ReFeree: Radar-Based Lightweight and Robust Localization using Feature and Free space

TL;DR

This work tackles robust radar-based place recognition under adverse weather while respecting onboard computational limits. It introduces ReFeree, a lightweight global descriptor of size $42\\times 1$ that fuses feature detections with free-space information, enabling rotational invariance and a semi-metric 1-DoF heading to boot SLAM. The system comprises R-ReFeree for fast place recognition and A-ReFeree for heading estimation, integrated with KD-Tree search, Nano-GICP-based registration, and pose-graph optimization, achieving strong performance across single- and multi-session datasets, including extreme weather. The approach yields high recall and AUC with significantly lower descriptor size than prior methods, and runs at about 4–5 Hz on Nvidia Jetson Nano, demonstrating practical feasibility for onboard autonomous systems. Overall, ReFeree offers robust radar-based localization and SLAM with lightweight descriptors, opening pathways for cross-condition, real-time radar robotics applications.

Abstract

Place recognition plays an important role in achieving robust long-term autonomy. Real-world robots face a wide range of weather conditions (e.g. overcast, heavy rain, and snowing) and most sensors (i.e. camera, LiDAR) essentially functioning within or near-visible electromagnetic waves are sensitive to adverse weather conditions, making reliable localization difficult. In contrast, radar is gaining traction due to long electromagnetic waves, which are less affected by environmental changes and weather independence. In this work, we propose a radar-based lightweight and robust place recognition. We achieve rotational invariance and lightweight by selecting a one-dimensional ring-shaped description and robustness by mitigating the impact of false detection utilizing opposite noise characteristics between free space and feature. In addition, the initial heading can be estimated, which can assist in building a SLAM pipeline that combines odometry and registration, which takes into account onboard computing. The proposed method was tested for rigorous validation across various scenarios (i.e. single session, multi-session, and different weather conditions). In particular, we validate our descriptor achieving reliable place recognition performance through the results of extreme environments that lacked structural information such as an OORD dataset.

Figures (8)

• Figure 1: The proposed place recognition successfully identifies the correct loop in radar images abundant with multipath and speckle noise. In contrast, Radar SC kim2020mulran exhibits vulnerability to multipath, often leading to the identification of incorrect loops. Also, 42$\times$1 shape of our lightweight descriptor enables us to perform the SLAM with Nvidia Jetson Nano in the KAIST sequence of Mulran dataset kim2020mulran.
• Figure 2: Our method's pipeline. In the frontend, we perform feature extraction and generate descriptors: R-ReFeree for place recognition and A-ReFeree for initial heading estimation. In the backend, we obtain a radar map through place retrieval and pose graph optimization. The white square and blue squares on the radar image in the ReFeree module are range-wise block and angel-wise block respectively. And the white and blue squares that make up the range-wise block and angle-wise block represent feature and free space respectively, and the red square represents the farthest feature for a unit angle.
• Figure 3: Query point cloud (red) and candidate point cloud (green) at frame $1492^{nd}$ and $1041^{st}$ in DCC 01, respectively. (a) The point clouds without estimating the initial heading. (b) After transforming the source point cloud with the initial heading. (c) The visualization after optimization with ICP in the state of (b).
• Figure 4: Recall@1 obtained by shifting the columns of the query radar image in the Hydro sequence. The radar image in the Hydro sequence implies that a shift of one column is about 0.9 degrees.
• Figure 5: Top, middle, and bottom are Bellmouth, Maree, and Hydro's F1-Recall and PR curve in order from the left.