Research On LiDAR-based 3D SLAM In Driverless Application

Simultaneous localization and mapping(SLAM)technology based on Lidar has become a research hotspot in the field of unmanned driving,due to it is less affected by illumination and weak txture,the accuracy of positioning and mapping is high,the reliability is good and there are other advantanges.However,all current lidars are not solid-state lidars.In the process of unmanned vehicle movement,the 3D data dynamically acquired by the lidar is not in the same coordinate system,which causes large errors in follow-up motion estimation and map construction.During long driving movement,the accumulation error will result in inaccurate pose estimation and low consistency map construction,and the excessive amount of lidar scan data has a great impact on real-time SLAM technology.In view of the above problems,in the existing 3D SLAM method based on lider,IMU rectified lidar is used for data fusion,and incremental segmentation is used for loop closure detection to achieve higher precision positioning and mapping.The work done in this thesis mainly includes the following aspects:(1)A lidar-based external parameter distribution calibration method based on defined area is designed.The ground and calibration rod data are extracted from the 3D point cloud by setting the area range,effectively reducing the influence of noise points on the calibration result.Based on the ground point extraction and plane fitting,the Particle Swarm Optimization(PSO)method is used to estimate the pitch angle,roll angle and height.The calibration rod data is k-means clustered and fitted to a straight line equation using least squares to estimate the vertical yaw angle of the lidar.(2)A data fusion method based on IMU compensation for lidar motion is proposed.The relative time to initial scan is obtained by the horizontal deflection angle of the point in each frame.Time is used as the key value,and the linear interpolation method is used to obtain the IMU state information corresponding to the current point.Point clouds are rectified by the conversion relationship between the current IMU state and the IMU state of the initial scan.The scan matching transformation is obtained by performing the Point-to-Plane ICP(PP-ICP)algorithm on the rectified 3D point cloud and converting it into a scan matching factor.The state information of the IMU is estimated by the dead reckoning to obtain the Odometry factor for later optimization calculation.(3)The loop closure is detected by the incremental segmentation point cloud.Firstly,the lidar point cloud is incrementally segmented,descripted and matched based on feature vector.Then,accurate loop closure detection is obtained after Random Sample Consensus(RANSAC)verification,and the result of loop closure is converted into the PR factor in the factor graph.Finally,the back-end optimization estimation is performed on the factor graphs that combine the scan matching factor,the odometer factor and the PR factor,so that more accurate positioning and mapping results can be obtained.(4)The experimental verification has been carried out on experimental platform based on ROS(Robot Operating System).The calibration results verify the effectiveness of the calibration method used in this thesis.The effects of dynamic pixel size and feature matching distance threshold on the SLAM results are explored through the KITTI dataset.The experimental results in the campus environment show that the improved method proposed in this thesis improves the positioning accuracy and improves the consistency of the map.Figures 55,Tables 12,and 55 references.