Evaluation and Deployment of LiDAR-based Place Recognition in Dense Forests

TL;DR

This paper analyzed the capabilities of four different LiDAR place recognition systems, both handcrafted and learning-based methods, using LiDAR data collected with a handheld device and legged robot within dense forest environments, and focused on evaluating localization where there is significant translational and orientation difference between corresponding LiDAR scan pairs.

Abstract

Many LiDAR place recognition systems have been developed and tested specifically for urban driving scenarios. Their performance in natural environments such as forests and woodlands have been studied less closely. In this paper, we analyzed the capabilities of four different LiDAR place recognition systems, both handcrafted and learning-based methods, using LiDAR data collected with a handheld device and legged robot within dense forest environments. In particular, we focused on evaluating localization where there is significant translational and orientation difference between corresponding LiDAR scan pairs. This is particularly important for forest survey systems where the sensor or robot does not follow a defined road or path. Extending our analysis we then incorporated the best performing approach, Logg3dNet, into a full 6-DoF pose estimation system -- introducing several verification layers for precise registration. We demonstrated the performance of our methods in three operational modes: online SLAM, offline multi-mission SLAM map merging, and relocalization into a prior map. We evaluated these modes using data captured in forests from three different countries, achieving 80% of correct loop closures candidates with baseline distances up to 5m, and 60% up to 10m. Video at: https://youtu.be/86l-oxjwmjY

Figures (10)

• Figure 1: Our place recognition pipeline: State estimation module (VILENS) provides odometry estimates using LiDAR-inertial measurements at 10 Hz and a pose graph SLAM is used to optimize poses after successful loop closure verification. The place recognition and verification server consists of three steps: 1. loop candidate proposals, 2. coarse registration, and 3. fine registration. At each step, proposed loop candidates go through verification checks (grey boxes) at the global descriptor-level, local feature consistency-level, and fine registration level, respectively. A verified loop candidate is integrated in the pose graph only if it passes these three steps.
• Figure 2: Our proposed cycle consistency check is general and applies to the online and offline multi-mission SLAM case, as well as relocalization tasks. We only need the relative transformation estimates and loop candidates between four nodes $n_i, n_j, n_k, n_l$ to verify the validity of a loop. Please refer to Sec. \ref{['subsubsec:coarse-registration']} for technical details.
• Figure 3: Pose graph formulation used for (a) online, and (b) offline multi-mission SLAM optimization. Each node $n_{i}$ has a 6DOF pose $\mathbf{x}_{i}$, which correspond to the main variables estimated on each case.
• Figure 4: Precision-recall curves obtained for Logg3dNet, STD, EgoNN and Scan Context in our four dense forest datasets. Only the top-1 candidate within 10m of the ground truth position is regarded as a true positive candidate.
• Figure 5: Heatmaps depicting descriptor distances for the Evo dataset. Yellow hues denote a high descriptor similarity between scans, whereas purple indicates low similarity. Patterns more closely resembling the ground-truth (top-left) indicate better descriptor performance. Logg3dNet descriptors show the most similar patterns, whereas ScanContext descriptors are the least discriminative among these models.