Self-Supervised Point Cloud Importance Awareness Network for 2-D LiDAR SLAM

Abstract

Data understanding of light detection and ranging (LiDAR) is essential for achieving embodied intelligence in mobile robots. However, the lack of semantic differentiation in 2-D point clouds hinders the potential capabilities of simultaneous localization and mapping (SLAM). To address this challenge, we propose a novel self-supervised learning approach for adaptive point cloud importance awareness. Firstly, we design an effective and lightweight awareness network that assign importance weights to highlight crucial points based on their inherent features and correlations. Secondly, we introduce a self-supervised triplet training strategy with multiple objective losses for efficient optimization, eliminating the dependence on manual annotations. Finally, we embed the importance-aware capability into the point cloud registration for SLAM enhancement. Extensive experiments demonstrate that our work achieves rational distribution of weighted point clouds, and also yields improvements in both accuracy and robustness across fundamental SLAM components including scan matching, mapping, and localization. The proposed approach effectively unlocks the latent potential of laser data, leading to superior performance for 2-D LiDAR SLAM systems.