Generation of Horizontally Curved Driving Lines for Autonomous Vehicles Using Mobile Laser Scanning Data

Abstract

The development of autonomous vehicle desiderates tremendous advances in three-dimensional (3D) high-definition roadmaps. These roadmaps are capable of providing 3D positioning information with 10-to-20 cm accuracy. With the assistance of 3D high-definition roadmaps, the intractable autonomous driving problem is transformed into a solvable localization issue. The Mobile Laser Scanning (MLS) systems can collect accurate, high-density 3D point clouds in road environments for generating 3D high-definition roadmaps. However, few studies have been concentrated on the driving line generation from 3D MLS point clouds for highly autonomous driving, particularly for accident-prone horizontal curves with the problems of ambiguous traffic situations and unclear visual clues. This thesis attempts to develop an effective method for semi-automated generation of horizontally curved driving lines using MLS data. The framework of research methodology proposed in this thesis consists of three steps, including road surface extraction, road marking extraction, and driving line generation. Firstly, the points covering road surface are extracted using curb-based road surface extraction algorithms depending on both the elevation and slope differences. Then, road markings are identified and extracted according to a sequence of algorithms consisting of geo-referenced intensity image generation, multi-threshold road marking extraction, and statistical outlier removal. Finally, the conditional Euclidean clustering algorithm is employed followed by the nonlinear least-squares curve-fitting algorithm for generating horizontally curved driving lines. A total of six test datasets obtained in Xiamen, China by a RIEGL VMX-450 system were used to evaluate the performance and efficiency of the proposed methodology. The experimental results demonstrate that the proposed road marking extraction algorithms can achieve 90.89% in recall, 93.04% in precision and 91.95% in F1-score, respectively. Moreover, the unmanned aerial vehicle (UAV) imagery with 4 cm was used for validation of the proposed driving line generation algorithms. The validation results demonstrate that the horizontally curved driving lines can be effectively generated within 15 cm-level localization accuracy using MLS point clouds. Finally, a comparative study was conducted both visually and quantitatively to indicate the accuracy and reliability of the generated driving lines.