Research And Implementation Of 3D Reconstruction Algorithm Based On Image

With the wide application of computer vision technology in the fields of games,medicine,surveying and mapping,virtual reality,etc.3D reconstruction algorithm based on image become a research hotspot.Although the passive monocular reconstruction algorithm has complicated process and a long running time,it has the advantage of high scalability,low cost,anti-environment light interference,and has many application scenarios.Therefore,it is favored by researchers.In feature matching,the accuracy of existing algorithms and the number of matching are low.the mismatching will affect the final reconstruction effect.In addition,in the dense reconstruction,there are many empties in the 3D point cloud recovered by the existing reconstruction algorithm,and the visual effect is unsatisfactory.Above those reasons,this thesis has conducted a deep study on the passive monocular reconstruction algorithm.The details are as follows.1.Research the camera model and imaging principle,analyze the basic model of 3D reconstruction,and explain in detail the rigid body movement in 3D space and the mutual conversion between camera and pixel coordinates.Analyze the effect of camera distortion on the reconstruction results.2.In feature matching,this thesis proposes a filtering method based on image local correlation.Experiments show that compared with threshold method,ratio method,twoway matching method,and RANSAC algorithm,it can effectively filter false matches.For the problem of too few matches,also fuses SIFT matching results and ORB matching results to get more matches.When using this result as the input to solve the camera pose,the accuracy of the solution is improved and the solving time is declined.3.This thesis takes the incremental SFM algorithm as an example to introduce the sparse reconstruction process in detail and verify the effect of Bundle Adjustment(BA)on the reconstruction results.According to the analysis of different BA descent algorithms,the RosenBrock function is used to verify the analysis results.4.Aiming at the problems of dense reconstruction,large reconstruction holes,and low density of reconstruction point clouds,a method based on optical flow and ORB feature fusion is proposed.It is proved by experiments that also fills the holes of the dense reconstruction of point clouds.It has more reconstruction points and better reconstruction results than C / PMVS and depth methods.5.An incremental SFM reconstruction system based on a monocular camera is implemented.When using the same number of pictures,compared with visualSFM,the visualization effect of this system is better.When the camera is about 0.5m away from the shooting target,the reconstruction accuracy can reach 1mm.When using this system to estimate the volume of unknown sands,the error is only 3%.