The Design And Implementation Of Rainy-scene Point Cloud Data Augmentation For Autonomous Driving Systems

Environmental perception is the important cornerstone to ensure the safety of autonomous driving,lidar is the main sensor for autonomous driving environment perception.Under severe weather conditions such as rainy scenes,the laser is absorbed,diffracted or refracted by water droplets or any particles in the air,and the point cloud data detected by the lidar will have measurement deviations.More critically,the effect of the autonomous driving model largely depends on the sample quality of the data set,including the number of samples and diversity.When the sample space is insufficient or the sample diversity is not rich enough,the generalization ability of the model may be insufficient.And there is a result that the recognition rate and accuracy rate are not high.In this thesis,the lidar point cloud data augmentation is carried out for the rainy scene of autonomous driving.It aims to provide point cloud data of severe weather with extremely high collection cost,increase the diversity of the data set samples of the autonomous driving scene,and improve the driving decision-making ability of the unmanned vehicle system.The degree of accuracy makes the mutation data of practical significance for the test of the automatic driving system.After a comprehensive analysis of the project background and technical status,this thesis derives the variation of laser transmittance with laser transmission distance under different rainfall intensities based on Lamberbeier’s exponential attenuation law,raindrop spectrum model and Mie scattering theory,and based on this The related algorithm of the change function design proposed a lidar point cloud data amplification technology for autonomous driving rainy scenes,and then based on the Spring Boot framework design and implementation of the corresponding point cloud data amplification tool.The tools are mainly divided into file upload module,amplification parameter configuration module,point cloud amplification module,visualization module and mutation data set evaluation module.Through this tool,users can upload original point cloud data to be amplified,configure rain intensity parameters in rainy scenes,and perform customized amplification tasks.In addition,the tool also provides 3D point cloud visualization and mutation data evaluation services.Through rigorous system testing,this tool basically achieves requirements,and the results meet the expected standards.This thesis uses the 3D target detection model and the KITTI public point cloud data set for experimental evaluation.The experimental results show that the variable point cloud data generated by the rainy scene amplification tool designed and developed in this thesis is true and effective,and conforms to the basic characteristics of laser attenuation in rainfall.The point cloud augmentation tool developed in this thesis greatly saves the cost of collecting extreme weather point cloud data,enriches the diversity of lidar point cloud data samples in autonomous driving scenarios,thereby improving the generalization ability and robustness of the model,and hopes to play a certain role in improving the safety of autonomous driving systems.