Research On Localization Technology For Unmanned Vehicles Based On Fusion Of Satellite Images And LiDAR Point Clouds

As a key link in environmental perception,vehicle positioning technology plays an important role in supporting the path planning,navigation,and environmental map construction of unmanned ground vehicles(UGV).This paper is mainly oriented to the unknown road environment under GNSS(Global Navigation Satellite System)rejection,and aims to integrate real-time laser point cloud and satellite image prior information,and conduct related research on positioning methods in structured and unstructured scenarios,so as to improve the positioning accuracy of UGV.The main research contents and innovations of this paper are as follows:1.A novel two-level detection algorithm for intersection attributes of laser point cloud is designed,and a matching strategy with satellite image intersection geometric information is proposed.This method firstly constructs a multi-candidate cluster detection model based on the beam rangefinder model and the ray filtering algorithm.After that,based on the grid location distance specificity,the secondary location of the intersection center attribute is realized,and a more generalized forward iterative growth algorithm is designed in the description of the road branch direction attribute.Finally,by fusing the intersection structured template information of satellite images,the association of intersection detection from local information to global information is completed.The algorithm has obtained good experimental results under the KITTI dataset,which is conducive to improving the positioning accuracy of UGV in intersection scenarios.2.A road area estimation method based on roadside fitting is proposed,and a surface matching strategy between point cloud ground information and satellite image road information is constructed.The method is firstly based on static binary Bayesian filtering and multi-frame information fusion,which further improves the robustness of ground segmentation results.Then,by combining the smooth feature,normal vector information and other information of the point cloud,the candidate position points of the road edge are preliminarily extracted.After that,based on the designed double boundary detection model,the secondary purification of road edge candidate points and the binary classification of boundary points are completed.In the fitting of road edge points,this paper adopts the fitting method based on random sampling consistency(RANSAC)and the least squares method to complete the extraction and smoothing of the road edge.Finally,the road area is estimated based on the roadside extraction results,which improves the matching efficiency of the point cloud ground points and the satellite image road information.3.A multi-threaded location algorithm based on satellite imagery and real-time point cloud is presented.The method starts with the establishment of a face match between the starting point cloud ground segmentation point or road estimation area and the fusion of satellite image road information,a point match between the point cloud junction detection information and the geometric information fusion of the satellite image junction,and a refined match matching between the point cloud reflectance normalized mutual information(Normalized Mutual Information,NMI)and the fusion of the grey scale information of the satellite image using three matching localization strategies for the multilayer mapping of the structured hierarchy and semantic hierarchy.Then,to address to some extent the possible matching degradation phenomenon along the longitudinal direction of the road,in this paper,we design a method to fit the matching weights based on a two-dimensional Gaussian distribution,and implement an online update to the error covariance in the localization filter.The experiment shows that the proposed method shows strong location robustness under both structured and unstructured scenarios.