Snow Removal for LiDAR Point Clouds With Spatio-Temporal Conditional Random Fields

Abstract

LiDAR sensors have been extensively used in numerous applications, including autonomous driving, owing to their ability to generate high-quality 3D point clouds. However, the sensor can be greatly affected in adverse weather conditions, resulting in noisy 3D points that impair LiDAR-based perception performance. This letter proposes a novel de-snowing formulation with Conditional Random Fields (CRF). The proposed approach first constructs the CRF based on k-nearest neighbors with the snow confidence derived from the physical priors of snow, such as intensity and distribution. Then, Iterated Conditional Modes (ICM) is applied for the confidence propagation from points with high certainty to nearby uncertain ones, thereby identifying the latter. Moreover, the method can be extended to fully utilize the temporal information of sequential scans for clearer snow removal. Extensive experiments on the real-scanned WADS dataset validate that our de-snowing approach significantly outperforms baselines and even the learning-based State-of-The-Art (SoTA) one. Furthermore, we demonstrate that our method has the potential to benefit the downstream 3D object detection task in snowy weather.