摘要
提出基于对极几何和单应映射双重约束的SIFT特征多尺度加权最小二乘匹配算法。算法首先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,并采用基于奇异值分解(SVD)的SIFT特征匹配、基于SIFT特征尺度和方位信息的自适应归一化互相关(NCC)匹配获得精度较高的初始匹配点用于立体像对的基本矩阵和单应矩阵估计。然后在对极几何和单应映射的双重约束下,基于自适应NCC及距离加权的多尺度最小二乘匹配算法进行扩展匹配并同时保留匹配定位精度较高的原始SIFT特征点对。算法综合应用基于积分影像的NCC快速计算、金字塔影像匹配等方法和策略。最后选取实际的宽基线序列立体影像进行试验并同原始的SIFT特征匹配算法、基于SVD的SIFT算法进行了综合对比分析。结果表明当影像间无显著亮度变化时该方法的匹配性能明显优于现有的方法。
Some new feature extraction and matching algorithms（especially the Scale Invariant Feature Transformation algorithm,SIFT） that are invariant to translation,scale and rotation changes have been widely used in digital photogrammetry and computer vision fields.However,SIFT features are not invariant to affine deformation,and the location precision may be low because the feature extraction and matching are independent.In practical applications,this disadvantage makes it not appropriate for wide base line stereo matching.In order to solve this problem and to make it invariant to affine transformation,a new SIFT features＇ matching algorithm based on duplicate constraints and least squares image matching method was proposed in this paper.In this algorithm,the optimal SIFT features with good spatial distribution and large information content was first selected,then these SIFT features was matched by using SVD-SIFT algorithm and adaptive NCC matching method based on scale and orientation information of SIFT features.The fundamental and homography matrix can be estimated by using these initial correspondences.Other SIFT features were matched by using duplicate geometric constraint and weighted least squares matching with multiple scale template window.At the same time,the least squares matching results were selected accordingly.Further,compared to the location error of the original SIFT key points,least squares matching results were determined to adopted or not.Last,two wide base line image sequences were selected to test the performance of proposed algorithm.Test results indicate that the proposed method has higher performance than the original SIFT methods and the SVD based SIFT methods when no significant illumination changes between stereo image pairs.
出处
《测绘学报》
EI
CSCD
北大核心
2010年第2期187-194,共8页
Acta Geodaetica et Cartographica Sinica
基金
地理空间信息工程国家测绘局重点实验室开放基金(200818)
关键词
尺度不变特征变换
特征提取
归一化互相关系数
单应映射
scale invariant feature transformation
feature extraction
normalized cross correlation
homography mapping
作者简介
杨化超（1977-），男，副教授，博士生，主要研究方向为摄影测量与遥感。
