Research On Point Cloud Simulation Technology Of Autonomous Driving Lidar Base Data Drive

Autonomous driving test technology plays an important role in autonomous vehicles.Its reliability determines the reliability of autonomous vehicles.At present,autonomous driving tests can be divided into two aspects,one is real road testing,and the other is Simulation test.However,the simulation data of the simulation test has the problem of low degree of authenticity,and the use of this type of data will lead to a decrease in the performance of automatic driving.This paper focuses on improving the authenticity of the automatic driving simulation point cloud dataset and expanding the diversity of the dataset.Based on the data-driven method,the roadcollected point cloud dataset and the virtual point cloud dataset are processed,and a large number of highly realistic datasets are reconstructed and synthesized.The point cloud dataset is used for automatic driving test,and the reliability of the dataset is verified through the test.This paper first establishes a lidar simulation model based on the principle of lidar,extracts the point cloud data of the virtual scene,and selects the existing KITTI dataset as the real point cloud data to ensure the authenticity of the synthesized point cloud.In addition,in order to expand the point cloud dataset,use Synthesize a large amount of real point cloud data based on the depth information of the depth image.Then on the basis of the existing dataset,after preprocessing the data,a static scene point cloud map is constructed based on the NDT matching algorithm.In order to ensure the accuracy of the constructed 3D map,this article adds GNSS constraints and The loop constraint effectively solves the problem of map ghosting.Next,a point cloud segmentation network is designed based on the full convolutional network to obtain the semantic information in the original point cloud and extract the traffic flow point cloud data.In order to solve the problem of point cloud loss caused by the traffic flow point cloud due to occlusion,it is based on generative confrontation The network has designed a point cloud repair network to achieve end-to-end point cloud repair.Then classify the extracted traffic flow point cloud according to distance and angle to obtain a traffic flow point cloud dataset,and then place the traffic flow point cloud on the static scene map based on the real traffic trajectory,and finally obtain a synthetic point cloud dataset through rendering.Finally,the reliability of the point cloud dataset obtained in this paper as the training data of the autonomous driving perception layer is evaluated.Using point cloud semantic segmentation as the evaluation method,using squeeze Seg as the semantic segmentation model,using the synthetic dataset,KITTI dataset,and CARLA dataset as the dataset to generate the corresponding weight file,and finally using the same dataset as the test dataset.Using the overall accuracy and intersection ratio of the test results as evaluation indicators,it is verified that the synthetic point cloud dataset obtained by the method in this paper has a higher degree of reality than the dataset collected by the game engine.In addition,the auto-driving ACC function is verified through real-vehicle tests,which proves the feasibility of using the dataset in this paper to test the auto-driving function.