Narrowing your FOV with SOLiD: Spatially Organized and Lightweight Global Descriptor for FOV-constrained LiDAR Place Recognition

TL;DR

This work tackles LiDAR-based place recognition under restricted FOV, where traditional descriptors struggle due to vacancy regions and computational constraints. It introduces SOLiD, a spatially organized, lightweight descriptor built from REC and AEC in range-elevation and azimuth-elevation bins, enhanced with an elevation-importance vector (IEV) to reweight features. The method yields R-SOLiD for loop detection and A-SOLiD for initial heading estimation, enabling real-time, onboard-capable recognition across single, multi-session, and multi-robot scenarios, with strong rotation robustness and low memory footprint. The results show SOLiD outperforms several baselines in accuracy and speed, while maintaining significantly smaller descriptor sizes, and the authors provide open-source code to facilitate adoption and integration with SLAM pipelines.

Abstract

We often encounter limited FOV situations due to various factors such as sensor fusion or sensor mount in real-world robot navigation. However, the limited FOV interrupts the generation of descriptions and impacts place recognition adversely. Therefore, we suffer from correcting accumulated drift errors in a consistent map using LiDAR-based place recognition with limited FOV. Thus, in this paper, we propose a robust LiDAR-based place recognition method for handling narrow FOV scenarios. The proposed method establishes spatial organization based on the range-elevation bin and azimuth-elevation bin to represent places. In addition, we achieve a robust place description through reweighting based on vertical direction information. Based on these representations, our method enables addressing rotational changes and determining the initial heading. Additionally, we designed a lightweight and fast approach for the robot's onboard autonomy. For rigorous validation, the proposed method was tested across various LiDAR place recognition scenarios (i.e., single-session, multi-session, and multi-robot scenarios). To the best of our knowledge, we report the first method to cope with the restricted FOV. Our place description and SLAM codes will be released. Also, the supplementary materials of our descriptor are available at \texttt{\url{https://sites.google.com/view/lidar-solid}}.

Figures (9)

• Figure 1: At the top is our loop matching between the 1631st (red) and 192nd (yellow) frame in the KITTI 00 datasets clipped with a 60$^\circ$ FOV. The loop matching exhibits slight differences in rotation. Our methodology (green) implies a low dominance range, however, Scan Context (SC) (red) kim2018scan signifies vacant information in the description. As the FOV narrows, it becomes more sensitive to subtle changes in small (purple) areas, directly correlating with success or failure. At the bottom are 3 phi charts that consist of recall@1, auc score, f1 max score, time, and descriptor size in KITTI 00 datasets with 60$^\circ$, 120$^\circ$, and 180$^\circ$ FOV. Recall@1, auc score, and f1 max score are higher values, the better performance. The higher values of speed and lightweight represent the low computational cost and the lightweight descriptor. We compare our method with well-known methods in LiDAR PR rohling2015fasthe2016m2dpkim2018scanvidanapathirana2022logg3dma2022overlaptransformerxu2023ring++ where discussed in Section II-A.
• Figure 2: Overall pipeline of our algorithm. Angled and rounded rectangles represent data and algorithms.
• Figure 3: Simple example schema of our methodology. The indices $i$, $j$, $k$ are determined along the $r$, $\theta$, $\phi$ axes. Utilizing the spatial organization, 3-D bins are projected onto REC and AEC according to their respective indices. Subsequently, R-SOLiD and A-SOLiD are generated by taking the matrix product with IEV.
• Figure 4: The process of generating R-SOLiD and A-SOLiD. First, when a 3-D scan is received, REC and AEC are generated through spatial organization. Second, summation is performed in the radial direction, followed by the creation of EC. Third, min-max normalization is applied to the created EC, resulting in the generation of IEV. Finally, the matrix product of IEV with REC and AEC respectively yields R-SOLiD and A-SOLiD.
• Figure 5: 500th frame from various angles in KITTI 00. Black represents full view (360$^\circ$) points. Blue, green, yellow, and red mean occlusion, 180$^\circ$, 120$^\circ$, and 60$^\circ$ points, respectively.