Centrala begrepp
A lightweight and robust radar-based global descriptor using feature and free space information enables reliable place recognition in diverse environments, including extreme conditions, and facilitates efficient SLAM with initial heading estimation.
Sammanfattning
The proposed method, called ReFeree, addresses the challenges of radar-based place recognition by leveraging both feature and free space information from radar images.
Key highlights:
Lightweight descriptor: ReFeree is at least 4x and up to 600x lighter than other methods, enabling efficient on-board processing.
Robustness to noise: By utilizing free space information, ReFeree mitigates the impact of multipath and speckle noise, achieving reliable place recognition even in extreme environments with sparse structural information.
Rotational invariance: ReFeree's range-wise block design provides rotational invariance, allowing it to recognize revisited places in both forward and reverse directions.
Initial heading estimation: ReFeree's angle-wise block design enables the estimation of the initial heading between revisited places, improving the efficiency and robustness of the registration process in the SLAM pipeline.
The method was extensively evaluated on various datasets, including Mulran, OORD, Oxford Radar RobotCar, and Boreas, covering diverse environments, weather conditions, and sensor configurations. ReFeree outperformed state-of-the-art radar-based place recognition methods across multiple metrics, including Recall@1, F1-score, and AUC. Additionally, the lightweight nature of ReFeree enabled its integration into a full SLAM pipeline, which was successfully tested on an NVIDIA Jetson Nano platform, demonstrating its suitability for on-board processing.
Statistik
The number of free space is significantly larger than the number of features in the radar images, with the DCC 01 sequence having 1,340,891 free spaces and 3,108 features, and the KAIST 03 sequence having 1,340,684 free spaces and 3,315 features.
Citat
"Unlike these methods, we propose a radar-based lightweight and robust global descriptor with a feature and free space called ReFeree by using the radar image in polar coordinates without a cartesian converting process."
"Also, the proposed descriptor that is at least 4× and up to 600× lighter compared to other methods and the KD-Tree searching process enhances usability on onboard computers."
"Unlike previous methods [11], our approach enables semi-metric localization by estimating the 1-DoF heading between the revisited place and the current place."