【作者】 刘亮；

【导师】 马华东；

【作者基本信息】 北京邮电大学 ， 计算机应用技术， 2009， 博士

【摘要】 目标定位跟踪技术是无线传感器网络众多实际应用的基础,具有十分重要的研究意义。在传统无线传感器网络中,目标定位跟踪技术已经积累了较多的研究成果。近年来,新型视频传感器网络逐渐成为无线传感器网络领域的研究热点,其感知信息丰富的特点更有利于实现目标定位跟踪功能。传统无线传感器网络中对目标定位跟踪的研究往往针对红外、声波、震动等感知数据;而传统视频监控领域中对目标定位跟踪的研究又仅仅只关注视频内容处理,不涉及网络方面的问题。因此,这些传统的方法都不能直接应用于视频传感器网络之中,这就需要我们基于图像视频信息,兼顾传感器网络基本特点,对目标定位跟踪方法及关键技术进行研究。针对视频传感器网络中目标定位跟踪的典型场景,本文从网络部署与覆盖、网络拓扑控制、定位跟踪算法三个方面提出一系列新模型和新方法。这些模型与方法以数学建模与分析为基础,并在大量计算机仿真和实验场景下进行了验证。主要创新点如下:(1)面向定位的视频传感器节点感知模型。对视频传感器节点的感知区域进行了合理抽象,并基于视频传感器的投影模型和高斯误差模型分析了观测值与目标位置的几何关系,进而建立了面向定位的视频传感器节点感知模型。该模型是解决定位跟踪问题的基础。(2)面向定位跟踪的覆盖分析。传统覆盖分析方法往往针对目标探测,并不适用于目标定位跟踪。在面向定位的视频传感器节点感知模型基础上,提出了面向定位覆盖(L覆盖)概念。基于满足二维泊松过程的节点部署策略,对K覆盖率与节点密度之间的数学关系进行了推导,进而利用贝叶斯估计与蒙特卡罗仿真相结合的方法,构建节点密度、面向定位覆盖率与节点视域特征之间的关系模型。利用该模型,能够针对所需的定位覆盖需求,计算网络的部署规模。(3)面向定位跟踪的视频传感器节点选择方法。对于视频传感器网络目标定位跟踪应用而言,网络中所有节点都处于工作状态既不必要又大大增加了网络的能耗。我们将视频传感器网络中的目标定位跟踪应用划分为目标探测阶段和定位跟踪阶段,针对不同的阶段设计不同的节点选择策略。对目标探测阶段,利用面向探测和面向定位的覆盖率关系模型,推导部署节点与工作节点之间的密度关系,在此基础上根据PEAS(Probing Environment and Adaptive Sleeping)机制设计相应的网络密度控制方法,以选择合适的工作节点集维持所需覆盖率。对定位跟踪阶段,利用目标位置后验概率密度函数的信息熵作为衡量定位精度的指标,构建节点集的定位收益函数;利用节点能耗模型构建节点集的消耗函数。通过收益函数和消耗函数的建立,将节点集选择问题转化为一个最优化问题,并提出了相应的最优节点选择算法,实现在满足定位精度需求的条件下,定位能耗最少。通过上述两种节点选择方法,实现定位跟踪质量和视频传感器网络能量消耗的平衡。(4)多节点协同定位跟踪算法及实现机制。首先根据目标运动模型,建立目标状态方程;根据视频传感器节点的面向定位感知模型,建立目标状态与节点测量值之间的关系方程。基于上述两个方程,利用序列蒙特卡罗方法估计目标位置。为了使得该算法在视频传感器网络中能够高效实现,我们进一步提出了一个基于动态簇的实现机制。该机制包括:a)簇头节点的选择;b)簇内协作成员节点的选择。在簇头节点的选择中,推导出下一时间点目标位置的概率密度函数,根据该概率密度函数和周围节点的位置,推选下一个簇头节点。在簇内成员节点的选择中,结合目标探测概率与期望的后验概率密度函数建立选择标准。被选中的成员节点,将观测值发送至簇头节点,更新目标状态的后验概率密度函数,得到目标的估计位置。更多还原

【Abstract】 The target localization and tracking is a fundamental technique for many applications in wireless sensor networks, thus it is worthy to be studied. In traditional wireless sensor networks, target localization and tracking are well investigated, and there are many existing works. On the other hand, a new type of wireless sensor network, video sensor network, becomes a research focus recently. Because of the rich sensing information, video sensor networks have more advantages to implement the target localization and tracking than the traditional wireless sensor networks.In traditional wireless sensor networks, the researches on target localization and tracking are always based on the infrared ray, sound, or quiver. On the other hand, in the area of traditional video surveillance, the researches on target localization and tracking only focus on the video content processing without considering the aspect of networking. These traditional methods about target localization and tracking cannot be applied for video sensor networks directly. Therefore, it is necessary to study the target localization and tracking problem on the basis of the video information and the features of sensor networks.This thesis aims at the typical scene of target localization and tracking in video sensor networks, and proposes a serial of novel models and methods from three different aspects: networks deployment and coverage, network topology control, and the localization and tracking algorithms. Our proposed models and methods are based on the mathematic analysis, and verified by extensive simulations and experiments. The main contributions of this thesis are as follows:(1) Localization-oriented sensing model for video sensor nodes. We reasonably abstract the sensing area of a video sensor node. Furthermore, by using the perspective projection model and the Gaussian noise model, we build the geometrical relationship between the measurement and the target location. Then, we design a localization-oriented sensing model for video sensor nodes which is the basic model of the target localization and tracking problem.(2) Localization-oriented coverage. The traditional methods about the coverage always aim at target detection, and then they are inapplicable for target localization. Based on the localization-oriented sensing model, we propose a novel concept, localization-oriented coverage. Moreover, by using the deployment strategy which follows the 2-dimensional Poisson point process, we derive the mathematic relationship between K coverage probability and the nodes density. Then, we utilize the Bayesian estimation and Monte Carlo simulations to build the relationship model among the nodes density, the localization-oriented coverage probability, and the FoV (field of view) of video sensors. According to this relationship model, we can estimate the size of deployed video sensor network for a given localization-oriented coverage probability.(3) Nodes selection for localization and tracking. For the localization and tracking applications of video sensor networks, putting all the video sensors in the working mode is not necessary and increases the cost of the network. Then, we study how to use a nodes selection scheme to balance the localization quality and the cost of network energy. We divide the target localization process into two phases: detecting phase and locating phase, and design two nodes selection schemes for the two phase, respectively. For the detecting phase, we utilize the detecting-oriented coverage probability and the localization-oriented coverage probability to derive the relationship between the density of the deployed nodes and the density of the working nodes. Based on the relationship about densities, we design a nodes density control method according to PEAS (Probing Environment and Adaptive Sleeping) mechanism. Our nodes density control method can be implied for selecting the set of working node to maintain required coverage probability. For the locating phase, we utilize the information entropy of the probability density function for the target location to set up the utility function. Meanwhile, we use the energy cost model of video sensor nodes to define the cost function. According to the utility and cost functions, we map the nodes selection problem to an optimization problem, and then develop an optimal selection algorithm to minimize the energy cost while satisfying the accuracy requirement of localization.(4) The multi-node collaborative tracking algorithm and the corresponding implementation mechanism. We build up the target state equation according to the target motion model, and build up the measure equation according to the localization-oriented sensing model. Based on these equations, we apply sequential Monte Carlo method to estimate the target location during the tracking process. In order to implement the tracking algorithm efficiently, we further propose a dynamic cluster based mechanism. Our mechanism include: a) election of cluster head, and b) selection of cluster members. We derive the probability density function of target location, and then elect the next cluster head according to the probability density function and the neighboring nodes’ locations. In order to select the cluster members, we build up a selection standard by combining the target detecting probability and the posterior probability density function. The selected video sensors send their measurements to the cluster head, and update the posterior probability density function of target state for estimating the target location.更多还原