On Mapping in Multilayer Environments: A Robust Graph-SLAM Approach Using LIDAR Intensity and Elevation Data

Abstract

An accurate map generation in multilevel environments is very challenging especially in cities packed with longitudinal bridges. This is because of obstruction and deflection of satellite signals of GNSS/INS-RTK (GIR) systems by the complex road structures. Even though using SLAM technologies, capability and implementation are rarely discussed in the literature because of difficulties to determine true correspondences between vehicle positions. In addition, these environments impose to address the issue of maintaining consistency in the Absolute Coordinate System between layers in order to enable many applications such as accurate estimation of elevation errors and sharing precise traffic information between autonomous vehicles. Accordingly, this paper proposes a Graph SLAM framework to generate precise LIDAR maps regardless the road structure complexity. The essence is to conduct the entire process in the image domain instead of 3D point cloud domain such as loop-closure detection, relative position error compensation and edge calculation. Moreover, a new concept of virtual layer edges is introduced and investigated to efficiently express the layer consistency in the optimization process. A critical highway and multilevel course in Tokyo has been scanned to check the reliability of addressing the above issues. The experimental results have verified the robustness, scalability and reliability of the proposed GS framework to generate accurate, coherent and consistent maps and outperform expensive GIR systems.