融合互补仿射不变特征的倾斜立体影像高精度自动配准方法

An Automated Registration Method with High Accuracy for Oblique Stereo Images Based on Complementary Affine Invariant Features

提出一种基于最大稳定极值区域(maximally stable extremal regions,MSER)与Harris&Hessian Affine的互补仿射不变特征高精度自动配准算法。算法分3个阶段:①融合MSER与Harris&Hessian Affine互补不变特征,采用最小生成树算法选取一定数量的最优互补特征集合,基于特征的仿射不变信息实现局部图像的仿射与方向归一化,特征匹配采用多层次自适应策略,首先基于SIFT描述符的欧氏距离比率测度获得初始匹配,继而估计影像间的基本矩阵与单应矩阵,然后在双重几何约束下利用归一化互相关(normalized cross correlation,NCC)测度进行扩展匹配,以增加特征匹配数量且最大限度地消除误匹配;②通过最小二乘匹配(least square matching,LSM)使匹配结果达到子像素精度,以提高配准精度,最小二乘匹配的迭代参数初值由同名仿射不变特征间的协方差矩阵与主梯度方位获得;③基于②的匹配结果和投影变换模型,完成影像的高精度配准。针对地面近景倾斜立体影像和无人机倾斜立体影像的试验结果证明了算法的有效性。

An automatic and precise registration algorithm is proposed based on complementary affine invariant features. The algorithm was involved in three stages： ① MSER and Harris＆ Hessian Affine were detected simultaneously on image pairs, and then a certain amount of optimal features were extracted by MST algorithm. Local image patch centered on centroid of affine regions were rectified based on the information of affine geometry and orientations. In order to increase the number of matches and identify the false correspondences, matching method was implemented from coarse to fine. Initial matching features were firstly obtained by the Euclidean distance ratio and SIFT descriptor, and then the fundamental and homography matrixes can be estimated between image pairs, the more matching features were acquired by using hybrid constraints based on normalized cross correlation measure. ② with the purpose of gaining subpixel accuracy, initial projection matrix was optimized by least square matching, and then the matching error was compensated according to the optimal projection matrix, the initial projection matrix was obtained based on covariance matrix and the principal gradient orientation of correspondent affineregions. ③ the high accuracy registration was obtained based on the matching result of the second stage and the projection transforma- tion model. Experiments on oblique images taken by ground close-range and unmanned aerial vehicle （UAV） demonstrate the effectiveness of the proposed algorithm.