Freespace Detection Based On 3D Lidar Point Cloud

In recent years,with the development of artificial intelligence and Internet plus,the intelligent network equipment has been pushed into the upsurge of the times.Intelligent network association vehicles have been constantly optimized,improved driving safety and reduced city traffic pressure,and also brought better driving and leisure experience to users.Since 2015,China has continuously launched relevant policies to guide the development of intelligent networked vehicles,and now it keeps pace with the international trend.Intelligent networked vehicle software system mainly includes environment perception and understanding,planning decision and motion control modules.Environment perception and understanding is the technical basis of intelligent networked vehicle,including freespace detection,target recognition,obstacle tracking and other parts.Freespace detection aims to obtain the direction and area range of vehicle safe driving,and provide reference for subsequent planning Decision making and motion control provide important reference,which is the basis of environment perception and understanding.At present,the detection of freespace based on 3D lidar mostly adopts road roadside detection or road surface detection.Road roadside detection can obtain smooth road extension trend by curve fitting of roadside points,and road surface detection can obtain road safety area by removing obstacle points.The two detection methods have different focuses,and can obtain more comprehensive road information by taking into account both of them.At present,3D lidar point cloud processing methods from artificial feature extraction to point cloud deep learning,the effect is constantly improving.At present,point cloud deep learning can obtain more accurate Road area through segmentation network,but can not obtain smooth road extension trend,which still needs to detect roadside based on artificial feature extraction method.This project takes the detection of freespace as the research object,only inputti ng3 D lidar point cloud,simultaneously detecting the roadside and road surface,in order to obtain smooth road extension trend and accurate road safety area,and provide reference for vehicle safe driving direction and safe driving area.The method based on artificial feature extraction is used to detect the roadside,and the method of point cloud deep learning is used to detect the road.The main contents of this paper are as follows(1)Construct the system and method architecture of freespace detection,and establish the freespace detection model.The freespace detection system includes:vehicle mounted 3D lidar sensor,detection algorithm calculation unit and result output module.Road detection method based on Cloud Architecture: road detection method based on artificial learning depth.The freespace detection model includes: vehicle,road edge and road area.(2)The method based on artificial feature extraction is used to detect roadside.The process of roadside detection includes: point cloud down sa mpling,candidate point extraction based on point cloud harness feature,roadside point partition and extraction,roadside fitting,and roadside point inter frame tracking.Voxelization is used to reduce the number of point clouds.Candidate points are ext racted by using the features of adjacent points in the harness: radial distance difference,gradient,angle and height difference.The roadside points are divided by the following base point base axis method,which is suitable for non-straight roads.The roadside points are extracted by the angle search method in the harness,and the angle between the candidate points in the harness and the corresponding base axis is calculated respectively.The least square method is used to fit the quadratic curve.Kalman filter algorithm is used to track the roadside points between frames,and prediction model and detection model are constructed to obtain the optimal estimation of roadside points.(3)The point cloud deep learning method is used to detect the road surface.The process of pavement detection includes: building Polar Net network structure,establishing loss function and selecting optimization strategy.Polar Net network structure includes: circle gridding of point cloud based on polar coordinate system,extraction of point cloud features in grid using Point Net,processing of circle feature grid using Ringcnn,and segmentation of point cloud using UN et architecture.The loss function is constructed by Lovasz Softmax,which is based on the cross parallel ratio IOU,and the discrete function is smoothed by using the mathematical tool Lovasz extension.The optimization strategy adopts Adam optimizer in the framework of Python.(4)Using KITTI data set to train and test the point cloud segmentation network Polar Net,using the group platform car roof 32 line lidar to collect environmental point cloud data,the point cloud data are input into the roadside detection subsystem and the road detection subsystem respective ly,to obtain smooth road extension trend and accurate road safety area.