结合网状描述符和单应约束的近景影像直线匹配

Close-Range Image Line Matching based on Mesh Descriptor and Homography Constraint

针对现有立体影像直线匹配方法中的线描述子只依赖局部灰度特征导致可靠性较弱的问题,本文提出了一种结合网状描述符和单应约束的直线匹配方法。(1)利用线特征检测算法(Line Segment Detector,LSD)提取参考影像及搜索影像中的直线段;(2)根据角度约束和核线约束确定候选直线,缩小直线搜索范围,并计算参考直线与候选直线的重叠部分以确保端点一致;(3)利用直线段固定邻域内的同名点对构建网状描述符,选择3组不同的同名点对分别计算直线相似值,取其中最大值作为直线段的最终相似度从而确定同名直线对;(4)将未搜索到同名点对的直线段利用单应性矩阵映射至搜索影像,并根据3个判别准则得到最终的匹配结果。为了验证算法的有效性及鲁棒性,本文选取国际公开的标准测试数据集中5组近景影像进行实验,并与现有3种具有代表性的算法进行对比。结果显示本文算法的准确度及有效性均优于对比的3种算法,在匹配准确率与运行效率上最高有16.1%与49倍的提升,对于不同条件下的影像均能取得良好的直线匹配结果。

Aiming at the problem that the line descriptor in the existing stereo image line matching methods only depends on local gray features,resulting in weak reliability,a line matching method based on mesh descriptor and homography constraint is proposed in this study.Firstly,we extract line segments and the corresponding point pairs in reference image and search image using the Line Segment Detector(LSD) and Scale-Invariant Feature Transform( SIFT) algorithms.And line segments that do not satisfy the length requirements to eliminate the noise caused by shadows or other ground objects are deleted.Secondly,after roughly estimating the rotation angle and homography matrix between the two images according to the obtained corresponding point pairs,the candidate lines are determined through the angle constraint and epipolar constraint,which can reduce the line search range in search image and exclude a large number of mismatches.In addition,the overlap between the reference line and the candidate line is calculated to ensure the consistency of the endpoints.Thirdly,a circular neighborhood using the midpoint of the line segment as the center and the length as the radius is constructed.The corresponding point pairs are searched in the neighborhood of the reference line and the candidate line to calculate the mesh descriptor.In order to improve the accuracy of line matching,for the line pairs whose number of corresponding point pairs is greater than three,we arbitrarily select three groups of different corresponding point pairs to calculate the line similarity value and take the maximum value as the final similarity of the line segment,so as to determine the corresponding line pair.Finally,the line segments that are not searched for the corresponding point pairs are mapped into the search image using the homography matrix.Three criteria are established according to the angle difference between the two lines and the vertical distance between the endpoints and the radiation information.Based on this,the final matching result is obtained,which can improve the number of corresponding line pairs.To verify the effectiveness and robustness of our algorithm,this paper selects five groups of close-range images in the international open standard test data set,and compares with three existing representative algorithms.The results show that the accuracy and effectiveness of our algorithm are higher than the other three algorithms.The matching accuracy and operation efficiency are improved by 16.1%and 49 times,respectively,and satisfactory line matching results can be obtained for close-range images under different conditions.