Research On Multi-feature 3D Reconstruction Algorithms

With the development of science and technology and the improvement of computational ability,the requirement of human for three-dimensional scene information promotes the development of three-dimensional reconstruction algorithm.Three-dimensional reconstruction is a process of obtaining three-dimensional point cloud and three-dimensional straight line by using the information of two-dimensional feature points,feature lines and their matching relationship provided by disordered images and computing camera position and attitude information through constraints.The main contents of this paper are as follows:Firstly,in order to obtain three-dimensional point cloud and three-dimensional straight line,feature points and straight lines are extracted and matched from the input image.In this paper,we use ASIFT algorithm instead of SIFT algorithm to simulate the image.After various affine transformations and ladder changes of the simulated image,feature points are extracted and matched.Through experimental analysis,the number of feature points extracted by ASIFT algorithm is 10-15 times that of SIFT algorithm.At the same time,ELSD is used to extract feature lines and LBD is used to describe feature lines,generating feature line descriptors with rotation,scaling and illumination invariance.According to the nearest neighbor distance ratio(NNDR),feature points and feature lines are matched,and then dense matching of multiple images is completed.Secondly,on the basis of dense matching results,the initial three-dimensional point cloud reconstruction chooses four images with the largest number of feature points matching each other.The last three images are calculated with the first one respectively.The position and attitude information of the last three images corresponding to the camera are obtained,and three groups of three-dimensional point clouds are calculated.Three groups of 3-D point clouds are re-projected,and the minimum error point is selected as the initial 3-D point clouds.On the basis of obtaining the initial three-dimensional point clouds,the corresponding relationship between the two-dimensional points of the new image and the initial three-dimensional point clouds is used to obtain the corresponding three-dimensional point clouds of the new image.Compared with two initial images,four initial images can obtain more accurate initial point clouds,which verifies the effectiveness and accuracy of the algorithm.Then,on the basis of the result of line matching,four images with the largest number of line matching features are selected,in which the camera coordinate system corresponding to the image with the largest number of line matching features is selected as the world coordinate system,and two images are selected arbitrarily to calculate with the first image to obtain the position and attitude information of the image corresponding to the camera,and then the three-dimensional line is obtained.The three-dimensional straight line is re-projected,and the three-dimensional straight line with the least error is selected as the initial three-dimensional straight line.Using the corresponding relationship between the initial three-dimensional straight line and the two-dimensional straight line in the new image,the position and attitude information of the new image corresponding to the camera are obtained,and finally the new three-dimensional straight line is obtained.Finally,the three-dimensional point cloud and three-dimensional straight line are fused,and coordinate system one is carried out by using the relationship between the initial cameras.According to the relationship between the camera and the initial camera,the position and attitude information of the camera are obtained under the condition that the overall projection error is minimum,so that the final fused three-dimensional point cloud and three-dimensional straight line can be obtained.