Research On Filtering And Registration Method Of LiDAR Point Cloud Data

In recent years,laser scanning has made a huge leap forward with the advent of LiDAR.The LiDAR scanning technology has the advantages of flexible structure,efficient processing and so on,and can store the collected mass point cloud data quickly.However,when the scale of the measurement scene is large,airborne LiDAR usually needs to fly multiple airstrips to obtain the complete point cloud data of the test area,so that the systematic errors in different airstrips are different,which leads to the spatial deviation of the coordinates of points with the same name in the airstrip overlap area.In addition,the data types of point cloud obtained under the environment of complex measurement area are complex,including ground object points and a small amount of noise points in addition to the surface points.In order to solve the above problems existing in the original point cloud data and improve the quality of subsequent 3D modeling,this paper takes the 3D point cloud obtained by LIDAR as the research object,and mainly studies the ground point cloud filtering and point cloud registration in the point cloud data processing.The main contents are as follows:(1)The point cloud filtering method based on convolution neural network is developed.In view of the problems that traditional methods in point cloud filtering are greatly affected by terrain features and have many parameter settings,this paper explores the application of deep learning to ground point cloud filtering and regards point cloud filtering as a binary classification problem.First to each point cloud data and its neighbor points of 3D coordinate by weighting function is converted to a fixed size of three-channel image,and then build a multidrop convolution and full connection layer of convolution neural network(CNN)model framework,and the resulting from the conversion of image data used for model training and prediction,finally using the trained model classifying test point cloud.The research results show that,compared with the traditional point cloud filtering method,the point cloud filtering based on deep learning is less affected by terrain features and has stronger anti-noise ability,and the total classification error of point cloud obtained on the standard data set of ISPRS is the lowest.(2)An iterative closest point(ICP)algorithm based on intensity threshold constraint is proposed.The traditional ICP point cloud registration algorithm will fall into the local optimal solution due to the unsatisfactory initial location of the point cloud,and the computational efficiency is low in the face of massive point cloud data.However,in view of the simple principle and wide application scenarios of the algorithm,this paper adds the intensity threshold to the ICP algorithm to realize the point cloud ribbon stitching.Firstly,this method extracted the key points in two navigation belt overlapping areas by using the intensity information of point cloud,and then searched the nearest neighbor points in the key points by kd-tree index.Finally,the ICP iterative method was used to solve the optimal transformation parameters.The results show that the improved ICP algorithm in this paper has greatly improved the registration efficiency,and can ensure the stability of the registration accuracy in the face of massive point clouds.