Dense Reconstruction Based On Monocular SLAM

With the rapid development of artificial intelligence,computer vision is an indispensable part of current artificial intelligence and attracts more and more researchers' close attention.3D reconstruction,on the other hand,belongs to a rather hot research point in computer vision.The main research of 3D reconstruction is to recover the three-dimensional information of the surrounding environment based on the images obtained by the camera.This thesis mainly studies in the following two aspects: the dense three-dimensional reconstruction based on direct method and the dense three-dimensional reconstruction based on the feature point method,and makes some improvements on the reconstruction effect.The completed work is as follows:1.Paving the technical principles needed in the subject such as the rigid body movement in three-dimensional space,the Euclidean transformation,the pole constraint,the essential matrix,the homography matrix,and the algorithmic basis of the feature-based and direct-method visual odometry.2.In the dense reconstruction algorithm based on the direct method,the whole process is emphasized,including the four main parts of photometric calibration of the global camera,visual odometer front end,data bridging and dense three-dimensional reconstruction back end.Aiming at the semi-dense point cloud provided by visual odometer and the camera pose algorithm combined with graph optimization and segmentation algorithm to realize the environment plane estimation,the dense map algorithm is used to optimize the result of dense point cloud.The proposed data point cloud selection strategy,a camera coordinate system strategy,data exchange storage pool to achieve better data fusion,improve reconstruction efficiency and reduce resource consumption.The SLAM public dataset was tested experimentally many times.The algorithm results show that the 3D reconstruction algorithm is efficient and robust.3.In the dense reconstruction algorithm based on the feature method,this paper mainly introduces the process of learning the edge information of the image by combining depth with the keyframes provided by SLAM for image segmentation and edge detection.We propose an image fusion algorithm to obtain a more obvious contour map and combine with the sparse point cloud for dense reconstruction.We use a fusion algorithm based on dictionary similarity calculation for the repetitive point cloud display,which reduces the point cloud storage space and improves the reconstruction display effect.We use the dataset in our experimental process,to achieve a relatively good reconstruction results and efficiency.The datasets used in this paper are all public datasets,and the experimental results show that our algorithm has high efficiency and practical value.