A two-stage approach for road marking extraction and modeling using MLS point clouds

Road markings are of great significance to road inventory management, intelligent transportation systems, high-definition maps (HD Maps), and autonomous driving. Most existing methods focus on extracting and classifying the road markings from mobile laser scanning (MLS) point clouds. Nevertheless, the performance suffers from the wear and incompleteness of road markings. Converting the extracted road marking points into a consistent representation with a sparse set of parameters needs extensive study as well. This paper presents a two-stage coarse-to-fine object detection and localization approach for automatically extracting and modeling road markings from mobile laser scanning (MLS) point clouds, which is robust to variations in reflective intensity, various point density, and partial occlusion. The first step is to use a general object detection network to detect bounding boxes with semantic labels of road markings on feature maps, which consists of information about intensity, elevation, and distance to the scanner. Next, accurate positions, orientations, and scales of candidate road markings are determined in the raw point clouds coordinate system through a shape matching operator that leverages the standard geometric structure and radiometric appearance of road markings. Finally, a re-ranking operator combining the coarse detection confidence and fine localization score is used to acquire the final road marking models. Comprehensive experiments revealed that the proposed method achieved an overall performance of 92.3% in recall and 95.1% in precision for extracting 12 types of road markings from urban scene point cloud datasets, even with worn and incomplete road markings. The modeling performance was 0.504 using the mIoU metric.