LiDAR-Based Localization in Tunnel From HD Map Matching With Pavement Marking Likelihood

Vehicle localization in tunnel is still a challenging task due to unavailable global positioning system (GPS) signal and degenerated structures, i.e., highly similar structures. This article proposes a novel high-definition (HD) map matching method based on point-to-likelihood association. The likelihood is generated with kernel density estimation (KDE) from pavement marking points by LiDAR intensity-based segmentation, called pavement marking likelihood (PML). The HD map matching results are then fused in a particle filter framework for vehicle localization, where the likelihoods consist of two parts. One is directly from the odometry localization results. The other is from the PML by applying point-to-likelihood association of HD map. By solving the maximum a posteriori probability (MAP) problem within the particle filter framework, it is ready to simultaneously localize the vehicle and detect the pavement markings. The proposed method has been validated with three tunnel sections from the open KAIST dataset. The experimental results demonstrate that the proposed method by fusing LiDAR, HD map, and other odometry sensors can achieve robust and accurate localization on each tunnel section. The translation and yaw angle errors are not more than 20 cm and 0.30° in three tunnel sections.