in the most active research fields of technology, is widely used in aviation, military and civilian fields of space represented. Because of the complicated data association technique without the traditional tracking method, random finite set theory for multi target tracking method by country based on the extensive attention of scholars and related research in the field of engineering and technical personnel. Based on the random finite set theory, using probability hypothesis density filter as the main tool, focusing on complex scene tracking multiple targets tracking and track maintenance issues carried out in-depth, systematic research work, mainly including the following aspects: 1. for close to the target tracking in the scene and the number of target state estimation problem, a multi object tracking algorithm is proposed to GM-PHD under the assumption of quasi linear Gauss GM-PH in the standard. The discrete time D filter target prediction strength measurement update after the detection and re distribution of the target by using the target weight redistribution method posterior weights unreasonable target component strength; strength test component in the target after the deletion phase, this paper proposes a new mixed component labeling method and component weight measurement of target component pruning and fusion method, to a certain extent can avoid fusion component important targets. Errors associated with the existing close to the multi-objective PHD filter algorithm has higher than the target state estimation.2. for multi target tracking unknown new targets in the scene prior strength estimation accuracy and accurate target number, the standard PHD filter is difficult to correctly estimate the target state and the number of in the scene, put forward a new target strength estimation based on multi-objective adaptive GM-PHD filtering algorithm. The new target strength Adaptive estimation method using PHD pre filtering technique and target speed characteristics, concentrated to get the highest possible new target based on real measurement set from the discrete time measurement, and then use the new object prior strength measurement modeling unknown. In addition, the update step into the measurement test driver update scheme in the standard GM-PHD filter the amount of each discrete time measurement set division survival target measurement set new target measurement set and clutter update step set. Different types of target strength prediction using its corresponding measurement set were updated and forbidden clutter set to update the target strength prediction. Simulation results show that the proposed algorithm not only with higher target state estimation error estimation accuracy and lower number of targets, but also has.3. stable and low computational cost to solve the inaccurate multi target detection probability environment. The problem, put forward a target weight fusion index attenuation function and updating strategy of multi frame target state extraction strategy of multi-objective GM-PHD algorithm. The target weights update strategy using exponential attenuation function and the target weight for a moment ago, by the target measurement in the state space of unreasonable distribution leads to the pseudo target weight of inertia detection in order to ensure the target posterior attenuation, the target component strength has a reasonable and effective weight; multi frame target state extraction strategy using a plurality of weights for each objective historical reference, drawn from the lower strength test environment because the detection probability of target measurement caused by the loss of the target from the discrete time really missed the target state estimation. Simulation results show that the proposed estimation accuracy of target state and the number of algorithm has high detection probability is not accurate under the environment of multi target tracking in the scene, and The filtering performance is relatively stable.4. in order to achieve close to the target tracking problem of multi target track maintenance, proposes a maintenance algorithm of multi target track Gauss mixture probability hypothesis density based filtering. Compared with the classical GM-PHD tracker, the target tracking algorithm to maintain Titus filter iterative integration track management scheme associated with the target state update strategy and irregular window multi target. Target prediction of multiple target state estimation history component in the intensity and the current measurement set based on target state associated with the update strategy to build a target for updating the predicted strength of the updating factor matrices. Update factor matrix to achieve the target strength and target and measurement update using the optimal relevance correlation measurement update step in the irregular window. Multi target track management method by making full use of the target track within a period of time the state estimation, Not only effectively maintain the continuity of the real target track, and effectively solve the false target track caused by false alarm or clutter filtering. In the process of multi-objective cross and parallel tracking show scene simulation, multi objective GM-PHD algorithm is proposed for track maintenance can not only improve the estimation accuracy and the number of targets close to the state, and the computational cost is relatively small and has excellent performance to maintain the target track.

【学位授予单位】：江南大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TN713

【参考文献】