| Multi-view 3D reconstruction is used the multi-view images as input,which has the characteristics of low cost,convenient operation and high real-time performance,therefore,it is widely used in visual positioning and navigation,unmanned vehicle environment perception,mapping and other fields.Sparse 3D reconstruction as an important role in multi-view 3D reconstruction,its key is to realize estimate camera pose parameters and recover sparse scene structure,and the main development trends of existing related technologies are as follows:on the one hand,the performance of traditional feature point detection and matching algorithm has been difficult to be further improved,and the traditional pose estimation algorithm often needs repeated iteration to eliminate the interference of outlier correspondences,resulting in low efficiency.On the other hand,with the in-depth application of deep learning in computer vision,more and more researchers focus on the application of deep learning technology in the field of 3D reconstruction,expecting further improvement in accuracy and efficiency compared with traditional algorithms.Based on the limitations and research trends of key techniques of traditional sparse3D reconstruction,this paper uses deep learning algorithm to study sparse feature matching and camera relative pose estimation in sparse 3D reconstruction,and the main content including the following three aspects:（1）In response to the problem that Super Point feature detection network can only extract pixel-level feature points,sub-pixel feature points detected by SIFT are combined with Super Point feature description to improve the precision of Super Point feature points while retaining its powerful feature description ability.Finally,the Super Glue feature matching network is combined to achieve high-performance sparse feature matching based on multi-algorithm fusion.The number of correspondences obtained by this method is 2.04times that of SIFT algorithm,and the proportion of correspondences with symmetric epipolar distance≤10-4 is increased by 7.2%compared with SIFT algorithm.（2）For the inevitable problem of outlier correspondences in the matching results,a correspondences classification network is designed based on OANet network for the elimination of outlier correspondences.The network is used to classify the matching results of the sparse feature matching algorithm based on multi-algorithm fusion on the DTU test datasets,and the proportion of inlier correspondences increases by 8.3%.（3）Aiming at the problem of low efficiency of the traditional camera relative pose estimation algorithm combined with RANSAC,a camera relative pose parameter solving network was designed,which took the correspondences as the input and the 6-parameter camera relative pose as the output.On the basis of screening out outlier correspondences by correspondences classification network,the relative pose parameter of camera is solved by using the network,which is 1.9 times faster than the traditional algorithm and has smaller error.Compared with the existing deep learning algorithm based on pixel semantic features,the average translation error is reduced by more than 2.271°.Finally,based on the camera relative pose parameters obtained by our algorithm,we test the sparse 3D reconstruction of left and right images on the DTU datasets with known camera internal parameters.The mean and variance of the sparse 3D point cloud reprojection error obtained are lower than those of other algorithms listed in this paper. |
