LIVW-Localization: A Multimodal Information Fused Vehicle Localization Method for Complex, Large-Scale, and GNSS-Denied Environments

Accurate and robust vehicle localization is a key issue in autonomous driving. However, traditional vehicle localization methods usually have issues such as excessive reliance on global navigation satellite system (GNSS), low localization accuracy, and poor robustness to sensor degradation environments. To solve the above problems, we propose a map-based vehicle localization method that fuses LiDAR, vision, inertial measurement unit (IMU), and wheel encoder, named LIVW-Localization. To fully utilize the complementary advantages of LiDAR and vision, we use a prebuilt LiDAR map and real-time visual patches to construct a local visual map. To extract more stable and high-quality visual feature points, we propose a method to reuse the nearest neighbor points retrieved from real-time LiDAR scan in a point cloud map as the initial selecting criteria for visual feature points. Moreover, we propose an efficient method to fuse a single wheel encoder to improve the vehicle localization accuracy. To validate the performance of the proposed method, qualitative and quantitative experiments are conducted in various complex scenes, such as tunnels, long corridor-like outdoor environments, and highways. The experimental results show that our method outperforms the state-of-the-art methods, and the root-mean-square errors (RMSEs) of translation and rotation are 0.395 m and 0.007 rad, respectively, in the 19.376-km highway LHW-2 dataset.