摘要
针对宽基线影像的高畸变特点,提出基于几何一致性约束的Harris-Affine高精度最小二乘匹配(Least Square Matching,LSM)方法。首先,用Harris-Affine算子在影像高斯多尺度空间中提取仿射不变特征,并根据特征区域灰度矩阵与主梯度方位来实现特征的几何归一化,继而提取特征描述符,接着采用一种由粗到精的特征匹配策略来渐进地获取正确率占优的特征匹配。然后,通过LSM迭代方法来补偿同名像点的定位误差,LSM迭代所需的良好的几何畸变初值由同名特征区域灰度矩阵与主梯度方位求取,而良好的辐射畸变初值通过同名区域像元灰度的最小二乘线性拟合法求取。实验结果表明:该算法鲁棒有效,且实现了亚像素匹配精度。
Due to the complex geometric and radiometric distortions on widely separated images, the traditional stereo algorithms were not suited for them, thus a novel high-accuracy least square matching（LSM） method based on Harris-Affine features and geometric consensus constraint was proposed. Firstly, Harris-Affine features were detected in multiple scale space of Gaussian images. Local features were normalized separately based on intensity moment and the principle gradient orientation of feature region, and then these normalized results were described respectively. Subsequently, a novel feature matching strategy from coarse to fine was implemented on candidates in order to improve the correctness of correspondences. Secondly, LSM algorithm was performed on these conjugate features so as to obtain the fine matches. As the LSM must require good initial value for iteration convergence, the geometrical distortion initial value was calculated based on intensity moment and the principle gradient orientation of correspondent affine regions, and the radiometric distortions initial value was obtained through least square fitting method which was implemented on normalized correspondent regions. Experiments on wide baseline images of real-world scenes indicate that the proposed algorithm has significant superiority of matching precision.
出处
《中南大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2014年第8期2661-2668,共8页
Journal of Central South University:Science and Technology
基金
国家自然科学基金资助项目(41272389,41371438)
江苏省基础研究计划(自然科学基金)青年基金资助项目(BK20130174)
江苏省普通高校研究生科研创新计划项目(CXZZ13_0937)
江苏高校优势学科建设工程项目(SZBF2011-6-B35)
