Driverless Vehicle Localization Method Based On LiDAR

Positioning technology is an important part of driverless technology system.Traditional positioning methods,such as satellite positioning and v2 x,are limited to complex traffic scenes like cities and tunnels.It has been a research hotspot in the field of driverless to realize the autonomous positioning by the Li DAR sensor.Registration algorithm is essential for localization based on Li DAR.Traditional point cloud registration algorithms,such as ICP and its improved algorithms,are faced with the disadvantages of narrow application and easy to fall into local minimum solution.Therefore,in order to adapt to the complex urban environment,this dissertation proposed a new point cloud registration algorithm,and designed a new lidar odometer and SLAM(simultaneous localization and mapping)algorithm based on the proposed registration algorithm.The main works of this dissertation are as follows:(1)A new point cloud registration algorithm is proposed based on the Point Net++ and ICP algorithm.The deep feature descriptors of point clouds are extracted by Point Net++,and the corresponding pairs of two point clouds are searched by the feature differences for improving the accuracy.(2)A new lidar odometer is proposed based on Point Net++&ICP,which uses rough and fine registration to realize motion estimation between adjacent frames.Firstly,the odometer preprocesses the point cloud with filtering and motion compensation to ensure the update frequency of odometer;Because Point Net++&ICP can overcome the initial value sensitivity problem,Point Net++&ICP can be used as a rough registration to provide a better initial value for the fine registration of ICP,so as to accelerate the convergence speed of ICP and avoid ICP falling into the local minimum solution.(3)This dissertation proposes a new SLAM algorithm,which uses Point Net++&ICPodometer as the front-end to improve the accuracy of localization.Meanwhile,the scan context algorithm is used to detect the loopback and send the signal to the backend for global optimization,so as to reduce the positioning error of the front-end.Experiments show that the proposed registration algorithm,Point Net++&ICP,is robust to different types of point cloud and can meet the requirements of scene change and converge rapidly;As a rough registration,Point Net++&ICP can significantly improve the positioning accuracy of odometer;The SLAM algorithm based on Point Net++&ICP-odometer and Scan-Context can successfully detect the loopback and further improve the positioning accuracy.