Research On Multi-view 3D Reconstruction Algorithm Based On Monocular Vision

Using two-dimensional images to reconstruct the three-dimensional point cloud model of the target has always been a hot topic in computer vision research.Researchers have proposed a variety of reconstruction methods,one of which is to reconstruct the 3D point cloud of the target by studying the correspondence between the multi-view images of the target object under the premise of known camera internal parameters.This reconstruction method has low requirements for equipment and data acquisition,and has wide research value both for laboratory research and industrial applications.The reconstruction result of the three-dimensional reconstruction method based on multi-view depends largely on the accuracy of the camera ’s internal parameters.At present,the widely used camera calibration method is to take some checkerboard images and then calibrate the camera by Zhang ’s calibration method.The selection of poor calibration images may lead to worse calibration results.In order to improve the accuracy of the camera internal parameter calibration results,this paper proposes a method based on adaptive image filtering.A small number of images are adaptively selected from the massive checkerboard images for camera calibration.The calibration results obtained by this method are more stable and accurate than the traditional calibration methods and calibration toolbox results.In order to determine the relationship between each view,it is necessary to extract and match the feature points of the view.The traditional feature extraction and matching uses artificial feature extraction operators such as SIFT to extract the image,and then uses the Euclidean distance and ratio verification method for feature matching.This method has the problems of insufficient number of feature points and mismatching.Therefore,this paper proposes feature point extraction based on Super Point network and further filters the purification matching results through reverse verification based on information divergence(ID)to improve the reconstruction quality.On the basis of feature extraction and matching,the pose of each view is restored by polar geometric constraints,and the reconstruction of sparse3 D point cloud is completed by triangulation.Finally,the sparse point cloud is densely reconstructed by PMVS algorithm.In order to verify the performance of the algorithm,experiments were carried out using the online three-dimensional data set of the Oxford University Vision Group.By comparing the results of feature extraction and matching,the proportion of feature points in the central target area verifies the performance of the improved feature extraction and matching algorithm.The reconstruction quality is analyzed by combining the reconstruction error with the proportion of feature points and reconstruction points.Experiments show that the algorithm further reduces the number of mismatches and obtains better reconstruction results.