Optimization Research Of 3D Reconstruction Algorithm Based On Depth Map

As one of the main research directions in the field of computer vision,3D reconstruction technology has been used in many fields.3D printing technology take a 3D model of a real object as input to achieve a 1:1 reduction of the real scene;Historical artifacts can be saved as 3D models in a digital way,which avoids damages by natural and human factors;Apply accurate 3D models to the film industry to achieve stunning audiovisual effects.At present,the model obtained by the 3D reconstruction algorithm mainly has the following problems:the error between the reconstruction model and the ground-truth data is large;the quality of the mesh patch is poor,and some irregular patches appears,such as:self-intersecting patches,Highly folded side patches.This paper mainly optimizes the above problems to improve the quality of the reconstruction model.Considering the limitations of 3D reconstruction based on depth camera and 3D reconstruction based on deep learning(hardware equipment,training time,scope of application,etc.),this paper takes depth map based 3D reconstruction algorithm,an implementation of multi-view 3D reconstruction,as the basis of research for reconstruct optimization.Through the analysis of the traditional depth map based reconstruction algorithm,the factors that have a negative impact on the reconstruction model are summarized,and the optimization design and implementation are carried out for these factors.This paper mainly does the following optimization work:Apply multiple constraints to the matching view selection.Through the baseline constraint,the view constraint and the pixel constraint,the best matching view of the reference view is obtained and used as the input of the BM(Bidirectional Matching)algorithm to obtain the initial depth information.The probabilistic model is applied to the 3D reconstruction process.Pixel-level view selection is implemented by setting hidden variables in the model and the initial depth information is optimized by the depth values of the neighborhood pixels.Through the angle between the normal vector of the corresponding space 3D point and the camera direction vector in the depth map,the noisy point can be eliminated,which accomplishes the preprocessing of the depth map fusion.The above optimization algorithms have been applied to the multi-view 3D reconstruction optimization system.Through the comparison experiments of 3D reconstruction on the standard public ToHoKu dataset and the real scene dataset,it is verified that the proposed optimization algorithm can simultaneously improve the accuracy of the reconstruction model and the quality of the mesh patch.