Research On Feature Point Mismatching Elimination Algorithm In Multi-view 3D Reconstruction

With the development of the times and the advancement of technology,3D reconstruction has been applied to all aspects of life.From the initial multi-view reconstruction to the currently popular RGB-D camera depth map reconstruction and LIDAR laser point cloud scanning,the 3D reconstruction system is constantly introducing new technologies and methods,while its overall complexity is also increasing.In terms of operability and cost,3D reconstruction based on multi-view images is still the primary choice.In this paper,we focus on sparse point cloud reconstruction for recovering structures from motion,using an incremental reconstruction framework with high accuracy,including steps such as feature point extraction and matching,camera pose estimation,triangulation,and bundle adjustment.As the first step of reconstruction,the accuracy and precision of feature point extraction and matching play an extremely important role in the subsequent steps.However,for datasets with ambiguity,the traditional process tends to have high errors.To address such problems,this paper starts by removing mismatches of feature points to improve the accuracy of the final matching features to enhance the 3D reconstruction accuracy.By introducing loop consistency constraints,this paper proposes two algorithms for different types of datasets:(1)Algorithm based on maximum spanning tree: Combining maximum spanning tree and minimum spanning tree,searching for edges that satisfy a certain consistency among the remaining edges and adding them to the maximum spanning tree,this process is continuously iterated until the number of edges in the maximum spanning tree satisfies a certain number.The main objective of this method is to remove only the absolutely wrong edges,allowing as many correct view pairs as possible to be retained on datasets with small amounts of data.(2)Algorithm based on orthogonal minimum spanning tree: The orthogonal minimum spanning tree is generated based on the initial epipolar geometry graph,and then it is extended several times to obtain the final view pairs.The main objective of this method is to preserve only the absolutely correct view pairs,allowing successful reconstruction of the model even on datasets with high ambiguity and large amounts of data.The loop consistency algorithm can obtain correct image matches,but cannot filter out the incorrect feature matches in them.To address such problems,this paper proposes a triangulation-based mismatch removal algorithm: By calculating the distance between the3 D points of the stitched image,the mismatches are removed using an adaptive thresholding method.The algorithm only uses the feature point information in this view pair,no other view related information is required and no parameters need to be set to achieve the purpose of mismatch removal.Experimenting with the above algorithm for different types of datasets,the algorithm proposed in this paper is more accurate than the original reconstruction framework and has certain advantages for the latest reconstruction systems and corresponding algorithms.