【作者】 项学智；

【导师】 赵春晖；

【作者基本信息】 哈尔滨工程大学 ， 信号与信息处理， 2008， 博士

【摘要】 利用时变图像序列进行运动分析来确定三维空间中运动物体的结构或物体与观察者之间的相对运动参数是计算机视觉领域中的热门研究课题。三维运动投影至二维图像平面也将形成运动,这种运动以图像平面上亮度模式的流动表现出来,称之为光流,而光流场是一种二维速度场。对光流场的计算是低层视觉计算技术中的一项重要任务。对光流计算技术的研究已持续了将近三十年,其间产生了一大批富有成效的光流计算方法,微分光流算法、匹配光流算法、能量光流算法、相位光流算法、小波光流算法等相继被提出,其中微分光流算法因具有完备的数学理论基础,实现简单,精度相对较高等优点而获得了广泛的应用。然而,这些光流计算方法都是基于图像中的灰度信息进行计算的,忽略了图像中蕴含丰富的彩色信息。光流估计是一个病态问题,彩色信息提供的附加约束可有效克服光流孔径问题带来的影响。彩色光流估计方法大致可分为灰度一致性方法和色彩一致性方法两大类,实验表明其光流估计精度优于经典的灰度估计方法。然而,总的来说,彩色图像序列的光流场计算技术仍处于起步阶段,利用彩色信息改进现有光流计算方法以获得更好的性能是一个值得进一步研究的课题。本文在回顾经典灰度与彩色光流计算方法的基础上,着重于研究利用彩色信息提高传统灰度光流估计算法的性能,同时对利用光流场进行运动目标检测也进行了探讨,其主要工作可概括为以下几个方面：第一,在基于高阶梯度的光流计算方法基础上,给出了一种基于色彩梯度恒常性的微分光流计算方法,该方法也可看成是一种基于二阶梯度进行光流求解的方法,只是其一阶梯度使用了基于色彩信息的彩色向量梯度。在假设色彩梯度恒常的基础上列写梯度约束方程,并施加全局平滑约束以求解光流。最后,进行了数值实验,实验结果表明了算法的有效性。第二,在研究了经典彩色光流算法及基于全局平滑约束的光流算法各自特点的基础上,给出了一种融合彩色光流计算与全局平滑灰度光流计算的方法。该方法通过矩阵条件数来判断彩色光流解的可靠性,对于不可靠光流位置,使用全局平滑光流解来进行光流替换,从而得到一种混合的光流解。同时,在上述算法的基础上,本文进一步研究了彩色局部光流计算模型和彩色全局光流计算模型。局部模型利用了邻域约束,可提高光流估计精度并有一定的抗噪性能,全局模型可得到稠密光流场但会模糊物体运动边界,参考前述的混合光流计算方法,本文提出了三种针对彩色图像序列的混合光流计算模型,并进行了数值实验,实验结果表明了算法的有效性。第三,回顾了小波的基本原理及其在光流估计中存在的问题,并简要介绍了利用复值小波克服相位震荡进行光流估计的方法,同时给出了一种利用色彩信息提高复值小波光流估计精度的方法,利用多通道彩色信息对基本的复小波光流方程组进行了扩展,利用光流解的稳定性判据筛选最可靠颜色通道进行光流计算,最后进行了实验对比,实验结果表明了算法的有效性。第四,给出了一种基于光流场与水平集曲线演化检测运动目标的方法。该方法首先利用运动内极线约束或规范化光流场确定运动目标数量及大致位置,然后通过Kmeans动态聚类方法进行运动目标区域分割。然而,由于光流估计误差及聚类分割误差,单纯依靠运动信息的检测方法通常都不能得到准确的目标轮廓线,因此本文在运动分割的基础上进一步加入基于空域信息的水平集分割步骤,利用彩色向量梯度定义了曲线演化停止函数,并利用快速行进算法加快计算速度。最后,对该算法进行了单目标与多目标,静止背景与动态背景的三组实验,实验结果表明了算法的有效性。更多还原

【Abstract】 It is a hot topic in computer vision field that determining the structure of moving object in 3D space or the relative motion parameters between the viewer and the object using image sequence. And it will be 2D motion when 3D motion is projected onto the 2D image surface. This motion will appear as the flow state of 2D brightness pattern, which is called optical flow. The optical flow field is a kind of 2D flow velocity fields. The computation of optical flow field is an important task of low level vision computation technology.The researching of optical flow computation technology has lasted for thirty years. And many effective computation methods have been produced. Many algorithms such as differential optical flow algorithm, matching optical flow algorithm, energy-based optical flow algorithm, phase-based optical flow algorithm and wavelet optical flow algorithm are proposed and developed one after the other. Among all of these algorithms, differential optical flow method is used broadly because it has the complete mathematics theories, can be realized easily and has high precision of estimation. But these algorithms metioned above are all based on the gray information and the color information is ignored dure the procedure of computation. Optical flow estimation is an ill-posed problem. The additional constraint of color information can be used to solve caperture problem. Color optical flow estimation method can be classified as gray consistency method and color consistency method. The experiments show that the precision of optical flow estimation is better than gray method. On all accounts, optical flow estimation based on color information is still immature. And it is a valuable topic that using the color information to improve the performance of existing optical flow estimation method. In this thesis, we reviewed the classic gray and color optical flow method, and focused on improve the performance of traditional gray optical flow methods using color information. At the same time, we also discussed the method of moving object detection using optical flow.The innovations of this thesis are as the follows:1. Based on the high order optical flow method, a new optical flow estimation method based on color gradient consistency is proposed. This method can be regarded as a second order optical flow method. And its first order gradient is just employed color vector gradient. The gradient constraint equation can be produced based on assumption of color gradient consistency and global smoothness constraint is used to solve optical flow. Finally, the numerical experiments were performed. The experiments show that the method is effective.2. Based on the characteristic of classic color optical flow method and global smoothness optical flow method, a new method fused color and gray optical flow method is proposed. The reliability of color optical flow is judged by condition number of matrix. The method adopts the smooth optical flow value to replace the unreliable color optical flow so that we can get the fixed optical flow field.At the same time, the color local model and color global model have been discussed in this thesis. The local model uses local constraint to improve the precision of estimation and has robustness to against the noise. The global model can produce density optical flow field and blur the edge of moving object at one time. Refereced the fuse method discussed above, three fuse models of color optical flow computation are proposed. And the experiments were exployed to contrast. The experiments show that the method is effective.3. The thesis reviewed the basic theory of wavelet and the problem in optical flow computation. Then complex wavelet is discussed and it can be used to overcome the effect of phase concussion. A method which used color information to improve the performance of complex wavelet is proposed. Multi-channel color information is used in this method to extend the basic complex wavelet optical flow equation and condition number is used to judge the reliability of every channel. The best stablest channel is selected to compute optical flow using complex wavelet. Finally, the numerical experiments were performed. The experiments show that the method is effective.4. A method of moving object detection based on optical flow field and level set is proposed. Epipolar constraint or standardization optcal flow field is used to determeming the number and initial motion region. Then Kmeans algorithm is used to get the region of moving object. However, the algorithm can not get the accurate edge of moving object because the error of segmentation and computation of optical flow. So level set method based on spatial information is adopted to get the final segmentation result. Color vector gradient is used to define the evolution ending function and fast marching method is used to improve the performance of level set method. Finally, the experiments which include single object, multi-object, state background and moving background were performed. The experiments show that the method is effective.更多还原