High Definition MAP Construction For Autonomous Vehicle Based On 3D Lidar

Auto-driving has gradually become one of the hottest research fields in the recent AI revolution,and the auto-driving technology route based on HD map has been widely considered as a reliable solution to realize L4 and L5 self-driving,this because HD map can provide prior map for autonomous vehicles,which plays an important role in localization,perception,planning and simulation.HD map usually has centimeter-level accuracy requirement,so the 3D lidar with the advantages of high measurement accuracy and wide detection range is often used to produce the initial map template of HD map,namely point cloud map.Combining school-enterprise cooperation project "Autonomous Valet Parking" of Jilin University ASCL and Shanghai Automobile City,this paper mainly studies the problem of HD map construction for autonomous vehicle based on 3D lidar and designs corresponding mapping algorithm in small-scale scenes without satellite localization signals such as underground parking lots and large-scale scenes with good satellite localization signals such as highways.The specific research contents are as follows:1.Automatic calibration algorithm for external parameters of 3D lidarCalibration of 3D lidar external parameters is the foundation of HD map construction,aiming at the disadvantages of common calibration methods,a step-by-step automatic calibration algorithm for 3D lidar external parameters is designed innovatively.The first step uses the flat ground as a reference to calibrate ?,?,?z,and transforms calibration problem into a nonlinear optimization problem,which is optimized by PSO algorithm.The second step uses the calibration rod as a reference to calibrate ?,extracts the rod when vehicle travels along a straight path,fits a straight line with the 2D centers of the rod bounding boxes,and calculates ? according to the line's slope.The whole calibration process doesn't need to measure any physical quantities manually,after entering the calibration field,it's automatically executed by the program,and the algorithm has high calibration accuracy without any sensors except lidar.2.Graph optimization mapping algorithm based on point cloud direct registrationAiming at mapping problem in the underground parking lots scene,a graph optimization mapping algorithm based on point cloud direct registration is designed,which uses Pose-Graph as optimization framework and uses G2 O as optimization function library.The algorithm sets the initial vehicle poses calculated by LOAM as the vertexes of the pose graph,uses ICP registration algorithm to correct the deviation of the relative position between the initial poses,uses the transformation matrix between the corrected poses to add local and global loop edges to the pose graph.Finally,the pose graph is optimized by the built-in solver of G2 O to obtain the optimal poses,so that the algorithm solves the problems of "ghosting" and "cumulative error" when direct mapping using LOAM in such scenes,and improves the mapping accuracy.3.Graph optimization mapping algorithm based on point cloud feature matchingAiming at mapping problem in the highways scene,a graph optimization mapping algorithm based on point cloud feature matching is designed,which returns Pose-Graph to the most basic nonlinear least squares problem.The algorithm uses the Ceres function library,uses the initial vehicle poses obtained by the fusion of RTK and IMU as the initial value of the optimization variables,extracts pole,sign and ground features from the point cloud,establishes the residual equations based on the feature matching,and then uses Ceres to solve the equations to get the optimal vehicle poses,so that the algorithm solves the problems of "ghosting" efficiently when mapping in highway scenes.In the above research process,the real vehicle experiments are carried out in the corresponding scenes with the self-driving experimental vehicle equipped with 3D lidar to verify the feasibility and effectiveness of the designed algorithms.