LiDAR point clouds segmentation in adverse weather conditions

Abstract

With the rapid advancement of 3D sensor technologies and the widespread use of Light Detection and Ranging (LiDAR) systems, point cloud data has become essential for describing real-world scenes. It is widely applied in various fields, including autonomous driving. Autonomous driving requires precise environmental perception, with point cloud segmentation being one of the core technologies. However, autonomous vehicles face various environmental conditions, such as rain, snow, and fog. These adverse weather conditions introduce disturbances into LiDAR data, resulting in significant challenges for point cloud segmentation. Existing point cloud segmentation algorithms often perform poorly under such conditions. To address this challenge, we propose an adverse weather segmentation network. In our network, the multiscale perception generation module includes components for multiscale feature extraction and multi-spatial pillar feature extraction, aimed at capturing multiscale spatial perceptual features and relationships. Additionally, a multiscale spatial fusion module integrates these features into the encoding stream, effectively enhancing feature representations. During data processing, we add instances to enrich the point cloud with coherent elements, ensuring environmental consistency. Experimental results show that our method achieves 64.1% mIoU under adverse weather conditions, demonstrating better performance than existing state-of-the-art approaches. These results highlight the robustness and effectiveness of the proposed method.