Cost-effective camera localization aided by prior point clouds maps for level 3 autonomous driving vehicles

Abstract

For navigation tasks, particularly within autonomous driving systems, accurate and robust localization is the critical aspect. While global navigation satellite systems (GNSS) are a widespread choice for localization, it has exhibited drawbacks such as susceptibility to issues like multipath and non-line-of-sight reception. Vision-based localization offers an alternative by relying on visual cues, circumventing the use of GNSS signals. In this study, we proposed a visual localization method aided by a prior 3D LiDAR map. Our approach involves reconstructing image features into multiple sets of 3D points using a localized bundle adjustment-based visual odometry system. Subsequently, these reconstructed 3D points are aligned with the prior 3D point cloud map, enabling the tracking of the user's global pose. The proposed visual localization methodology boasts several advantages. Firstly, the aided prior maps contribute to improving the robustness in the face of variations in ambient lighting and appearance. Additionally, it capitalizes on the prior 3D map to confer viewpoint invariance. The key idea of point cloud registration for the proposed approach determines geometric matching to establish the accurate position and orientation of a camera within its surroundings. This is achieved by contrasting the geometric features present in the camera's image with those stored in a reference map. The method identifies and aligns the geometric points between the camera image and the prior 3D point cloud map. Notably, our method is also conducive to the utilization of cost-effective and lightweight camera sensors by end-users. The experiment results show the proposed methods are accurate and frame rates without the need for supplementary information.