SG-LPR: Semantic-Guided LiDAR-Based Place Recognition

Place recognition plays a crucial role in tasks such as loop closure detection and re-localization in robotic navigation. As a high-level representation within scenes, semantics enables models to effectively distinguish geometrically similar places, therefore enhancing their robustness to environmental changes. Unlike most existing semantic-based LiDAR place recognition (LPR) methods that adopt a multi-stage and relatively segregated data-processing and storage pipeline, we propose a novel end-to-end LPR model guided by semantic information—SG-LPR. This model introduces a semantic segmentation auxiliary task to guide the model in autonomously capturing high-level semantic information from the scene, implicitly integrating these features into the main LPR task, thus providing a unified framework of “segmentation-while-describing” and avoiding additional intermediate data-processing and storage steps. Moreover, the semantic segmentation auxiliary task operates only during model training, therefore not adding any time overhead during the testing phase. The model also combines the advantages of Swin Transformer and U-Net to address the shortcomings of current semantic-based LPR methods in capturing global contextual information and extracting fine-grained features. Extensive experiments conducted on multiple sequences from the KITTI and NCLT datasets validate the effectiveness, robustness, and generalization ability of our proposed method. Our approach achieves notable performance improvements over state-of-the-art methods.