Automated Road-Elements Modelling and Centerline Generation for High Definition Maps Utilizing 3D Point Cloud

Abstract

Recently High Definition Maps (HD Maps) become additional aiding information for autonomous vehicles to improve road safety and are under rapid development Compared with the conventional digital navigation maps the accuracy requirement is quite higher as well as HD Maps record more detailed road information and rich road elements including road edges lane lines centerlines traffic signs and traffic lights Owing to the efficiency and highly accurate geospatial measurements the commercial Mobile Laser Scanner (MLS) is competent to data acquisition of HD Maps generation However the subsequent process of maps generation relies on manpower Therefore this thesis is contributed to proposing an automated approach to generate road edge lane line and centerline defined in HD Maps from point clouds The proposed method can be divided into road elements extraction and modelling Firstly the road edges are obtained to determine the range of the road surface On the other hand the road markings are extracted based on intensity filtering Subsequently the road markings are classified by the pre-defined geometric thresholds Next the extracted road edges and lane lines can be modelled based on the cubic spline interpolation algorithm Therefore centerlines are generated pass through translating from the modelling lane lines Finally the results are assessed by the verified HD Maps provided by the professional surveying company to guarantee the absolute precision and quality of generated results Moreover this thesis also implements the proposed approach on the point cloud collected by the low-cost sensors to validate the flexibility and usability Although the quality and the performance of low-cost point cloud cannot be as well as MLS point cloud the data acquisition method on the basis of using low-cost payload can be an alternative way if the accuracy can be proved to achieve the accuracy requirement On the other hand it is not practical to process entire experimental areas at the same time to generate HD Maps since the attribute of the same road might change and the data volume and the computing operation need to be considered In fact the road will be segmented into a certain length in advance and then merging the modelling results with two consecutive road sections Thus this thesis also evaluates the proposed method on the road scenario of road segments From the results exhibited in this thesis the overall recall of road markings extraction is relatively unstable that results in insignificant performance However most road markings can be classified into correct categories and most precisions are more than 90% Additionally the experimental results illustrate that the modelling road elements and centerlines can be successfully generated The RMSEs of modelling lane lines and lane lines in both commercial MLS point cloud and low-cost point cloud are lower than 0 2 m in 3D space while the RMSEs of modelling road edges in MLS point cloud are also lower than 0 2 m in 3D space In conclusion this thesis not only identifies the road markings but also models the road markings to generate reliable and precise road elements in HD Maps for autonomous driving applications

Date of Award	2020
Original language	English
Supervisor	Kai-Wei Chiang (Supervisor)