Research On Dense Propagation Based Three-dimensional Reconstruction Algorithm

3D reconstruction based on image reconstructs 3D information of scenes or objects through multiple visual images obtained from different viewpoints,which is a multidisciplinary research topic involving computer graphics,stereoscopic vision,and image processing.It has great application potentials in virtual reality,scene understanding,cultural relics restoration and protection,target recognition and tracking,and other fields.In this thesis,we focus on recovering 3D point cloud data from obtained image sequence.After summarizing the reconstruction principle and implementation idea of the existing methods,a method based on the combination of feature point filtering and 3D patch propagation strategy is proposed to realize the dense reconstruction,solving the existing problem of low reconstruction accuracy and reconstruction completeness.The thesis mainly completed the following work:1.TZNCC dense diffusion.Firstly,the calibration image sequence taken at different angles is taken as the input set,and the Harris and Do G feature detection operators are employed to extract the corner points and spots respectively.The image sequence is selected on the basis of reference image,an image with a smaller angle between the main optical axis and the reference image is reserved as a candidate image for feature matching;the matching feature points are used as the initial seed,and the matching points are optimized by the parallax gradient and the confidence constraint.The characteristic information such as light,texture shadow and concavity and so on are taken into account.The weighted value of TZNCC between the zero-mean normalized cross-correlation coefficient and cross-correlation coefficient of texture as the matching criteria in the diffusion process and diffuse to the characteristic neighborhood.The camera parameters and matching points are restored to the 3D model point.2.Patch propagation.3D dense reconstruction method based on patch is proposed to densify the discrete 3D points.The initial patch is generated by the TZNCC sieve candidate point based on the 3D model point recovered by the TZNCC diffusion process.The generated initial patch is propagated according to the normal vector and position relation of the adjacent patch,and the dense space is obtained from the relatively sparse seed point Then,for the propagation of the patch,use the geometric consistency and image gray consistency constraints to filter and remove the noise.Then,generate accurate dense three-dimensional point cloud model.Finally,the recovery of the point cloud data is used to reconstruct grid and generate a complete surface model.Proposed 3D reconstruction method recovers accurate and dense point cloud data via two dense propagation processes for matched sparse feature points.Experiments on multiple datasets is carried out to evaluate the performance.Results show that proposed dense propagation strategy increases 2.3 times for initial seed feature points,effectively filters the noise points,and further recovers detailed 3D point cloud model.Compared with PMVS and Visual Sf M,proposed method is more efficient and robust,and reconstructed points owns better distribution,which means superiority.