Objective Road markings provide a wealth of guidance information to their drivers on daily basis, and they are also the fundamental of lane-level navigation maps used in advanced driver assistance systems and unmanned driving technologies. With the rapid development of autonomous driving technology, swift and accurate identification and extraction of road markings are essential. Traditional methods use total station measurement to obtain information about road marking. This method has low collection efficiency and high personnel costs, making it unsuitable for acquiring road marking information quickly. Many researchers have attempted to extract road markings from images, but the effect is highly dependent on image quality. Vehicle-mounted mobile laser scanning, as a high-tech surveying technology that has developed rapidly in recent years, can efficiently and accurately obtain three-dimensional spatial information and echo intensity information of roads and surrounding features, thereby providing a new method for obtaining road marking information. In this paper, combining the characteristics of the road structure and the intensity information of the point cloud echo, a road marking extraction algorithm based on the vehicle-mounted laser point cloud is proposed. Methods Firstly, the cloth simulation filtering (CSF) algorithm is used to obtain the ground point cloud from the original point cloud; secondly, the pavement point cloud is extracted using the normal vector-based area growth method; then, the inverse distance weighted interpolation method is used to project the pavement point cloud into an intensity feature image, dividing it into multiple subimages, using the largest interclass variance to divide the subimages into pure blocks and miscellaneous blocks. The pure block determines the segmentation threshold based on the relative size of the gray average value of the original image, while the miscellaneous block determines the segmentation threshold based on the Otsu algorithm. Finally, the road marking point cloud is obtained after threshold segmentation, morphological filtering and denoising, and point cloud back projection. Results and Discussions CSF algorithm and normal vector-based area growth method are used to generate the pavement point cloud. The only difference between pavement points and nonpavement points is the presence of curbs, indicating that the pavement point cloud is correctly extracted (Fig. 5 ). To test the performance of the dynamic threshold segmentation method proposed in this paper, it is compared to the maximum entropy threshold segmentation method, the block maximum entropy threshold segmentation method, and the adaptive threshold segmentation method. The comparison results show that the dynamic threshold segmentation method proposed in this paper can simultaneously solve the problem of uneven intensity distribution and the size of image blocks, the extracted road markings are more accurate and complete with less noise (Fig. 8 and Table 1 ). The test results on three sets of data show that the recall rate, accuracy, and comprehensive evaluation index of the road markings extracted by the method in this paper are all greater than 90%, and the effect of extracting the markings is better ( Fig. 9 and Table 2 ), and it has some advantages over methods proposed by Yu and Yao et al (Fig. 10 and Table 3 ). Conclusions This paper proposes a road marking extraction algorithm based on a vehicle-mounted laser point cloud. Based on the ground attributes of road markings, CSF algorithm is used to obtain ground point clouds, eliminating the influence of nonground point clouds on the extraction of markings; according to the difference of normal vectors on both sides of the road, the area growth method is used to segment and extract road points from the ground point cloud, eliminating the interference of curbs and ground objects on both sides of the road; to speed up the processing efficiency of the algorithm, the pavement point cloud is projected to generate intensity feature images; for the uneven intensity distribution in the extraction of markings, the intensity feature images are divided into multiple subimage to reduce the incident angle and range of the image. To avoid misclassifying subimages that are all roads or markings, the proposed algorithm use the largest interclass variance to classify the subimages and select different thresholds based on the subimage category. Dynamic threshold segmentation is used by the calculation method. Finally, three actual road sections?????????????????? point cloud data are chosen for experimental verification in this paper. The results show that the recall rate, accuracy rate, and comprehensive evaluation index of road marking extraction are 92.8%, 96.8%, and 94.8%, respectively, suggesting that the algorithm in this paper is effective and feasible, has some practical application value, and supports the production of high-precision maps in unmanned driving.
