Accurate Intelligent Vehicle Localization Method Based On Vehicle-Bone LiDAR And GPS Fusion

In the 21 st century,with the continuous development of new technologies such as information communication,big data,artificial intelligence and their wide application in the automotive industry,the traditional automotive industry is seeking transformation.“Intelligent Vehicle” has become a hot research direction in the field of intelligent vehicle academia and industry,and high-precision vehicle positioning technology is the foundation and core of intelligent vehicle technology.According to the types of sensors,intelligent vehicle positioning methods can be divided into GPS-based positioning methods,vision-based positioning methods and LiDAR-based positioning methods.Because of the blind area and low precision of GPS signal,GPS-based positioning methods often need to be combined with high-cost and high-precision IMU(Inertial Measurement Unit)sensors.Vision-based positioning methods have problems on vulnerability to environmental impacts,poor accuracy and reconstructing 3D images.The LiDAR-based positioning methods have attracted more and more attentions from smart car manufacturers and academia because LiDAR can directly obtain dense 3D environmental perception information and high positioning accuracy.In recent years,with the production cost of LiDAR greatly reduced,its application prospects in intelligent vehicles are also recognized.This paper presents a high precision positioning method for intelligent vehicle based on LiDAR data: Firstly,road scene representation model based on LiDAR data is built: A distance-weighted LiDAR cloud projection method is proposed to generate 2D scene image and 3D structure from LiDAR clouds,and using visual feature descriptors to extract scene features of 2D scene images,and using the LiDAR Odometry method to obtain vehicle trajectory information,finally high-precision road scene representation model is built,the model includes high precision trajectory information,2D scene features,LiDAR clouds and their 3D structures.In the positioning stage of intelligent vehicle,multi-scale positioning strategy is adopted to achieve positioning purpose,which contains rough positioning,node-level positioning based on scene feature matching and metric-level positioning based on LiDAR cloud and its 3D structure.The proposed method has been tested by using the public KITTI database and the actual data collected in the field,which were collected by two different types of LiDAR sensors,Velodyne HDL-64 e and Velodyne VLP-16,respectively.Experimental results show that the high precision positioning method of intelligent vehicle based on LiDAR data proposed in this paper can achieve the 25 cm accuracy of positioning and can be applied to different sensors and different scenes.In addition,a dynamic calibration and fusion method of LiDAR and GPS is proposed: firstly,the space calibration based on trajectory of LiDAR and GPS is carried out,then,the LiDAR trajectory and GPS trajectory are dynamically calibrated and fused by linear Kalman,which can optimize LiDAR positioning trajectory.The main contributions of this paper are as follows:(1)A distance-weighted LiDAR cloud projection method is proposed.In this paper,a distance-weighted LiDAR cloud projection method is proposed.The method divides the LiDAR point cloud into grids with different resolutions according to the LiDAR model,which determines the pixel values of the scene image according to the number of points within the grid and the threshold related to the distance,in addition,the 3D structures of LiDAR clouds are calculated by computing the average height of points in the grid.(2)Road scene representation model based on LiDAR data is built.The road representation model based on LiDAR mainly contains three elements: high-precision trajectory information,LiDAR clouds and their 3D structures,and scene features of 2D scene image.Trajectory information is collected by LiDAR Odometry method and corrected by high-precision IMU.The 2D scene features are obtained by projecting the LiDAR clouds into 2D scene images by using a distance-weighted LiDAR point cloud projection method and extracting scene features by using visual feature descriptors.The 3D structures of LiDAR clouds are calculated by computing the average height of points in the grids of LiDAR clouds.(3)A multi-scale and high precision intelligent vehicle positioning method based on road scene representation model is proposed.In the high-precision intelligent vehicle positioning method,multi-scale positioning strategy is applied,that is,initial positioning,frame-level positioning based on scene feature matching,and metric-level positioning based on 3D structures.(4)A dynamic calibration and fusion method of LiDAR and GPS is proposed.Firstly,the space calibration of LiDAR and GPS based on trajectory is carried out,and then the LiDAR trajectory and GPS trajectory are dynamically calibrated and fused by linear Kalman method,which can optimize LiDAR positioning trajectory.