Research on adaptive threshold lane detection methods based on LiDAR

Abstract

As one component of the environment perception part for autonomous driving, lane marking detection is of paramount importance for its direct influence on autonomous vehicle’s safety. In the literature, the lane marking detection methods are camera-based and relatively mature. However, performance is affected by environmental lighting conditions. The LiDAR sensor collects high precision and three-dimensional environmental information without being affected by the ambient light, making the detection method based on LiDAR become the focus of research in recent years. Nevertheless, state-of-the-art detection methods mainly rely on 64-beam or 32-beam LiDAR, resulting in an extensive amount of data that makes it impossible to guarantee real-time performance. By considering time consumption and cost control, we proposed an adaptive threshold lane marking detection method based on a 16-beam LiDAR. The contributions of this dissertation are: Firstly, we proposed a curb points filtering algorithm based on segmented points density. A random sample consensus algorithm with constraints is used to extract the original road data, and most of the invalid background data are eliminated. Then according to the distance, road surface points are clustered to different scanlines. By analyzing the statistical characteristics of the spatial distribution of curb data within each scanline in a specific direction, a segmented point density method is proposed, and curb data is filtered out from the original road surface to optimize results of road data extraction further. Then, an adaptive threshold lane marking detection algorithm is proposed. We note that lane marking echo intensities are higher than the intensities of road data, and the echo intensity of lane marking data decreases with the increase of the distance. Therefore, we improve the Otsu algorithm by defining the threshold interval. The value of the threshold is determined based on the echo intensities of points in each scanline. Rather than affected by the echo intensity of adjacent scanline, which reflects the adaptive characteristics of thresholds, our method improves the efficiency and accuracy of threshold selection. Finally, we tested the proposed lane marking detection method by using five different datasets collected by the 16-beam LiDAR. Through the statistics and analysis of detection results, the adaptive threshold selection lane marking detection method based on a LiDAR sensor proposed in this paper can efficiently and accurately complete the function of lane marking detection, which is feasible and practical.