Research On Vehicle Localization Method Based On Semantic Map Matching

In autonomous driving systems,accurate and real-time acquisition of the vehicle’s position and orientation is crucial,as precise pose estimation forms the basis for decision-making in planning and control.Typically,a combination of global navigation satellite systems(GNSS)with an inertial measurement unit(IMU)and Real-Time Kinematic(RTK)technology can provide accurate real-time position information to autonomous driving systems.However,satellite signals can be weak and susceptible to interference or obstruction in urban environments,leading to discontinuity and significant errors in the resulting combined navigation system.This severely affects the accuracy of localization and renders the resulting position estimates unsuitable for subsequent decision-making in autonomous driving systems.Therefore,a localization method that can operate in scenarios where satellite signals are unavailable or intermittently lost is essential for autonomous driving systems.Currently,in such scenarios,pre-built high-precision maps are often used to solve the localization problem.Currently,high-precision maps are usually constructed by LIDAR or vision sensors.In this paper,we address the problems in the map construction method and map matching method in the scenario of no satellite signal or long time signal interruption,and construct a semantic map matching-based localization method to meet the localization requirements by means of monocular vision sensors based on the trade-off between the accuracy and cost of map construction and localization.The main research contents and contributions are as follows:(1)In this paper,the method of map construction is studied and a static semantic map construction method based on monocular camera is constructed.The semantic information required in the map construction is obtained through images,the dynamic regions are excluded,and the 3D point cloud map is converted into a 2D map by projection according to the vehicle driving in a common scene,which only needs to obtain the 2D coordinates in the road plane,reducing the dimensionality of the point cloud matching and enhancing the stability of the semantic features of the map to realize the semantic map construction.(2)To address the scale uncertainty of monocular camera,a method is designed to recover the scale of motion trajectory,and the ground points obtained by semantic information are combined with camera height and pose information to obtain depth,and then complete the recovery of motion scale to make the map matching process scale uniform.(3)In this paper,the localization method is studied.A localization method based on semantic map matching is constructed.The method uses the classical Iterative Closest Point algorithm and introduces semantic information to constrain the selection of nearest point pairs to align the local map established in the localization process with the pre-built global semantic map to obtain the current position.(4)A positioning method that fuses the positioning results of visual odometry and map matching is constructed.By fusing the two different positioning information,the disadvantages of each of the two positioning methods are reduced,the accumulated errors in the visual odometry are corrected by using the map matching results,and the poses provided by the visual odometry are used as the initial values for matching in each map matching,so that more accurate positioning results can be obtained and the absolute positioning accuracy reaches 0.419 m with a maximum error of no more than 0.850m;the relative positioning accuracy reaches 0.139 m with a maximum error of no more than 0.190 m.