High-Definition Map Construction And Application In Localization

Autonomous driving is one of the hottest research areas in the artificial intelligence rev-olution in recent years.With the continuous updating and iteration of autonomous driving technology,the current research hotspot is how the vehicle can complete almost all driving operations under the limited road and environment,at this time the traditional navigation map can no longer meet the needs of the technology,autonomous driving cars need the map to pro-vide richer road element data information,more accurate road element coordinate information,more road.To adapt the development of autonomous driving technology,High-Definition Map（HD Map）came into being,and the route of autonomous driving technology based on HD Map is also regarded by academia and industry as a reliable solution for the continued development of autonomous driving technology.Therefore,the research of HD Map related technologies has very important research significance and value.This thesis carries out systematic and in-depth research on the automatic generation tech-nology of HD Map and its application in localization,which effectively improves the efficiency of large-scale HD Map annotation and further improves the lateral localization accuracy of ve-hicles on the road,and the main work and contributions of the thesis are as follows.（1）In order to obtain the lane line information with height information,this thesis uses Li DAR to detect the ground to estimate the ground equation,and subsequently discusses the camera inverse perspective transformation equation combined with the ground equation.In-spired by cubic Hermite spline curves,a parametric curve fitting method based on asymptotic approximation is proposed to fit spatial 3D points to spatial cubic parametric spline curves.The algorithm indirectly reduces the number of parameters of the spline curve while ensuring that the curve is as close to the sampling point as possible,and the curve also has the advan-tage of global~1continuity.The spatial parametric curve fitted by the proposed algorithm has wider applicability than the two-dimensional ordinary curve equation,which better meets the needs of the actual autonomous driving scenario and improves the accuracy of path planning for autonomous vehicles.（2）This thesis discusses the storage technology of HD Map,analyzes the data required for a complete HD Map,and implements the storage of these data on the open-source Lanelet2map.For the problem of uneven sampling points of HD Map,this thesis uses the numerical integration method to calculate the arc length of spatial cubic parametric curves to achieve isometric sampling on the arc length.Subsequently,an update method applicable to HD Map is also proposed,providing a reliable implementation path for practical applications.（3）This thesis categorizes the information in HD Map into three basic elements:poles,lane lines,and signs.The three basic elements are used to develop their similarity evalua-tion metrics,and the coarse matching of features is completed by considering the number of matches,similarity,and local structure consistency.For the problem of discontinuous match-ing results in time sequence,this thesis combines the confidence of past detection results and sliding.Finally,this thesis combines odometry and HD Map to derive the localizational opti-mization equation in graph optimization equation,and uses robust kernel functions to reduce the impact of incorrect matching,which effectively improves the accuracy of lateral vehicle localization of autonomous vehicles.Finally,the proposed algorithm is tested in KITTI dataset.The experiment proves that the HD Map automatically generated by the algorithm in this thesis meets the accuracy requirement of providing localization information for autonomous driving,and the localization algorithm incorporating the HD Map is more accurate than the common pure odometry-based localization algorithm.