SCOPE: Spatial Context-Aware Pointcloud Encoder for Denoising Under the Adverse Weather Conditions

摘要

Reliable LiDAR point clouds are essential for perception in robotics and autonomous driving. However, adverse weather conditions introduce substantial noise that significantly degrades perception performance. To tackle this challenge, we first introduce a novel, point-wise annotated dataset of over 800 scenes, created by collecting and comparing point clouds from real-world adverse and clear weather conditions. Building upon this comprehensive dataset, we propose the Spatial Context-Aware Point Cloud Encoder Network (SCOPE), a deep learning framework that identifies noise by effectively learning spatial relationships from sparse point clouds. SCOPE partitions the input into voxels and utilizes a Voxel Spatial Feature Extractor with contrastive learning to distinguish weather-induced noise from structural points. Experimental results validate SCOPE’s effectiveness, achieving high Intersection-over-Union (mIoU) scores in snow (88.66%), rain (92.33%), and fog (88.77%), with a mean mIoU of 89.92%. These consistent results across diverse scenarios confirm the robustness and practical effectiveness of our method in challenging environments.