【作者】 杜海；

【导师】 李木国；

【作者基本信息】 大连理工大学 ， 信号与信息处理， 2009， 博士

【摘要】 流场信息的测量是流体研究中的一项重要内容。相对于传统接触式单点测量方式,粒子测速技术可以非接触地获得瞬间、全场的流动信息。该技术对于流动结构研究极为有益,现已成为现代流场测试技术之一并且得到了广泛的应用。根据空间粒子图像的特点,本文基于双目视觉原理对粒子图像测速仪(ParticleImaRe Velocimetry,PIV)的匹配技术进行了研究。在对每个相机的粒子图像二维匹配技术研究的过程中,着重根据研究对象——流体以及置入观测场中的示踪粒子的特点,对粒子运动图像以及相机像平面上粒子数据的匹配结构进行了研究,其中包括:(1)根据粒子分布不均匀且分布结构较复杂以及不同空间、时间的复杂流场会出现不同区域特征的特点,对原PIV技术中的定窗分析技术进行改进,提出了一种改进的处理方法:包括利用加权平均的方法对后续图像的预测处理,以及根据窗口内数据的相关程度与粒子分布浓度来对分析窗进行迭代选择。(2)流体具有连续性,散布在流场中较密集的粒子群中邻近粒子的运动具有很强的相关性。根据这种相关性,提出了基于自组织映射(Self-organized Mapping,SOM)神经网络的粒子图像测速算法。经SOM网络改进的测速算法首先利用相关后的结果进行网络构建,然后使用逐次相关的方法对候选匹配点进行筛选。该算法不仅消除了粒子密度与灰度分布的敏感性,而且也降低了相关时对分析窗口尺寸的敏感。(3)根据粒子图像中粒子聚团分布的特点,针对粒子追踪技术中的粒子误对应问题与粒子相关法对粒子区域平均化处理所带来的误差问题,制定了用于粒子图像测速的细胞分裂准则,并提出了基于细胞分裂的粒子追踪匹配算法模型:将兴趣区域作为细胞体,并以细胞体内单点匹配相关程度与粒子的近邻程度作为判断准则对分析窗口内的点进行分裂;分裂形成的子团之间进行竞争,并将优胜团代表本分析区域参与其他区域优胜团的竞争;根据优胜团先后位置变化得到矢量位移场。在对粒子图像三维匹配技术分析的过程中,着重根据机器视觉测量原理对两个相机的空间匹配方法以及像平面上有效数据的提取方法进行了研究,其中包括:(1)针对运动粒子在具有一定体积的流场中透视成像后易出现的运动重叠问题,研究了PIV技术中近邻粒子的弱刚性运动,并在PIV技术中近邻性假设的基础上,推导出用于粒子筛选的空间粒子运动测量约束条件以及实际中便于应用的三点式结构约束条件。该约束条件基于PIV的近邻性准则并以邻近的粒子团作为匹配对象,利用视平面内相同运动的特征直线具有公共点的性质对前后两帧图像中的粒子进行运动检测。(2)针对粒子测速系统中空间点立体匹配难、空间信息难以表示的问题,构建了基于双目视觉的三维粒子测速一般模型。模型中提出并采用了分层提取的逐层相关法,即先对三维空间进行二维切分,然后结合PIV技术与粒子追踪测速(Particle TrackingVelocimetry,PTV)技术对不同帧的二维粒子层面进行相关处理并求出粒子运动矢量。(3)提出了基于遗传算法的三维粒子图像匹配方法及实际中的优化设计。该方法以视差为编码,结合SSD(Sum of Square Difference)与SAD(Sum of Absolute Difference)法对结果进行评估,并通过单点、多点的双交叉、变异与序列的混沌化处理达到对分析空间的搜索;最后,局部处理结合全局唯一性迭代检测进一步增加了匹配结果的可信度。(4)提出了可适用于体积光照明的双目PIV系统三种匹配结构并对其进行分析,同时根据匹配结构的特点提出了双相机像面上数据的匹配融合方法。此外,基于上述匹配方法,构建了体视双目PIV系统,同时对原二维PIV系统中伪矢量的后处理方法进行改进,使之可以适用于体视PIV匹配结构。本文在研究这些匹配技术的同时做了大量的实验,实验结果表明所提匹配算法切实可行具有较强的适用性。与此同时,使用本文所构建的体视双目PIV系统进行了实际流场分析,实验表明该PIV系统可对流场进行较好的观测。更多还原

【Abstract】 Measurement of fluid information is very important in researches of fluid characters.In contrast with the touching single point measurement methods,techniques of particle image velocimetry(PIV) can be used to obtain instantaneous information over the entire field of view,while leaving the fluid flow undisturbed.It has thus gained a reputation as one of the most promising measurement techniques,and has been used in a wide range of research.According to the characters of particle images,matching methods of PIV systems have been studied based on binocular vision theory.When investigating the matching methods in image plane,according to the characters of fluid and tracing particles,the particle images and the matching structures of particles in the CCD(Charge Coupled Device) camera plane have been analyzed.Firstly,distributions of tracing particles are not well-proportioned in the fluid field and the complex fluid field may have different region features.In order to improve the analysis capabilities,the fixed interrogation window technique of PIV is modified.The current results are averaged as estimations of next process.The size of the interrogation window is selected iteratively according to the correlation degree and the distributions of particles.Secondly,fluid is continuous and there exists high correlation among adjacent particles. A modified PIV method based on Kohonen self-organized mapping(SOM) neural network is presented.First,the results of cross-correlation are used to build networks.Second,tracking method is used to select matching points.The new PIV algorithm based on SOM network can reduce the dependence on particle density,intensity distribution and interrogation window size.Thirdly,false matching in particle tracking velocimetry(PTV) and low-pass feature of PIV will increase processing errors.To overcome the two problems,cell segmentation theory (CST) is presented in this paper according to the clustering characters of particle images. Furthermore,the processing model based on CST is described as:First,the interrogation fields are divided into different local spaces named as cells,and these cells continue to segment into sub-cells according to the correlation degree and neighbor degree.Second,these sub-cells compete against each other and the final victorious one attends competition in other fields.Third,the velocity vector field is gotten according to the position alteration of the victorious cell. While studying 3D matching methods in space,according to vision theory,the extracting methods of valid data in images and the matching methods have been analyzed.Firstly,the rigid motion hypothesis in PIV is studied because of the overlap problem of perspective particle images.The movable particle measurement constraints(MPMC) and the structure constraint based on three points are presented.MPMC is based on the nearest assumption and the adjacent particles with certain numbers are regarded as objects to match. The motion feature points and the motion feature lines are defined in MPMC.The motion feature lines with the same movement have the same feature point,which can be used to detect particle movements.Secondly,a new 3D measurement method based on binocular vision is proposed.Using the proposed method,problems of spatial stereo matching and information representation can be solved.A 3D space is first segmented and aligned to many 2D planes.And then 2D-PIV and PTV technologies are applied to different image frames which include the information of vision disparity.Thirdly,the 3D particle image matching method and optimal design are proposed based on genetic algorithm(GA).Disparity is aligned to 1-D data array and encoded.And then the sum of square difference(SSD) method or the sum of absolute difference(SAD) method is applied to evaluate the results.A modified genetic algorithm is employed to stereo matching. First,the crossover and mutation methods are modified;second,the sequences are made chaotic;third,the uniqueness is detected based on iterative algorithm.Fourthly,three matching structures,which are suitable for binocular PIV system with volume light,are presented and analyzed.Meanwhile,according to the characteristics of matching structures,a new matching method of data in two CCD images is proposed.Furthermore,based on the aforementioned matching techniques,the binocular PIV system is constructed.The post-processing method for 2D-PIV is modified in order to fit for the proposed 3D-PIV matching structure.In this paper,lots of particle images are tested and the errors are analyzed.These experiments demonstrate the effectiveness and practicability of the proposed methods and 3D-PIV system.更多还原