视觉基础矩阵估计方法的性能比较与分析  

Performance Comparison and Analysis of Fundamental Matrix Estimating Methods for Computer Vision Applications

作　　者：蔡涛[1] 段善旭[1] 李德华[2] 

机构地区：[1]华中科技大学电气与电子工程学院,武汉430074 [2]华中科技大学图像识别与人工智能研究所图像信息处理与智能控制教育部重点实验室,武汉430074

出　　处：《计算机科学》2009年第12期243-247,289,共6页Computer Science

基　　金：国家自然科学基金(69775022);863计划(863-306-ZT04-06-3)资助

摘　　要：基础矩阵描述了单个场景的2幅视图之间的对应关系,在计算机视觉领域中扮演着极其重要的角色。首先讨论了三维视觉系统中的极线几何理论,随后论述了几类基础矩阵的估计方法,并对这些方法作了详细的比较和评价,最后实现了各种算法且使用仿真数据以及真实图像数据对各自的性能作了分析和总结。比较结果说明:1)如果图像特征点定位精确并且匹配无误,则线性方法的结果相当好;2)迭代算法可以解决定位的高斯噪声,但是当数据存在错误匹配时,性能很差;3)鲁棒算法能够同时解决定位误差和误匹配两类问题。此外,模拟实验和真实图像实验的结果表明,基于特征分析的正交回归最小二乘法的计算结果优于经典的线性最小二乘法。The fundamental matrix （F matrix） relates corresponding points across two different viewpoints and defines the basic relationship between any two images of the same scene. Therefore, the F matrix plays an important role in most computer vision applications. Some important computing methods for the F matrix were introduced and analyzed after describing the epipolar geometry in computer vision. At last, these methods were implemented and their performances were evaluated systematically based on simulated data and practical images. The test results proved that 1） the linear methods will work well on precisely located point-pairs without no mismatch; 2）the iterative nonlinear methods can conquer the Gaussian noise in positions of point-pairs, however, have poor performance for mismatched points;3）the robust methods can resolve the problems brought by noise and mismatching. Furthermore, the results also showed that the eigen-analysis based orthogonal regression methods outperform the conventional least squares methods.

关 键 词：计算机视觉 极线几何 基础矩阵 鲁棒估计 图像匹配 

分 类 号：TP242.62[自动化与计算机技术—检测技术与自动化装置]