Feature Line Matching Algorithm Based On Perspective Transformation And LBD Descriptor Constraint

With the development of theories such as photogrammetry and remote sensing,close-range images have gradually become the main data source for 3D reconstruction based on images.Since the close-range image contains rich feature line,it becomes a solution idea to use line matching to realize the reconstruction of 3D models,and the determination of the corresponding line becomes the key link in line matching,which will directly affect the final quality of line matching results.At present,line matching usually uses the matching algorithm jointly constrained by geometric relations and support domain descriptors.As a common algorithm in geometric relations constraint,the homography constraint exists certain limitations,when the image does not exist a complex non-planar scene,the region within the feature line neighborhood is considered as planar,only then the feature attributes within the neighborhood can satisfy the homography constraint;the line band descriptor,as a common descriptor,has the problem that the support domain is difficult to be unified,which makes it unable to cope with images with complex textures,and the final line matching results obtained are generally not reliable.To address the above problems,this paper proposes a feature line matching algorithm based on perspective transformation and line band descriptor constraint.Compared with the existing algorithms,the main research of this algorithm includes:(1)Introducing perspective transformation constraint instead of homography constraint.In the initial matching,the matching area is limited by using the perspective matrix to realize the position constraint between the target line and the line to be matched.(2)Effectively constrain the matching search range.Based on the perspective transformation constraint limited area,the initial matching is achieved jointly with geometry-related constraints,including the target line endpoint constraint,vertical distance constraint and midpoint constraint to achieve the effect of streamlining the number of candidate lines,achieving the matching of most of the feature line and obtaining the initial corresponding line.(3)Improving the line band descriptors.Subsequent secondary matching is performed for feature line that are not involved in the initial matching,and the line band descriptor is improved by constructing a unified support domain with multiple constraints.Among them,in order to solve the problem of large differences in the neighborhood information of the corresponding line due to the inconsistent endpoints of the corresponding line,the overlapping line segments of the matched lines are calculated by using the epipolar constraint,and then the support region is constructed with the overlapping line segments,so as to improve the matching accuracy.(4)Corresponding line checking.The Euclidean distance constraint and the nearest neighbor distance ratio constraint are used to measure the similarity,to discriminate the quadratic matching corresponding lines,and to add the obtained corresponding lines to the initial corresponding line set to obtain the final corresponding line set.Twelve representative images are selected for experimental testing,and the results show that the algorithm shows good robustness and universality in most cases,especially in dealing with complex scenes with blurred texture,low resolution,mostly repetitive similar texture information,and drastic viewpoint changes.