Autonomous Vehicle Self-Localization Algorithm Based On Lidar

Accurate localization is one of the key technologies for autonomous vehicles.In complex urban environment,it is difficult for global navigation satellite system(GNSS)to provide accurate localization for autonomous vehicles.What's more,simultaneous localization and mapping(SLAM)cannot guarantee accurate localization because it suffers from cumulative error in long driving distance.In order to solve the localization problem in the deep urbanized area,this paper proposes a map-based localization system based on LiDAR.Firstly,for point cloud registration,we select feature points based on their curvature and divide them into edge feature points and flat feature points.In order to compute the transformation between two frames of point clouds,the cost function is defined according to the distance from edge feature points to their corresponding lines in the reference frame and the distance from flat feature points to their corresponding planars.The motion of the vehicle will be recovered by minimizing the cost function.Secondly,for mapping,we propose a pose-graph mapping method based on adaptive information matrix using LiDAR and GNSS data.The positions of the vehicle are nodes of the graph.And the edges of the graph are constraints between two positions created from measurements.We add GNSS data as prior position information.According to the point cloud registration result,we construct the information matrix of the constraint edge between the LiDAR-odometry nodes.When the loop closing is detected,an edge is added between the two frames according to the registration result of the two frames.The pose graph is essentially a least square problem,and the prior 3D point cloud map can be generated by stitching point clouds according to the optimized poses.The adaptive information matrix and GNSS prior information improve the accuracy of pose graph.Then,for localization,we propose 3D curvature features-Monte Carlo Localization algorithm(3DCF-MCL).3DCF-MCL is composed of three parts: initialization,prediction and updating.For initialization,we use GNSS data whose signal is good to provide the initial position of the autonomous vehicle in the world coordinate system,and then initialize the particles with Gaussian distribution.As for prediction,we use the Odometry Motion Model as our vehicle motion model,and the wheel odometry is obtained by wheel encoder information.In updating part,we update the autonomous vehicle position according to the matched result between the feature points scanned by LiDAR and the prior map generated by our method.The advantage of 3DCF-MCL is that it combines the accuracy of our 3D feature point registration and the robustness of particle filter.Finally,we validate our algorithm using 3D LiDAR data gathered from real outdoor environments with a vehicle.Experiments show that 3DCF-MCL can provide accurate,real-time and robust localization for autonomous vehicles with 3D point cloud map that generated by our method.Compared with authoritative LiDAR-based localization algorithms,it demonstrates that the localization performance of 3DCF-MCL is better than them.