Method for classifying and segmenting sparse point clouds by utilizing point cloud density and rotation information

Abstract

The invention discloses a method for classifying and segmenting sparse point clouds by using point cloud density and rotation information. The method comprises the following steps of preprocessing thepoint clouds; analyzing the coordinate information of the sparse point cloud; analyzing the space and density information of the sparse point cloud; and point cloud classification and segmentation are performed in combination with the point cloud space coordinate features, the density and the rotation features. According to the method, through a point cloud preprocessing mode and mining and utilization of sparse point cloud space information and rotation information, the working mode of point cloud classification and segmentation is improved, classification and segmentation of point clouds (sparse point clouds and dense point clouds) with different densities are compared by adopting the algorithm, and the efficiency is improved by more than ten times; compared with the classification andsegmentation of the traditional machine learning method (filtering), the classification and segmentation of the sparse point cloud by using the method have the advantages that the precision is improved by 7% and is improved from 87% to 93%; and compared with a PoineNet algorithm and a 3DCNN algorithm, the precision is improved by 4%.

Description

technical field [0001] The invention relates to the technical field of point cloud classification and segmentation. Specifically, it is a method for classifying and segmenting sparse point clouds using point cloud density and rotation information. Background technique [0002] Lidar has been widely used in surveying and mapping engineering because of its unprecedented advantages in real-time and accurate acquisition of surface information. The point cloud is the most direct output of the lidar sensor, and it is an objective and detailed expression of all objects seen by the sensor's vision. However, exhaustive means large amounts of data and long-term calculations. Most researchers don't need exhaustive data, just focus on point clouds of regions of interest. Taking mining detection as an example, reclamation researchers pay attention to the growth status of vegetation in mining areas, and deformation monitoring researchers focus on changes in surface morphology during mi...

AI Technical Summary

Problems solved by technology

[0016] The 3DCNN algorithm is a method based on voxel feature search, which indirectly converts the point cloud into a voxel grid for storage and then performs calculations. The accuracy is not high, not suitable for the use of point clouds in mining areas

Images