Research On Multi-beam LiDAR Point Cloud Simulation

Because of the precise ranging capability,LiDAR(Light Detection and Ranging)is increasingly used in geological mapping,3D modeling,and autonomous driving.With the improvement of the ability of deep learning to process LiDAR data,the demand for lidar data is also increasing.Compared with image data,the cost of LiDAR data acquisition and post-processing is much higher.And the LiDAR data is sparse and disordered,making it difficult to label the data.Therefore,how to efficiently generate LiDAR data has important research significance.In addition,the working principle of different types of LiDAR is different,and the obtained LiDAR data will have great differences.In order to have better adaptability to different types of LiDAR,deep learning models need to be trained on different LiDAR data.In order to efficiently generate different types of LiDAR data and take advantage of existing data,it is a feasible solution to use existing LiDAR data to simulate different types of LiDAR data.In this context,this paper proposes two different methods of LiDAR point cloud simulation.The main content of this article includes:1.This paper presents a multi-beam LiDAR point cloud simulation based on highdensity point cloud data downsampling.This method first needs to model the main parameters of the LiDAR.The model parameters are the resolution,detection distance,and the number of beams of the LiDAR.Then downsampling the high-density point cloud data to generate simulation data with the characteristics of the original highdensity LiDAR data and the characteristics of the LiDAR model.This method can generate simulation data based on the requirements of different types of LiDAR on the existing LiDAR data.2.This paper presents a multi-beam LiDAR point cloud simulation based on deep learning upsampling.This method performs spherical projection on the original lowdensity LiDAR data first,then learns the features of the LiDAR data through a neural network with an auto-encoder as the main architecture,and upsamples the features to finally generate a higher-density LiDAR data.The data simulated by this method is not only enhanced in density but also retains the characteristics of the original LiDAR data.In this paper,we conduct experiments on the aforementioned research content on different data sets collected by LiDAR,and quantitatively and qualitatively analyze the experimental results.After experimental verification,the LiDAR point cloud simulation method proposed in this paper has good reliability and versatility.