节点文献
基于计算机立体视觉的双目立体成像研究
Study on Binocular Stereo Imaging Based on Computer Stereo Vision
【作者】 刘然;
【导师】 朱庆生;
【作者基本信息】 重庆大学 , 计算机应用技术, 2007, 博士
【摘要】 双目立体成像(Binocular Stereo Imaging)是指根据双目立体视觉(Binocular Stereo Vision)的基本原理,获取、产生和传输一个场景,并将这个场景展现出具有立体感的景象。基于计算机立体视觉(Computer Stereo Vision)的双目立体成像不仅是实现虚拟植物可视化立体展现的关键技术,而且在虚拟现实、机器视觉、多媒体教学、数字娱乐、产品外观设计、雕刻与建筑等领域都有着广泛的应用。本文研究的目的,就是通过对基于计算机立体视觉的双目立体成像进行研究,为它的应用提供理论和技术支持。论文首先对基于计算机立体视觉的双目立体成像相关技术及理论进行了研究,然后在此基础上给出了一个基于监视器的双目立体成像模型,详细讨论了这个模型所具有的成像性质。最后针对该模型结合实验深入研究了三种立体图像对(stereo pair)获取的方法:基于Hou方法的立体图像对的生成方法、已有三维模型的情况下立体图像对的生成方法和基于单目视频图像序列的立体图像对的生成方法。论文的主要研究成果包括以下几点:针对遮挡问题,本文提出,可以利用空间关系的定性表示与推理来表示物体之间的位置关系以及让计算机理解这种位置关系。在空间关系的定性表示方面,我们给出了一个SR表示模型,并提出了距离关系的定性表示模型和方向关系的定性表示模型。在空间关系的定性推理方面,我们定义了基于SR表示模型的组合运算,给出了相应组合运算表;此外,还构造了定性三角推理所需要的QTR组合运算表、距离定性减法表和角度定性减法表。与定量的方法相比,采用定性的方法大大降低了时间复杂度和空间复杂度。在成像模型方面,本文给出了一个基于监视器的双目立体成像模型,详细研究了这个模型所具有的成像性质。我们给出了立体图像对坐标的计算公式、讨论视差及其性质、像点立体成像规律以及直线段立体成像规律,并设计实验验证了相关的结论。理论的分析和实验表明,当观察者保持双眼平行地远离或靠近监视器屏幕,他看到的各点深度都相同的图像虽然也会移动,但不会产生变形;但是,各点深度不同的图像则会产生形变,并且遵循一定的成像规律;所成的虚像在负视差的情况下是可测量的。针对Hou方法,本文给出了评价转换后的立体效果的定量指标(交叉熵和均方根误差),以及产生符合要求的随机数算法ProduceKmn,构造了平移像素法、均匀分布、三角分布、分段均匀分布、拟合灰度值分布、正态分布等六种分布,通过实验详细讨论了各种分布对立体效果的影响。最后,研究了图像子块个数对立体效果的影响。实验和分析表明,使用Hou方法时,平滑与立体感是互相矛盾的;随机变量服从正态分布时立体效果较好;相对于其他方法,Hou方法简单且效果明显,但适用范围有限。论文介绍了已有三维模型的情况下立体图像对的生成的方法,详细讨论了影响立体效果的几个重要因素,包括目标摄像机与三维模型的位置关系、基线长度、成像位置、成像变形等等,给出了相应的成像规律。此外,论文还研究了视差对深度感知的影响、对象的提取、时序与帧间隔等关键问题。在此基础上,提出了一种单目视频图像序列立体化的方法,并通过实验对该方法进行了检验。实验显示,当摄像机或场景中的物体做水平或者近似水平方向的运动时,由该方法生成的立体视频具有较好的立体效果。
【Abstract】 Binocular stereo imaging means capturing, producing and transmitting a scene, and then displaying it with stereoscopic perception based on the basic principles of binocular stereo vision. Binocular stereo imaging based on computer stereo vision is not only the key technique to realize the stereoscopic display of the virtual plant, but also widely used in many fields such as virtual reality, machine vision, multimedia education, digital entertainment, appearance design of industrial products, sculpture and architecture etc. The purpose of this dissertation is to probe into binocular stereo imaging based on computer stereo vision, which will provide the theories and techniques support for its applications.In this dissertation, the theories and techniques related to binocular stereo imaging based on computer stereo vision are firstly studied. Based on this foundational work, a binocular stereo imaging model based on monitor is proposed and the properties of this model are thoroughly discussed. Finally, with the help of the model and subsequent experiments, three kinds of methods for acquiring stereo pair for the model combined with experiments are elaborated: stereo pair creation based on Hou method, stereo pair creation when 3D models are given, and stereo pair creation based on monocular video frequency image sequence.The main research results of this dissertation are summarized as follows.With respect to occlusion, the dissertation argued that, the theory of qualitative representation and reasoning of spatial relations can be used for the representation of the positional relations among objects in computer and the comprehension of this kind of relations by computer. As for qualitative representation of spatial relations, a SR representation model is presented and further provided is the qualitative representation models of distance and directional relations. As for qualitative reasoning of spatial relations, the composition operation is defined based on SR representation model and the corresponding composition operation table is also presented. Furthermore, we create qualitative subtraction table of distance, qualitative subtraction table of angle and the QTR composition operation table which is needed for qualitative trigonometry reasoning. Compared with the quantitative method, the qualitative method can considerably reduce the time complexity and space complexity. As for imaging model, the dissertation introduces a binocular stereo imaging model based on monitor. Then it focuses on the imaging properties of the model. The formula for computing the coordinates of stereo pairs is given; the parallax and its properties are discussed; we also present the rules of stereo imaging of point and line segment, and experiments have been carried out to prove the conclusions. Theoretical analysis and the experimental results show that, the observer can see the image still move when his eyes move away from or towards the screen when the depth of each point of the image is the same, and the image won’t be distorted; on the other hand, if the depth of each point of the image is different, the image will be distorted by following certain imaging rules. The stereo images are measurable in the case of negative disparity.The dissertation presents two quantitative criterions, cross-entropy and root-mean-square error, to evaluate the stereo effect for Hou method. Besides, an algorithm named ProduceKmn, which is used to produce random numbers that accord with the requirement, is also presented. Moreover, we constructed six kinds of distributions for Hou method, such as Moving Pixels without Stretch, Uniform Distribution, Triangular distribution, Piecewise Uniform Distribution, Fitting Gray Value Distribution, and Normal Distribution. Meanwhile, on the basis of experiments, the impacts that the above-mentioned six distributions may have upon stereo effect have been discussed. In addition, the interrelation between the stereo effect and the number of sub-blocks has been also studied. The experimental results and the analysis show that, a) smoothness is incompatible with stereoscopic perception; b) Generally, the stereo effect is preferable when random variables have normal distribution; c) In contrast with other methods, Hou method is simple, and the stereoscopic perception is distinct, but its application scope is limited.The dissertation introduces a stereo pair creation method when 3D models are given. Meanwhile detailed discussion is focused on the important factors that exert great influence on stereo effect, such as positional relations between the target cameras and the 3D Model, length of the baseline, position of the virtual image and imaging distortion, etc. Moreover, corresponding rules of imaging are also presented.Last but not least, the dissertation also touches on the influence that parallax has upon depth perception, object extraction, time sequence & frame interval, etc. Accordingly, we presented a method to convert a monocular video frequency image sequence to a stereo one. Experiments have been carried out to test it. The results reveal that, better stereo effect will be perceived from the stereo video converted by this method when the camera or the objects in the scene make horizontal movements or approximately horizontal movements.
【Key words】 Stereo Vision; Computer Stereo Vision; Binocular Stereo Imaging; Stereo Pair; Parallax; Stereo Video;
- 【网络出版投稿人】 重庆大学 【网络出版年期】2007年 05期
- 【分类号】TP391.41
- 【被引频次】37
- 【下载频次】3269