Research On SfM Algorithm For 3D Reconstruction Of Multi-view Images Based On Point And Line Features

Multi-view-based 3D scene reconstruction technology is currently a key research direction in the field of computer vision.This technology constructs high-precision 3D models by extracting geometric information from multiple views,and has broad application prospects in many fields such as virtual reality,robot navigation,medical biology and aerospace.Features are the main means of extracting geometric information,of which there are numerous mature algorithms based on point features,but they still perform poorly in the face of repetitive textures and weakly textured scenes.In addition,how to use line features with intuitive semantics to improve reconstruction accuracy is still one of the problems facing 3D reconstruction.In contrast,there are a large number of point and line features in artificial scenes,and making full use of point-line relationships to constrain the 3D reconstruction process will effectively improve the reconstruction accuracy.This article takes multi-view-based 3D reconstruction technology as the research object,focusing on improving the sparse reconstruction part of 3D point clouds,including feature mismatching rejection,incremental motion recovery structure(Structure from Motion,SfM)mechanism for point-line fusion and line structure reconstruction,aiming at critical guidance for subsequent model reconstruction and improving model refinement.The main research work of this article includes the following aspects:(1)There are problems of visual similarity and dense repetitive structures in multi-view images,and the resulting mis-matches cannot be filtered out using the global constraints of the camera motion model.To address the above problems,this article proposes a position geometry-based mismatch rejection algorithm,in which a quantity-adaptive four-directional region selection algorithm is designed to construct a triplet to calculate the relative position geometry consistency in order to determine mismatches.In order to make this location geometry constraint universally adaptable to various point-line features,two determination algorithms are proposed to cope with both single point feature extraction and simultaneous extraction of point-line features to achieve fast rejection of mismatches and improve the correct image matching rate.(2)The existing line reconstruction algorithm has a large time overhead and the reconstruction process of point-line features is not easily unified.To address this problem,this article proposes an incremental SfM mechanism for point-line fusion.Firstly,this article designs a camera pair selection scoring algorithm to select the best initial camera seed pair by a comprehensive measure of the number of matches,camera angle and planarity.Subsequently,considering the characteristics of line reconstruction which requires high matching quality and difficult reconstruction,this article introduces the intersection ratio property to convert line matching into matching of points on a line,and uses point features to represent line features,thus line features are successfully introduced into the SfM system based on point reconstruction to complete sparse reconstruction and obtain accurate poses and sparse point clouds.(3)The line structure has stronger semantic expression capability,but the existing point-line fusion mechanism only retains the spatial location information of the points on the line and fails to effectively recover the overall structure of the line features.To address this problem,this article improves the clustered line reconstruction algorithm.Firstly,the tentative line structure of each line match is recovered by image poses,then it is 3D scored and the false matches are filtered out according to the scores.Finally,the retained line structures are clustered to retain as much endpoint information as possible and determine the sparse structure of the lines,thus representing the scene structure information more intuitively.In summary,the point-line fusion SfM algorithm proposed in this article provides more accurate sparse reconstruction results for monocular multi-view reconstruction,can intuitively outline the scene structure,and provides a new research idea for high precision 3D reconstruction.The false match rejection algorithm is universal and can be used as a separate library of algorithms to quickly complete the matching refinement step,which is applicable to a wide range of point-line features.