Localization and Offline Mapping of High-Voltage Substations in Rough Terrain Using a Ground Vehicle

TL;DR

We address autonomous navigation of a ground vehicle in rough terrain by enabling localization within a pre-built 3D map while concurrently updating the map via a hybrid SLAM-localization framework. The method extends FAST-LIO2 to localize against a prior map using a loaded $ikd$-tree and an ICP-based initial alignment, solving for the rigid transform $T = (\mathbf{R},\mathbf{t})$ that minimizes the residuals $\sum_i ||\mathbf{R} p_i + \mathbf{t} - \hat{q}_i||^2$ with $\mathbf{R}\in SO(d)$. A two-step map crafting pipeline—uniform voxel sampling and MLS smoothing—produces a de-noised, sharp point cloud; traversability is then estimated by CSF ground extraction and a Grid Map-based 2D costmap using normals, slope, and roughness. Field tests on HVSS data demonstrate that the proposed FAST-LIO-LOC approach achieves higher localization accuracy (mean error around $0.026\,$m) than ICP or LIORF-based methods while remaining computationally efficient, and the authors release open-source code for community use.

Abstract

This paper proposes an efficient hybrid localization framework for the autonomous navigation of an unmanned ground vehicle in uneven or rough terrain, as well as techniques for detailed processing of 3D point cloud data. The framework is an extended version of FAST-LIO2 algorithm aiming at robust localization in known point cloud maps using Lidar and inertial data. The system is based on a hybrid scheme which allows the robot to not only localize in a pre-built map, but concurrently perform simultaneous localization and mapping to explore unknown scenes, and build extended maps aligned with the existing map. Our framework has been developed for the task of autonomous ground inspection of high-voltage electrical substations residing in rough terrain. We present the application of our algorithm in field trials, using a pre-built map of the substation, but also analyze techniques that aim to isolate the ground and its traversable regions, to allow the robot to approach points of interest within the map and perform inspection tasks using visual and thermal data.

Figures (6)

• Figure 1: Generated Map of HVSS.
• Figure 2: Overview of the general pipeline of the system.
• Figure 3: Application of our method within a HVSS.
• Figure 4: Raw and filtered part of map (upper and lower correspondingly).
• Figure 5: (UP) Ground point cloud extracted from CSF. (DOWN) Grid map with overground structures.