Application And Research Of Large Scale 3D Reconstruction Algorithm Based On Multi-view Stereo Matching

3D reconstruction technology converts a scene into a digital representation in a computer by calculating the depth information of the scene.There are many methods for 3D reconstruction,such as using laser radar scanning to obtain scene depth through triangulation,using structured light method to infer scene depth through light deformation and multi-view stereo matching method is used to restore 3D information from images from different perspectives of the scene.In this paper,multi-view stereo matching method is adopted to realize 3D reconstruction,which is relatively cheap compared with the equipment used by other methods.At present,it is also a hot research direction of computer vision.This method requires only one camera to take photos from multiple angles of the scene.First of all,feature point detection described by SIFT descriptors is performed on all images.Then matching the feature points of all pictures.The position and pose of all photos are recovered by structure from motion algorithm,and then the depth map of each photo is obtained by dense reconstruction.Each depth map is fused to generate a 3D point cloud,in order to enhance the representation of the model,surface reconstruction is also required,that is,connecting the points in the point cloud with patches to generate a surface.Finally,texture mapping is performed,and the texture in each reconstructed patch is given to form a vivid 3D model.We use UAV to take photos of large scenes,including squares,rockeries,various buildings,sports ground,amusement park and vehicles.Using the 3D reconstruction algorithm to reconstruct the scene.We use the deep learning method to improve the dense reconstruction in the above algorithm,that is,restore the depth of the image and fuse the depth map to generate a point cloud.RGB images have three channels,and the feature map obtained through deep learning greatly enriches the number of channels having more global semantic information.This enables the depth learning method to learn more information about highlights,reflections,and texture-less areas,so as to achieve better depth estimation.However,most networks use regression methods to obtain depth maps,resulting in inaccurate depth maps and pool confidence maps.To this end,we propose GDINet,using a probabilistic method to obtain depth maps and confidence maps to enhance the reconstruction effect and generalization ability.We build a Gaussian distribution for each pixel to get a Gaussian map,and iterate it to achieve gradual optimization.The parameters of Gaussian distribution are estimated by deep learning method.The mean value is the depth of pixel estimation.The variance and geometric consistency are used for filtering.We design a loss function with good convergence to train the model and improve the reconstruction effect.We design the initialization module to calculate the parameters of Gaussian distribution for each pixel,controlling the mean and variance are within a reasonable range.We teste our results on the DTU dataset set,Tanks&Templates dataset,and BlendedMVS dataset.Our reconstruction effect is among the best compared with other methods,and greatly improve the performance of dense reconstruction.