Large-scale Structure From Motion Research Based On Multiple Video Sequences

In the era of big data,the demand for 3D geographic information is becoming more and more urgent.Traditional 2D geographic information cannot fully meet the needs of social development.The acquisition of conventional 3D data cost so much;therefore,how to reconstruct 3D information using Structure from Motion(SfM)based on images obtained by low-cost equipment is very significant and promising.SfM is mainly based on multi-view geometry,and the reconstruction procedures mainly includes feature detection,camera pose estimation,and bundle adjustment etc.In this paper,we studied the whole process of SfM and its subsequent dense reconstruction.We improved the steps of feature detection and bundle adjustment.The main research contents and conclusions of this paper are shown as follows:(1)For the problem that the correct matching will be eliminated in conventional feature extraction and matching method,this paper constructs a feature matching method based on essential matrix.Under the depth scene,the shortcomings of conventional matching purification method using homography matrix.Such as false elimination of correct matching points,concentrated distribution of matching points after purification and poor purification effect.In order to improve the speed and efficiency of purification,this paper uses the ORB algorithm for feature detection and matching.The ORB algorithm has high matching efficiency but will generate many mismatches.In order to solve this problem,this paper adopts the epipolar geometry constraint as the model of matching point purification.We compared and analyzed the accuracy of the homography matrix,the fundamental matrix and the essential matrix for matching point purification in various scenarios such as depth and planes.Through experimental verification,the commonly used homography matrix will eliminate a large number of correct matches in scenes with the depth scene.While the essential matrix and the fundamental matrix have fewer correct matches.Moreover,the essential matrix can eliminate incorrect matches better than the fundamental matrix,and the purification effect is better.(2)Aiming at the weakness of the same camera parameters were optimized into different values in sparse reconstruction of 3D point cloud,this paper constructs a bundle adjustment method with shared camera parameters.In the process of 3D point cloud reconstruction,it is necessary to perform multiple bundle adjustments for optimization.The conventional bundle adjustment operation is very complicated.In order to carry out reconstruction and optimization better,this paper constructs a bundle adjustment method to share the camera's internal parameters,which can obtain reliable camera's internal parameters through camera calibration or self-calibration.Through mathematical deduction,our bundle adjustment method is sparse.It can be solved quickly by mathematical methods.We reconstruct the target scene by SfM based on multi-video sequence.(3)In the process of 3D point cloud reconstruction,in order to solve the problem of fewer reconstruction points in the blind area of the top field of view,multiple video sequences are used in this paper.The data obtained on the ground and in the air contain rich information on the side and top of the ground feature respectively.Using multiple video sequences as data sources can solve the problem of holes on the top of the ground feature reconstruction.In order to get dense point cloud,this paper uses the CMVS/PMVS method to intensively reconstruct the ground objects,and finally reconstructs a dense,high-quality point cloud with a relatively uniform surface.