Research On 3d Reconstruction Method With Local Information Constraint

The 3d reconstruction technology for oblique images(oblique 3d reconstruction)is a research hotspot in the field of photogrammetry and computer vision.Because of its flexible,unique type of oblique photography,It not only can provide better stereo image pairs,higher spatial resolution,more available observations and texture information of object side face,but also can complete 3d reconstruction task of large scene with texture,just rely on only one source of oblique images.Compared with other 3d reconstruction methods,such as laser scanning and artificial modeling,It has the characteristics of high data collection efficiency,large scene adaptation,low cost and low equipment requirements,therefore,it is favored by relevant scholars and professionals.Correspondingly,oblique 3d reconstruction also has problems such as large difference in scale,large difference in radiation,redundancy in data volume,redundancy in data expression,obvious noise at the edge of object side,unprominent feature structure,texture inconsistency and high time and space complexity.These problems seriously affect the development of oblique 3d modeling,restrict its potential development potential,and limit the real-time and verisimilitude of 3d products.Aiming at these problems,this paper proposes an oblique 3d reconstruction method with local information constraint,which makes full use of local information of object to reduce data volume redundancy,data expression redundancy and highlight structure information of object.The main research contents and contributions of this paper are as follows:1)This paper discusses the advantages and disadvantages of the oblique image 3d reconstruction technology and the problems that need to be solved at the present stage,introduces the research status,basic theory and key technologies of the oblique image 3d reconstruction,and probes into the core essence and existing problems of the existing relevant algorithms.2)In view of the existing SIFT algorithm's sensitivity to local perspective changes,this paper proposes to use frequency division technology to detect the main direction,so as to make up for the influence of large intersection angle on gradient histogram statistics.In addition,by using the robustness of the structure,the geodesic distance weighting method of anisotropy is used to replace the original gaussian isotropic distance weighting method.At the same time,in order to reduce the influence of geodesic distance control coefficient,this paper referred to ASIFT algorithm for sampling its coefficient space,and used this multi-layer description of structure to perceive object structure information.3)This article in view of problems that exist in the dense matching,namely the observation information redundancy,computing redundancy,expression of redundancy,noise edge information,etc.,first,using object space's sparse point cloud to detect the most preferential match image collection for each reference image,in order to reduce computational complexity,decrease the space complexity of fusing point cloud.Then,the depth map of the reference image is filtered,weighted,optimized and supplemented with invalid areas by taking advantage of the fact that the object noise fluctuation has a limited gaussian range.At the same time,small holes and isolated islands are eliminated to ensure the integrity and accuracy of depth map.Next,the object local plane and edge are detected in the depth map,after that,sparse sampling and noise point removal are carried out for plane region and edge region respectively,in this way,the expression redundancy of point cloud data is reduced and the point cloud accuracy in the structure area is improved.Finally,point cloud data with low data redundancy and high precision are generated.4)In this paper,a solution is proposed to solve the problems of the existing watertight meshing,such as high time complexity,serious mesh redundancy and unprominent structure.Firstly,the initial watertight mesh is constructed from the point cloud data with low data redundancy and high precision by using the soft constraint graph-cut algorithm that takes into account the local object information of the point cloud.Then,analyze the discrete differential geometric properties of the curved mesh,and use the change of curvature to measure the geometrical precision loss in the simplification process of the mesh.so as to resist the edge collapse operation across the structure,retain the edge precision while highlighting the edge structure,and greatly reduce the redundant expression and data amount of the linear piecewise mesh.In addition,this paper considers that two major influencing factors of texture consistency expression in texture mapping are geometric consistency expression and radiative consistency expression.Aiming at the radiation distortion and scale change caused by oblique image,this paper firstly uses the image fuzzy detection function to eliminate the image with poor quality.Then,uses multi-label optimization technology to select clear and realistic images as the texture source of object elements.Then,the global and local radiation correction methods are used to adjust the radiation difference between adjacent surface elements.Finally,the 3d model products are released by means of those methods,such as texture packaging.In order to reduce the difference of texture radiation between adjacent blocks,a pretreatment radiation adjustment function is proposed to homogenize the whole image set.