AdWeatherNet: Adverse Weather Denoising with Point Cloud Spatiotemporal Attention

Haozheng Han, Dongyu Du, Jie Luo, Xin Jin

Abstract: Adverse weather introduces disruptive noise into LiDAR data within autonomous driving systems, compromising the accuracy and range of 3D perception. Mitigating this challenge for high-precision noise removal becomes intricate due to the varying noise distributions at different distances. A novel spatiotemporal denoising network, AdWeatherNet, is proposed to address this problem. The Spatial Encoder module dynamically encodes spatial features using a designed density evaluation model. Additionally, the Temporal Differential Attention module effectively leverages temporal variation in adjacent point clouds to identify and accurately remove noise. To drive the research, we also introduce an adverse weather dataset, named the AdScenes dataset, which features point-wise annotations and a wide variety of weather conditions, making it one of the largest comprehensive datasets in this domain. The experimental results demonstrate the effectiveness of our method, with a remarkable improvement of +9.5% of IoU in rainy scenes, +5.9% of IoU in snowy scenes, and +1.9% of IoU in foggy scenes. Compared to the SOTA, AdWeatherNet enhances the mAP of object detection by an average of +1.8% across all weather conditions. Our method contributes to the development of reliable LiDAR perception systems, fostering the development of autonomous vehicles.