【作者】 李霞；

【导师】 李定主； 熊风光；

【作者基本信息】 中北大学 ， 计算机科学与技术， 2019， 硕士

【摘要】 在计算机视觉领域,曲面匹配是一项非常重要的基础研究,在三维识别、逆向工程等领域都有着非常广泛的应用。本文的主要研究方向是基于特征的曲面匹配算法,其核心处理过程可以分为三个步骤:特征点的检测、特征点的描述和特征匹配。针对这三个步骤,本文提出了一种基于协方差矩阵的多尺度特征描述子,用于特征点的描述;并将相关性降采样方法引入到所提出的特征描述子中,进行特征点的检测;同时,还提出了一种基于演化博弈论的误匹配剔除算法用于提升特征点匹配的效果。具体研究内容如下:(1)提出了一种结合三维几何信息和视觉信息的基于协方差矩阵的多尺度特征描述子。该描述子通过特征点及其邻域的颜色和稳健的几何特征来构建多尺度的协方差矩阵,以实现对特征点的描述,能够解决目前特征描述子因只注重特征点几何形状信息描述而引发的抗噪性较差等问题。实验结果表明:在匹配复杂度较高的三维模型时,本文特征描述子与其它特征描述子相比,具有更好的抗噪性,对分辨率的变化具有更强的鲁棒性。(2)提出了一种根据相关性降采样方法的特征点检测算法。该算法通过计算候选特征点的关联度划分相关性区域,并将该区域内显著性的值为极大值的点作为整个区域中的特征点,实现对候选特征点的降采样,能够解决目前检测算法结果中的特征点集群现象,进而导致算法运行效率不高的问题。实验结果表明:本文特征点检测算法能够在不影响匹配准确度的基础上,将用于之后特征匹配的点的数量减少了70%以上,提高了算法的运行效率。(3)提出了一种基于演化博弈论的误匹配剔除算法。首先,该算法通过计算任意两对初始对应关系之间的特征描述子相似性,距离和角度一致性来构建收益矩阵,并计算出整体的平均收益。然后,根据感染与免疫动力学方程动态模拟演化过程。其中,获得较大收益的初始对应关系生存下来,获得较小收益的初始对应关系则被淘汰,据此逐步迭代出最终的演化均衡状态。最后,该演化均衡中剩余的初始对应关系就是最终的匹配对,以此实现了对错误对应关系的剔除。实验结果表明:本文所提出的误匹配剔除算法与现有剔除算法相比,具有更高的运行效率和更强的抗噪性。更多还原

【Abstract】 In the field of computer vision,surface matching is a very important basic research,and widely applied in three-dimensional recognition,reverse engineering and other fields.The main research direction of this paper is surface matching algorithm based on the feature.Its core processing process include three steps: feature point detection,feature description and feature matching.In view of these three steps,a multi-scale feature descriptor based on covariance matrix is proposed to describe feature points.The correlation downsampling method is introduced into the proposed feature descriptor to detect feature points.At the same time,an error matching elimination algorithm based on evolutionary game theory is proposed to improve the performance of feature point matching.Specifically,the following aspects of work have been done.(1)A multi-scale feature descriptor based on covariance matrix combining three-dimensional geometric information and visual information is proposed.The descriptor constructs a multi-scale covariance matrix by using the color and robust geometric features of the feature points and their neighborhoods to describe the feature points.It can solve the problem of poor noise resistance caused by the current feature descriptor only focusing on the description of geometric information of the feature points.The experimental results show that when matching complex three-dimensional models,this feature descriptor is more robust to the noise and the change of resolution than other feature descriptors.(2)A feature point detection algorithm based on correlation downsampling is proposed.The correlation region is divided by calculating the correlation degree of the candidate feature points,and takes the points whose saliency value is the maximum as the feature points in the region,the de-sampling of the candidate feature points is completed,which can solve the clustering phenomenon of the current feature points,and the inefficiency of the current detected algorithm.The experimental results show that this feature detection algorithm can solve the clustering phenomenon of feature points without affecting the matching accuracy,and the number of points involved in later feature matching is reduced more than 70%,which improves the efficiency of the algorithm.(3)An error matching elimination algorithm based on evolutionary game theory is proposed.Firstly,the algorithm constructs the payoff matrix by calculating the similarity,distance and angle consistency of the feature descriptors between any two pairs of initial correspondence relationships,and calculates the overall average payoff.Then,the evolutionary process is dynamically simulated according to the infection and immune dynamics equation,the initial correspondence which obtains larger payoff survives while the initial correspondence which obtains smaller payoff is eliminated,and the final evolutionary equilibrium state is gradually iterated,and the final evolutionary equilibrium is iterated step by step.Finally,the remaining initial correspondence in the evolutionary equilibrium is the final matching pair,which achieves the elimination of error correspondence.The experimental results show that the algorithm has the advantages of high efficiency and strong anti-noise.更多还原