AdverseNet: A Unified LiDAR Point Cloud Denoising Network for Autonomous Driving in Adverse Weather

In the field of autonomous driving, a pressing issue is how to enable LiDAR to accurately perceive the 3D environment around the vehicle without being affected by rain, snow, and fog. Specifically, rain, snow, and fog can be present within the LiDAR's detection range and create noise points. To address this problem, we propose a unified denoising network, AdverseNet, for adverse weather point clouds, which is capable of removing noise points caused by rain, snow, and fog from LiDAR point clouds. In AdverseNet, we adopt the Cylindrical Tri-Perspective View (CTPV) representation for point clouds and employ a two-stage training strategy. In the first training stage, generic features of rain, snow, and fog noise points are learned. In the second training stage, specific weather features are learned. We conducted comparative experiments on the DENSE dataset and the SnowyKITTI dataset, and the results show that the performance of our method on both datasets is significantly improved compared to other methods, with the Mean Intersection-over-Union (MIoU) reaching 94.67% and 99.33%, respectively. Our proposed AdverseNet enhances the LiDAR sensing capability in rain, snow, and fog, ensuring the safe operation of autonomous vehicles in adverse weather conditions. The source code is available at https://github.com/Naclzno/AdverseNet.