【作者】 赵玲玲；

【导师】 马培军；

【作者基本信息】 哈尔滨工业大学 ， 计算机应用技术， 2011， 博士

【摘要】 点目标跟踪问题是指在杂波环境中,根据受噪声干扰的传感器测量序列,对可以忽略形状、颜色等外观信息的目标状态进行估计,并估计出目标航迹的过程。点目标跟踪在定位、导航、工业控制等军事和民用领域有着广泛的应用。根据监测区域内目标的个数,点目标跟踪可以划分为单目标跟踪和多目标跟踪,单目标跟踪中如何利用测量进行有效的滤波,实时给出准确的状态估计是其研究的核心内容;根据目标数目的情况,多目标跟踪又分为数目已知且固定的多目标跟踪,数目未知且随时间变化的多目标跟踪,后者在实际中有着更为广泛的应用范围和更高的技术难度。在存在杂波对测量的干扰和影响的情况下,利用带有噪声的传感器测量,对运动方式复杂多变的目标进行连续跟踪具有很大的不确定性,是十分具有挑战性和具有重要研究价值的课题。本文分别针对单目标和未知时变的多目标跟踪问题进行了研究。对于单目标跟踪问题,由于不必考虑测量与目标的关联关系,因此以跟踪算法即滤波算法研究为核心,针对具有典型非线性特征的雷达测角等测量模型,采用以粒子滤波为核心的滤波跟踪方法,重点解决粒子滤波算法中的计算效率问题,同时通过解决样本枯竭问题,降低重采样算法中的多样性损失,提高粒子滤波的估计精度,从而提高对单目标的跟踪性能;对于未知时变的多目标跟踪问题,则以基于概率假设密度(Probability Hypothesis Density, PHD)滤波的目标状态估计和连续跟踪方法为主要研究内容,通过推导权值域的单目标PHD分解形式来估计多个目标的状态,同时结合粒子标签和航迹管理来形成目标的连续航迹。首先,在基于粒子滤波的单目标跟踪算法研究方面:(1)针对粒子滤波为解决退化问题引入重采样而导致的样本枯竭和多样性损失问题,提出了基于准蒙特卡罗(Quasi-Monte Carlo, QMC)的采样重采样算法,用在大权重样本邻域内生成具有低差异性的QMC序列来代替一般重采样算法中对大权重样本的复制过程,一方面解决了重采样后样本的枯竭和多样性损失问题,降低样本的目标失跟概率;另一方面利用QMC序列的低差异性,获取高于一般蒙特卡罗采样方法的估计精度,从上述两个方面提高粒子滤波的跟踪精度。(2)针对粒子滤波的样本容量大,跟踪效率受样本容量影响明显的问题,在多分辨粒子滤波的基础上,提出基于滤波系统状态检验的样本容量控制算法。利用粒子滤波的样本集在空间域中的多分辨分解形式,在相似样本中提取关键样本,以降低总样本容量。为了监测系统是否出现滤波失效,定义了粒子滤波的系统拟测量误差等相关统计量,通过检测系统状态,对样本容量进行自适应调节,保证滤波精度的稳定,同时提高粒子滤波的跟踪效率。其次,在基于粒子概率假设密度滤波的多目标状态估计与连续跟踪方法研究方面:(1)针对概率假设密度滤波过程仅输出多目标分布的概率密度,无法直接输出目标的状态估计的问题,提出粒子单目标PHD滤波方法,通过构造PHD样本权值向量,从PHD的状态预测和测量更新公式中推导出单目标的粒子PHD分解公式,从混合的多目标后验PHD中同时分离出单个目标的后验PHD形式,从而利用粒子权值分量描述的单目标PHD实现所有目标的状态提取,从原理层面上解决基于粒子概率假设密度滤波的多目标状态估计问题。(2)针对PHD滤波的输出状态的集合无序性,无法给出目标时间域上的状态关联问题,在提出的单目标PHD滤波的基础上,提出一种结合标签的粒子层状态关联方法,定义了基于权值向量的状态间关联矩阵,同时建立了结合一步预测的关联规则,通过追踪样本权值向量在测量域上的移动来判定目标状态的前后帧关系,进而给出目标状态的标识,形成连续航迹;在包含杂波的广义状态估计的基础上,构建了状态假设、伪航迹剪枝、状态估计反馈的环状粒子概率假设密度滤波跟踪器模型,解决了数目未知时变的多目标的连续跟踪问题。更多还原

【Abstract】 Punctual target tracking problem is to estimate the target states, whose appearance features such as shape and color can be ignored, and provide their trajectories from a discrete set of noisy measurements in clutter. Punctual target tracking are wildly applied in military and civil fields such as location, navigation, industrial control and so on.Punctual target tracking is decomposed of single-target tracking and multiple target tracking according to the number of targets existing in the survillance region. Besides, multi-target tracking scene includes a fixed and known number of target tracking and a time-varying and unknown number of target tracking, obviously, the latter has a wilder application and is harder to resolve. Furthermore, when targets move complicatedly in clutter, it is a challenge and valuable issue to obtain their continuious trajectories from noisy measurements.This dissertation mainly addresses the issues that single target and a time-varying unknown number of targets tracking problem. In single target tracking, since data association is not necessary to consider, the key issue is the tracking algorithm, that is, filter algorithm. This dissertation uses particle filter to handle target tracking problem with nonlinear or non-Gaussian dynamic model from some classical sensors such as radar. In order to improve the performance of particle filter in single target tracking, we resolve the sample impoverishment, improve the computational efficiency and reduce the loss of variation of samples in resampling step. In the field of an unknown time-varying number of targets tracking, this dissertation focuses on the probability hypothesis density (PHD) filter and mainly addresses the issues of multi-target state estimation and track continuity. The single-target PHD decomposition is constructed in weight domain. Based on single-target PHDs from particle-PHD filter, this dissertation estimates the states of targets and identifies target trajectories by combining particle-labelling association and track management.Firstly, in single-target tracking based on particle filter,(1) Aimming at the sample impoverishment caused by rempling algorithm, this dissertation proposes a new resampling algorithm exploiting quasi-Monte Carlo (QMC) method. This algorithm generates QMC sequences around the particles with heavy weights instead of multiplying them. This scheme improves the tracking performance of particle filter from two aspects, on one hand; it solves the impoverishment of samples caused by resampling and reduces the probability of target miss-tracking. On the other hand, the estimation accuracy can be improved by taking advantage of low-discrepancy of the QMC sequence.(2) A large number of particles propogated in the iterations of particle filter result in low tracking efficiency. Based on multi-resolutional particle filter, this disserstein proposes a sample-size control method combining with filtering performance detecting. This algorithm extracts representative particles from all sets of similar particles and reduces the number of particles. Several statistic values are further defined, including quasi-measurement error, to detect whether the particle filter falls into failure. Based on these performance parameters, a sample set control algorithm for a multi-resolvational particle filter is proposd. In this scheme samples are increased if the failure has been proved to happen, which maintains the performance of filtering, and meanwhile improves the efficiency of tracking.Secondly, in particle probability hypothesis density (PHD) filter for multi-target state estimation and track continuity,(1) Particle filter implementation of the PHD filter has demonstrated a feasible suboptimal method for tracking multi-target in real-time. To obtain the target states, the peak-extraction from the posterior PHD particles needs to be implemented. This disserstein derives a decomposition of single-target PHD form. Based on it, a new state estimation method is proposed, which doesn’t need to extract the PHD peaks by clustering analysis. The method provides a single-target PHD expression derived from the updated PHD equation. The target states can be directly estimated from the single-target PHD sequentially.(2) In the aspects of track continuity, PHD filter can not provide the track-valued estimates. It is a drawback of PHD filter when it is necessary to identify the individual target trajectories. State association among frames is needed to estimate the individual target trajectories. A new particle PHD filter tracker is proposed. The method extends particle-labeling association to the single-target PHD filter, associates target location estimates among time frames by combining particle-labeling association with state prediction. Furthermore, in order to filter out clutter, a new single-target PHD filter tracking model is constructed, which includes state candidate estimation, pruning, and state estimates feedback. This framework resolves the problem of multi-target tracking continuity.更多还原