Hierarchical Matching Method For Wide-baseline Images Via Junction-point-based Constraints

Image matching is a key step in many visual tasks such as three-dimensional reconstruction.It establishes feature correspondences for multiple images under different conditions(illumination,camera position,viewpoint,etc.),thus providing prior information for subsequent tasks.Unfortunately,satisfactory matching results between wide-baseline images are difficult to achieve due to considerable viewpoint,scale,and illumination variations.How to achieve reliable feature matching between wide-baseline images is a difficult problem to be solved urgently.This study is oriented to urban wide-baseline images without high-precision positioning information,and presents a hierarchical feature point matching method with local structure constraints to solve the problems caused by the inconsistency of the image content of the feature region calculated by the traditional method for the corresponding points.First,we use image coplanar line segments to construct junction-points,and uses the junction-points and interest points obtained by traditional feature point detection operators as matching primitives.In view of the strong robustness of the local structure of the junction-points to the image viewpoint variations,in the initial matching,the junction-points are used as the matching primitive,and viewpoint invariant feature region is constructed based on the geometric position relationship of the line segments corresponding to the junction-points and other line segments in the neighborhood.Then,feature descriptors that are robust to viewpoint variations are constructed and similarity measurement is performed to obtain junction-point matches.Afterward,interest-point matching is performed on the basis of junction-point matches to obtain more point correspondences.Lastly,the junction-point matches and interest point matches are merged and input into a RANSAC framework for outlier elimination and generate the final matching result.Experimental results demonstrate that the proposed method can obtain better matching performance than existing feature matching methods between wide baseline images with considerable viewpoint variations without any prior information,especially in terms of the number of correct matching feature points.