Online 6DoF Global Localisation in Forests using Semantically-Guided Re-Localisation and Cross-View Factor-Graph Optimisation

TL;DR

This work tackles global localisation of ground robots in GPS-denied forests, where canopy and long-range odometry drift hinder reliable navigation. It introduces FGLoc6D, a framework that combines semantically-guided re-localisation with a cross-view ground-aerial factor-graph to achieve online 6DoF pose estimation against an aerial LiDAR map, aided by a semantically-informed keypoint regression loss that anchors features to tree trunks. The main contributions are (1) online 6DoF geo-localisation in forests, (2) a semantically-guided keypoint regression loss to improve keypoint repeatability, (3) integration of ground-to-aerial unary factors in a factor-graph with LiDAR-inertial odometry, and (4) extensive forest experiments showing drift-free, robust localisation under dense canopies. The results demonstrate superior accuracy and robustness compared to baselines like MCLoc3D and LIO-SAM, with online performance suitable for real-time robotic navigation in challenging forest environments.

Abstract

This paper presents FGLoc6D, a novel approach for robust global localisation and online 6DoF pose estimation of ground robots in forest environments by leveraging deep semantically-guided re-localisation and cross-view factor graph optimisation. The proposed method addresses the challenges of aligning aerial and ground data for pose estimation, which is crucial for accurate point-to-point navigation in GPS-degraded environments. By integrating information from both perspectives into a factor graph framework, our approach effectively estimates the robot's global position and orientation. Additionally, we enhance the repeatability of deep-learned keypoints for metric localisation in forests by incorporating a semantically-guided regression loss. This loss encourages greater attention to wooden structures, e.g., tree trunks, which serve as stable and distinguishable features, thereby improving the consistency of keypoints and increasing the success rate of global registration, a process we refer to as re-localisation. The re-localisation module along with the factor-graph structure, populated by odometry and ground-to-aerial factors over time, allows global localisation under dense canopies. We validate the performance of our method through extensive experiments in three forest scenarios, demonstrating its global localisation capability and superiority over alternative state-of-the-art in terms of accuracy and robustness in these challenging environments. Experimental results show that our proposed method can achieve drift-free localisation with bounded positioning errors, ensuring reliable and safe robot navigation through dense forests.

Figures (7)

• Figure 1: Our method, FGLoc6D, incorporates a Semantically-Guided Re-Localisation method for cross-view place retrieval and robust keypoint registration. Following re-localisation, we employ a factor-graph optimisation framework that integrates LiDAR-inertial odometry and ground-to-aerial registration factors, obtaining accurate global 6DoF pose estimation within the aerial reference map.
• Figure 2: Block diagram of our proposed pipeline. FGLoc6D is composed of a semantically-guided deep module to re-localise the ground robot against an aerial map. Once re-localisation is done, the method estimates 6DoF robot poses using a fixed-lag smoothing factor graph structure. The factor graph is populated with a prior factor, odometry factors obtained from LiDAR-inertial odometry, and unary factors representing ground and aerial submap registrations.
• Figure 3: EgoNN keypoint degeneracy in ground and aerial submap pairs. Keypoints visualised on ground query (bottom, coloured) and aerial candidate (top, greyscale), colourised by RANSAC inliers (green) and outliers (red), with lines between correspondences. Circled regions highlight the superiority of our semantically-guided keypoints, improving consistency for cross-view re-localisation.
• Figure 4: Precision-Recall curves for our deep re-localisation module. We omit Forest I as both methods achieve $\sim$100% $F1_{\mathrm{max}}$.
• Figure 5: Trajectory comparison between our proposed localisation method FGLoc6D, ground truth (offline Wildcat SLAM ramezani2022wildcat with loop-closures), our previous approach MCLoc3D de2023air, and the LiDAR-inertial odometry LIO-SAM shan2020lio with loop-closure in Forest I (top) and Forest III (bottom).