三维物体抗仿射变换特征匹配方法

为解决三维物体视角变化下的图像匹配问题，提出一种实现三维物体任意仿射变换间的图像匹配方法. 在抗仿射变换的Affine-SIFT算法基础上，提出了多变换矩阵mASIFT（multi Affine SIFT）算法. 使用多平面随机抽样一致性（multi-plane RANSAC，mRANSAC）几何变换约束方法，从描述子粗匹配结果中提取正确匹配对. 比单一变换矩阵方法更加符合目标物体的立体多平面特性，匹配对数普遍是ASIFT的5~9倍. 降低每个单应矩阵的误差阈值，使内点提取中双向变换误差阈值小于3个像素（或2个像素），更有效地剔除误匹配，可计算出更精确的变换矩阵，实现更加准确的匹配. 

Image Matching Method of Three-Dimensional Object with Affine Invariant Feature

In order to realize the image matching of three-dimensional object under different perspective, a method which is robust to affine transformation is proposed. Based on the fully affine invariant image comparison method, named Affine-SIFT (ASIFT), mASIFT (multi ASIFT) was put forward, in which a geometric transformation constraint method named mRANSAC(multi-plane RANSAC) was used to extract the right matching points from the coarse matches. It is corroborated that, compared to the singular homography constraint, the multi-plane constraint excluding mismatching method is more proper to the multi-plane characteristics of 3D objects. It could increase the number of matching pairs up to 5~9 times of ASIFT under the same condition. Further, the mutual error threshold of every homography would lower 3 pixels or 2 pixels. As a result, the proposed method could remove mismatches more efficiently, calculate out more accurate transformations and obtain more exact matching.