Study On Generating Desert Vegetation Area DEM Algorithm Based On UAV LiDAR Point Cloud

Unmanned aircraft light detection and ranging(LiDAR)can actively acquire high-density three-dimensional spatial information,and digital elevation model is one of the important terrain products,so it becomes one of the data sources for generating high-precision digital elevation model(DEM).The key technology of LiDAR point cloud processing to generate DEM still needs to be further studied.This paper uses the unmanned airborne LiDAR system to collect the point cloud data of desert vegetation area,and research on three aspects:point cloud reduction,point cloud filtering and point cloud interpolation expands from point cloud simplification,point cloud filtering and point cloud interpolation?The box plot detection method is used to process the point cloud attribute information to realize the rapid presentation of non-ground points.The point cloud normal vector is clustered by K-means++method to realize point cloud reduction.At the same time,a K-means clustering filtering method combined with echo intensity is proposed to obtain the ground point cloud.The point cloud elevation prediction model based on RBF neural network method is further established.DEM is generated by Delaunay triangulation interpolation.The main research contents and research results are as follows.(1)Aiming at the problem of a large number of non-ground points in the original LiDAR point cloud,a method based on box plot to remove non-ground points is proposed.Combining the three kinds of attribute information of point cloud echo frequency,height and point cloud echo intensity,the box plot test method is used to obtain the non-ground point cloud,and the point cloud data set before and after the detection is filtered and compared by using the repeated triangle network method.In order to reduce the amount of data involved in point cloud filtering and improve the filtering speed for the test area,the box plot can be used to quickly remove the non-ground point cloud to reduce the filtering time to 32.2s.(2)Aiming at the discrete,sparse and disordered data structure of the original LiDAR point cloud,a KD tree based distributed scattered point cloud is proposed.At the same time,aiming at the massiveness of the original LiDAR point cloud,a K-means++clustering reduction method based on point cloud normal vector is proposed to realize point cloud storage and fast processing.According to the number of point cloud echoes,multiple echo point clouds are removed,and based on the zero-mean normalization method to obtain the standard value of the point cloud attribute,the KD tree is used to establish the point cloud index,and the point cloud K neighborhood is further constructed and estimated by principal component analysis(PCA).The point cloud normal vector is used to determine the optimal number of clusters by the elbow method.Finally,the point cloud is simplified by K-means++clustering method.At the same time,the reduced results are generated into Delaunay triangulation and converted into raster data,and the validity of the method is verified by the correlation coefficient.When the reduction rate is 18.167%,the highest correlation coefficient is 0.892,which indicates that the clustering method can effectively implement point cloud reduction.(3)Aiming at the low efficiency in the process of point cloud filtering,a K-means clustering filtering method based on echo intensity is proposed.Based on the K-means clustering of the point cloud three-dimensional coordinates to obtain different clustering results,the original values of the point cloud echo intensity of different clustering results are normalized,and the echo intensity standard values of different ranges are obtained.According to different clustering results,different range of echo intensity standard values are selected to obtain the corresponding ground point cloud,and the ground point without clustering result is combined to obtain the final ground point cloud,and the effectiveness of the test method in the test area is intercepted.The results show that the method is better.Maintaining the terrain contour and reducing the amount of data on the ground point cloud lays the foundation for the rapid establishment of high-precision DEM in the later stage.(4)Aiming at the problem that the original LiDAR point cloud can't continuously express the surface information due to the discreteness and data loss,a point cloud elevation prediction model based on radial basis function(RBF)neural network is proposed,and a Delaunay irregular triangulation network is proposed.Discrete point cloud elevation values are interpolated using Delaunay triangulation to generate DEM.Under the constraint of point cloud echo intensity,K-means clustering is used to obtain the ground point cloud in the experimental area.The Kriging method is used to interpolate the ground point cloud data with different streamlining rates to obtain better and better simplification rate.The point cloud elevation value is used as a variable to establish the RBF neural network prediction model,and the model is analyzed by linear regression test.The Delaunay triangulation is used to interpolate the discrete point cloud elevation values to generate DEM.The results show that the decision coefficient R~2 of the point cloud elevation prediction is 0.887,and the root mean square error RMSE is 0.168,which indicates that the RBF neural network has a good prediction effect on the point cloud elevation value.