基于SIFT的宽基线立体影像密集匹配

Dense Matching for Wide Base-line Stereo Images Based on SIFT

提出基于对极几何和单应映射双重约束及SIFT特征的宽基线立体影像多阶段准密集匹配算法。算法包括三个阶段:①基于特征点的空间分布和信息熵选取一定数量的最优SIFT特征点集并进行最小二乘初始稀疏匹配及立体像对的基本矩阵和单应矩阵估计;②对于其余特征,利用同名核线倾斜角及SIFT特征的尺度信息对匹配窗口的仿射变换参数进行迭代优化及变形改正、提取仿射不变SIFT特征描述符,并基于双重约束信息及欧氏距离测度进行匹配;③考虑宽基线立体影像较低的特征提取重复率,对第②步左右影像中未能成功匹配的特征点,基于双向搜索策略,采用基于盒滤波加速计算的SSD测度在变形改正后的双重约束区域中进行匹配,并对匹配结果进行加权最小二乘拟合定位。实际的宽基线立体影像试验结果证明了算法的有效性,可为后续的三维重建提供较为可靠的密集或准密集匹配点。

A novel multi-stage quasi-dense matching algorithm for wide base-line stereo images is introduced based on SIFT and dual constraints of epipolar geometry and homographic mapping. The proposed algorithm includes following three stages （1）The optimal SIFT features with good spatial distribution and large information content are first selected, and matched by using the least squares matching method, then the fundamental and homographic matrix can be estimated by using these initial sparse correspondences with higher precision; （2） For the other SIFT features, the affine transformation parameters between matching windows are iteratively optimized by using the slope angle of correspondent epipolar lines and scale information of SIFT features, and affine invariant feature descriptors are extracted from the corrected matching windows, then correspondences can be determined by Euclid distance and dual constraint information; （3） Considering the lower repeatability rate of feature detection for wide base-line stereo images, for the unmatched points extracted from left and right images of stereo pairs, matching can be carried out by adopting two-way search strategy from left to right image or from right to left image based on the rapid SSD similarity cost function and affine rectified dual constraints region, and the least squares curve surface fitting weighted by Gaussian-distance algorithm is adopted to improve the precision of matching results. Test results using practical wide base-line image pairs indicate the proposed algorithm is effective and can provide reliable dense or quasi-dense matching points for subsequent 3D reconstruction.