三维重建中立体匹配问题研究

【作者】 陈海波；

【导师】 罗三定；

【作者基本信息】 中南大学 ， 计算机应用技术， 2005， 硕士

【摘要】 立体匹配问题一直是立体视觉研究的瓶颈问题。基于区域的立体匹配算法在计算一点的深度信息时,需要从邻域获得支持。因此,问题的关键就是确定具有较强适应性的支持窗口的形状和大小。本文从分析立体匹配的一般计算原理入手,分析每个过程对于立体匹配计算准确度和时间复杂度的影响。在此基础上提出了一种新的基于区域增长的动态窗口增长策略,该动态增长策略可以很好的获取不同大小,不同形状的支持窗口。同时给出了一个具有广泛适应性的匹配测度函数,该函数理论上可以很好的避免图像中的线性和非线性误差。 在立体匹配过程中,边界问题和遮挡问题一直影响着立体匹配的准确程度。本文采用视差函数值和视差唯一性约束策略相结合的方法来解决这一问题。同时针对立体匹配算法时间复杂度高的缺点,给出了点采样、核心窗口视差近似、动态伸缩和边界约束策略,有效地减低了算法的时间复杂度。为了对实验的结果进行准确度的检验,本文还引入整体准确率和边缘准确率的测度函数,并且根据该测度对实验的结果进行了验证。结果证明,本文的算法具有良好的视差计算准确性和较低的计算时间复杂度。最后,给出了一个立体匹配算法的实验测试平台。更多还原

【Abstract】 Stereo matching is the key problem in Stereo Vision. Area based stereo matching needs to get information from the neighbor region when computing the depth of a point. How to select size and shape of supported window with strong adaptive activity is the key problem in area based stereo matching. This thesis firstly analyzes the every computing stage effect on accuracy and time cost for computing disparity image. Based on this analysis, a new stereo matching algorithm of adaptive windows via region growing is proposed. This region growing method can obtain appropriate size and shape of supported windows adaptively. In addition to this, a new cost matching function which have good adaptability is proposed. This function can eliminate the linear errors and the non-linear errors in theory.In stereo matching, the edge problem and the occluded areas problem always affect the accuracy of the disparity image. Here, the value of DSI function and the uniqueness assumption are mixed together to solve it. Aimed at reducing time cost, estimating the scope of disparity, approximating the disparity of the core window, adjusting the step dynamically, restricting edge are used to improve the algorithm. The result proves that all of them can decrease the time cost effectively. In order to verify the accuracy of the result, tow standard functions are introduced into this paper. The experiment result proves that this strategy can get good accuracy disparity image. Finally, a test bed is given to evaluate the algorithms that are appeared in this thesis.更多还原