3D Feature Point Clouds-based Research On Mapping And Localization In Urban Environments

3D Mapping and localization play a key role in improving autonomous vehicle's adaptable capacity for different environments.The dissertation is grounded on a major research project of the National Natural Science Foundation of China-"Key Technology of the Autonomous Vehicle and Integrated Testing Platform",focuses on 3D feature point clouds-based highly precise mapping and real-time localization in urban environments.The main contributions and innovations are as follows:(1)A new method to estimate an initial position of the current vehicle in the pri-or map,by matching lines in the 2D infrared reflectivity intensity map built with the point clouds of lidar,is proposed,in order to overcome the problem of not having enough overlapping data to register induced by the occlusion of GPS/INS signal.To evaluate the consistency of the lines between the highly precise prior map and the smooth current map,we combine the local feature description,pairwise geometric attribute and structural likelihood to construct an affinity graph for the candidate line pairs and employ a spectral technique to solve the graph efficiently.It overcomes the drawbacks of great gradient changes in the sparse regions and rich similar patterns of the intensity maps.Experiments in the scenes,with GPS/INS signal being occluded,show that our method can always provide an accurate initial position with meter-level accuracy.(2)A robust scan registration algorithm using a few quantity of 3D feature point clouds is proposed.We adjust the raw distorted points by modeling the lidar motion as constant angular and linear velocities within a scan interval;the extraction of feature points located in edge lines and planar patches of the objects not only makes the matched points be more stable but also reduces the number to a lower level;in contrast to the clas-sical point-to-point distance,the reliable plane-based distance measurement can efficient-ly overcome the negative effect of the noise disturbance and the changed measurement views;the new combinative optimal framework of coarse scan-to-scan motion estimation and fine bundle adjustment between sequential scans ensures the robustness and precision of the registration.The algorithm has been validated by a large set of qualitative tests on our collected point clouds on the campus and urban environments,and quantitative com-parisons on the public KITTI urban and highway lidar odometry datasets.(3)A novel 3D map representation method based on K dimensional binary tree struc-ture,feature point clouds and circulant voxel is proposed.Compared with the center of the voxel,the centroid can more precisely describe the object.The trick of circulant voxel makes the memory consumption be constant.Our method has such advantages as no need to initialize large map volumes beforehand,low-memory consumption,memory-efficient usage.(4)A localization algorithm using feature points is proposed.It combines the high frequency but low precision lidar odometry between scans and low frequency but high precision registration between the scan and prior map,achieving the highly precise,real-time localization.GPS/INS is exploited for initialization only and is not required during the subsequent localization process.The experiments on the campus and complicated urban environments show that our algorithm is superior to GPS/INS and it can limit the average longitudinal and lateral errors in the range of 0.3 meters,meeting the need of the autonomous vehicle.The above mentioned algorithms have been successfully used in the autonomous vehicle designed by our team,which greatly improve the adaptable capacity for different environments.