IR-MFGL: Image-Represented Magnetic Field Global Localization in Repetitive Environments

Global localization is an essential ingredient for autonomous mobile robots. However, existing global localization systems primarily rely on Global Navigation Satellite System (GNSS), infrastructures, or visual/LiDAR-based place recognition, which suffer from enclosed/semi-enclosed GNSS-denied environments, offline and costly installation and calibration, or texture/geometry repetitive scenarios, respectively. To address the aforementioned challenges, this article proposes an infrastructure-free global localization system, which can achieve simultaneous place recognition and global pose estimation in GNSS-denied repetitive environments using ambient Magnetic Field (MF) produced by inherent ferromagnetic objects around the robot. The key idea of the proposed method is to represent the ambient MF in an image formulation, and transform the MF global localization problem into a cascading place recognition and pose estimation problem from the perspective of MF images. An end-to-end descriptor extraction network is designed to achieve place recognition, eliminating the inefficient sampling and searching used in conventional MF-based global localization systems. Moreover, a robust estimator is derived by introducing the graduated non-convexity optimization algorithm into global pose estimation, enabling the determination of the global minimum of the global pose on $\text{SE}(3)$ without requiring any prior knowledge. To validate the effectiveness of the proposed Image-Represented MF Global Localization method, called IR-MFGL, extensive benchmarks are conducted in real-world GNSS-denied repetitive environments, and the results show that the proposed IR-MFGL outperforms State-of-the-Art (SOTA) global localization methods.