A Comparative Field Study of Global Pose Estimation Algorithms in Subterranean Environments

In this article, we perform a novel and extended field evaluation of the state-of-the-art algorithmic frameworks’ performance on global pose estimation. More specifically, we focus on relocalizing a mobile robot in a pre-built 3D point cloud map of a large subterranean environment. The evaluation is divided into two parts. The first part consists of multiple simulations performed in two different Gazebo SubT worlds, where one is flat with various types of features, while another has uneven structure and is more textured. The second part is an experimental evaluation and takes place in a real-world underground tunnel. In all evaluation tests, the robot’s pose is selected in such a way that we can test the robustness, as well as the feature extraction capability, of each algorithm. The evaluation is carried out using three ROS packages: a) hdl_global_localization using both BBS and FPFH+RANSAC, b) LIO-SAM_based_relocalization, and c) Fast-LIO-Localization. Our goal is to have a clear view of each algorithm’s efficiency in terms of CPU load, memory allocation and time to relocalize, while we increase the size of the map and transverse the robot in different parts of the map.