Research On Road Ellement Extraction Method Based On Vehicle LiDAR

With the growing rise of driverless technology,the real-time storage and analysis of high-precision maps is particularly important.Among them,road traffic elements such as road surface,traffic marking and typical poles are the most basic elements of precision map preservation and real-time analysis.Based on the traditional way to obtain road related elements is low efficiency,large security risks and difficult to efficiently store and analyze the actual time.As a relatively new way of data collection,on-board point cloud mobile scanning technology can obtain high-precision ground spatial information data and other road scene information in a large range and high efficiency in the process of carrier vehicle travel,which greatly promotes the future development of unmanned driving and intelligent transportation.Affected by the driving speed of the carrier vehicle,the scene point cloud density of the data collected by LiDAR(Light Detection and Ranging)is relatively sparse,which is easy to cause the lack of ground information in the process of extraction.Therefore,in order to fully extract the information of road elements and ground objects,this thesis uses the extraction cloud data of road field attractions based on vehicle LiDAR to explore the automatic classification method of road elements in complex scenarios.The main research contents are as follows:(1)For the complex and sparse road scene based on on-board LiDAR,the road surface is extracted by using the virtual scanning line progressive segmentation method.Filters according to the spatial characteristics of non-ground point clouds to reduce the difficulty of subsequent processing.Improve the simplicity of ground information acquisition and reduce the calculation complexity;use the scanner path information to organize the point cloud into a virtual scanning line system;use the road boundary points from the virtual scanning line to complete the automatic extraction of the road surface.(2)In view of the sparse and discrete point cloud data,and characteristic intensity noise points,it is difficult to extract traffic lines from the road surface,a traffic line extraction classification method based on image adaptive threshold segmentation is proposed.First,the point cloud data is used to generate geographic reference images to detect traffic marking;then retrieve the marking point cloud,segment the connected objects,and refine the traffic marking by Otsu threshold method.Finally,the marking point clouds such as road boundary lines and rectangular lines are classified according to the model matching to realize the vector quantification of line marking.(3)In view of the difficulty of extracting typical rods such as street trees and signs in tedious large traffic scenarios,an extraction and classification method for detecting typical rods based on structural hierarchy is adopted.Based on the object distribution features,the ground points are filtered,and Eurostyle is clustered to divide the ground points into independent objects.For the adjacent objects produced by rough clustering,the non-pole part and the Inertial Measurement Unit cutting method are used to analyze the roughness of the pole,and finally classify the pole by using the spatial distribution features.In this thesis,the most experimental data of about 457 meters long and 70 meters wide collected by vehicle LiDAR are used to verify the road surface extraction method based on virtual scanning line progressive segmentation,the traffic marking extraction classification method based on image adaptive threshold segmentation,and the typical rod extraction classification method based on structural hierarchical detection.Experimental results show that the recall of road surface extraction 98.50%,precision 98.84%,actually 98.67%,recall of traffic marking extraction 89.45%,precision 94.97%,actually 92.13%;recall of typical rod extraction 97.17%,precision 95.23% and actually 96.19%.In conclusion,the three methods can efficiently extract road surface,traffic lines and typical rods from relatively large and complex field attraction cloud data.