Research On The Key Technology Of Point Cloud Sampling

With the rapid development of 3D laser scanning technology,the point cloud data in real scenes can be acquired efficiently.The acquired data can have a large number of noise points,huge data volume,and missing data,which cannot be used directly.Therefore,it becomes necessary to sample the point cloud data for research.In this paper,the sampling research is carried out for the point cloud data in the field of civil engineering and construction.For the existing filtering sampling algorithms,which have the problem of incorrect filtering or retention in the sampling process,a combined filtering algorithm is proposed to extract the target point cloud completely.In addition,a deep learning-based approach is used to further complete the missing point cloud data on the basis of preserving the geometric structure of the object,and to improve the quality of urban architecture point clouds to make them more complete.The main research of this paper is as follows:1.A combined filtering-based point cloud extraction method for bridge is proposed.The non-target points are filtered out by analyzing the characteristics of color,distribution shape and range size of point clouds based on civil construction site scenes.Firstly,the vegetation point cloud is coarsely filtered by using the dispersion method based on the feature of dispersion of vegetation point cloud distribution;secondly,the radius filtering algorithm is improved based on radius filtering and making full use of the idea of color and elevation features to filter the remaining vegetation point cloud finely;finally,the ground point cloud is filtered by using the normal filtering model.The experiments show that the combined filtering algorithm can extract the bridge building point cloud more completely and with higher accuracy compared with the existing filtering algorithms.2.A deep learning based point cloud upsampling method is proposed.For the problem that the geometry of object holes is overfilled and the boundary is blurred,we design a weighted graph convolutional network by graph feature enhancement and boundary information weighting.Among them,the graph feature enhancement module is used to reduce the similarity between different nodes,which in turn solves the problem of overfilled holes.The boundary information weighting is used to calculate the feature similarity between neighboring nodes based on the information of both point cloud space and features,which is used as the boundary weight of the point cloud graph,thus solving the problem of object boundary blurring.Experiments show that this method has better upsampling effect.