| In autonomous driving systems,accurate global positioning results are the prerequisite for achieving safe vehicle driving and reasonable path planning,so positioning of autonomous vehicles is a key link in autonomous driving systems.At this stage,most positioning systems adopt multi-sensor fusion positioning solutions,and the mainstream sensors include:Global Navigation Satellite System(GNSS),LiDAR,Camera,Radar,etc.Multi-sensor fusion can achieve the complementary advantages of each sensor and improve the accuracy and robustness of the positioning system,but it also leads to the problems of diverse data types and complex system implementation.In this paper,a fusion localization system based on laser odometry localization and point cloud descriptor matching localization is proposed,focusing on the global localization of L4 autonomous vehicle using only 3D LIDAR.Research works below has been perform:(1)A new 3D point cloud descriptor is proposed.The descriptor encodes the point cloud structure information in z-height direction in space in binary,converts the encoded binary code to the corresponding decimal number,and stores the descriptor information in a two-dimensional matrix representation.A descriptor-based matching method is also proposed to achieve global localization of the autonomous vehicle in the environment by matching the current frame point cloud descriptor with a database containing all historical frame point cloud descriptors.Since the descriptor matching establishes the association between current and historical information,it can also be used for loop detection of SLAM systems.(2)A SLAM system based on graph optimization is constructed.The system mainly includes four parts of work:laser odometry construction,back-end positional map optimization,loop detection and map construction based on known poses.Based on the curvature of the laser points,the boundary points and planar point are extracted from the point cloud after the distortion removal process,and the feature points are used to registration the point cloud;the descriptors proposed in this paper are used in the loop detection of the SLAM system to correct the accumulated errors;the open-source tool g2o is used in the back-end to optimize the front-end constructed positional map containing the odometry constraints and the loop constraints;in the map construction session,a 3D point cloud description of the environment is constructed based on the optimized known poses.(3)The accuracy of the constructed SLAM system is evaluated in the KITTI public dataset.First,the performance of the loop detection of the point cloud descriptors constructed in this paper is evaluated and compared with IRIS,SC,and M2DP point cloud descriptors,where our descriptors achieve the best precision-recall performance.The vehicle trajectories estimated by the SLAM system are evaluated and compared with LOAM,V-LOAM,and IMLS-ICP methods.The average translation error of our method is 0.53%and the average rotation error is 0.0021°/m,where the translation error is the smallest among the four methods.(4)A LiDAR-only global localization method is proposed,which fuses real-time LiDAR odometry localization results with a priori map-based matching localization results.The experimental evaluation of the localization performance of autonomous vehicle is conducted in public datasets and real scenarios,respectively.The experimental results of the public dataset show that the translation error is:0.038 m and the rotation error is 0.031°.The localization experiments were conducted on two routes of the real scenario,and the average translation error was 0.067 m and 0.063 m,respectively.In the two routes cumulatively run more than 10 km,the proportion of the distance of the two experimental routes with translation error less than 0.14 m is 99.54%and 100%,respectively. |
PDF Full Text Request