ReFeree: Radar-based efficient global descriptor using a Feature and Free space for Place Recognition

TL;DR

The paper addresses radar-based place recognition under adverse weather by proposing ReFeree, a lightweight global descriptor built from a feature-rich radar image and free-space information. It employs a three-step pipeline: (1) feature extraction from polar radar images, (2) aggregation of free-space density, and (3) construction of an $oldsymbol{\alpha}$-dimensional descriptor $\mathbf{K}$ via range-wise blocks, with a retrieval strategy based on a KD-tree using both descriptor and translational distances. The key contributions are the efficient ReFeree descriptor, its semi-metric properties, and extensive single- and multi-session validation showing up to 497× fewer descriptor bytes while achieving superior or competitive recognition performance compared with RadarSC, Raplace, and RadVLAD. The approach enables robust place recognition with low computational cost, making it attractive for real-time radar localization and map merging in challenging weather and environments.

Abstract

Radar is highlighted for robust sensing capabilities in adverse weather conditions (e.g. dense fog, heavy rain, or snowfall). In addition, Radar can cover wide areas and penetrate small particles. Despite these advantages, Radar-based place recognition remains in the early stages compared to other sensors due to its unique characteristics such as low resolution, and significant noise. In this paper, we propose a Radarbased place recognition utilizing a descriptor called ReFeree using a feature and free space. Unlike traditional methods, we overwhelmingly summarize the Radar image. Despite being lightweight, it contains semi-metric information and is also outstanding from the perspective of place recognition performance. For concrete validation, we test a single session from the MulRan dataset and a multi-session from the Oxford Offroad Radar, Oxford Radar RobotCar, and the Boreas dataset.

Figures (4)

• Figure 1: Top is The bottom is the Area Under the ROC Curve (AUC) score per descriptor's weight in single-session (S) kim2020mulran and multi-session datasets (M) kim2020mulranbarnes2020oxfordburnett2023boreas. We compare the proposed method with Radar Scan Context (RadarSC) kim2020mulran, Raplacejang2023raplace and Open-RadVLAD (RadVLAD) gadd2024open used in Radar.
• Figure 2: Our proposed pipeline. The navy rectangle represents data, and the blue rectangle represents the algorithm. After acquiring Radar polar images from the sensor, we first proceed with feature extraction and counting free space to obtain the final descriptor called ReFeree. Second, we put the proposed descriptor in the database. Finally, we build a KD-Tree to recognize revisited places through fast searching.
• Figure 3: Matching graph where F1 score is highest and a Precision-Recall (PR) curve in single session datasets. In the matching graph, the green color represents true loops located within 20m.
• Figure 4: Matching graph where F1 score is highest and a Precision-Recall (PR) curve in multi-session datasets. In the matching graph, the green color represents true loops located within 20m.