Three-dimensional Reconstruction Of Outdoor Scenes Based On Laser Point Clouds

Three-dimensional reconstruction of outdoor scenes is an essential task in the application of autonomous driving and other robots.Accurate environmental representation is a necessary condition for the safe interaction between vehicles or machines and the environment.Since depth cameras have been widely used,3D reconstruction algorithms based on volume depth fusion are often applied in 3D reconstruction due to their robustness and high-quality reconstruction.However,the lidar sensor is the primary way to obtain the depth information in outdoor scenes.This paper studies the point cloud registration and 3D model reconstruction using laser point clouds.Point cloud registration is the basis of 3D model reconstruction,which can be improved in the initial registration stage based on the traditional registration algorithm,and 3D model reconstruction based on truncated symbolic distance function for volume depth fusion needs to solve the problem of discontinuous triangular surface caused by the sparse laser point clouds in outdoor scenes.Based on the above problems,this paper carries out research on the 3D reconstruction of outdoor scenes based on laser point clouds.The main research contents are as follows:(1)The point clouds are preprocessed by filtering and downsampling.Because the laser point clouds in outdoor scenes have much noise and a large number of points,which will affect the accuracy and efficiency of calculation.The statistical filter and grid filter in the PCL library are selected as the method of point clouds filtering and downsampling,respectively.(2)Point cloud registration of outdoor scenes based on sample kurtosis is studied.An initial registration algorithm based on sample kurtosis is proposed.Firstly,A point cloud is partitioned into several groups,and the sample kurtosis of each group is calculated as the feature of the group.Then,the point cloud subsets with similar features are matched,and the subsets are registered.Last,this algorithm is combined with ICP precise registration algorithm to provide the initial pose for precise point cloud registration.The experimental results show that the new initial registration algorithm can accelerate the convergence speed of ICP precise registration stage,and the conclusion is that the point-to-plane ICP algorithm is better than the point-to-plane ICP algorithm.Finally,the point-to-plane ICP algorithm is selected to be applied in 3D model reconstruction.(3)The 3D model reconstruction of outdoor scenes based on volume depth fusion is studied.A method to improve the surface discontinuities caused by sparse laser point clouds is proposed.Firstly,some continuous point clouds are aligned at a position through local registration in order to increase the density of the point cloud.Then,empty pixels are filled in the depth map after the projection of the point cloud to make the depth information denser.The experimental results show that the 3D model obtained by volume depth fusion using the new depth map dramatically increases the density and continuity of the triangular patches and the authenticity of the scene,compared with the results after fusion of the sparse point clouds.