Robust Object Detection Under Snowfall Weather Condition from LiDAR Point Cloud

Robust 3D object detection is a pivotal task in numerous applications, including autonomous driving. However, adverse weather conditions have the potential to significantly impact the efficacy of LiDAR in collecting point clouds, thereby posing challenges to the safety of autonomous driving. In this paper, we undertake a comprehensive analysis of the impact of snowfall conditions on LiDAR imaging and the nature of interference to the object. We then propose a Snow Point Feature Filter (SPFF) block which mitigates the interference from noise points by performing feature-level filtering at the point feature aggregation stage. Our block can be seamlessly integrate into any voxels feature aggregation module. Empirical evidence from experimental trials substantiates the effectiveness of our methodology in curtailing the impact of noise points in voxel features during the feature aggregation stage, thereby enhancing the reliability of subsequent feature extraction. The efficacy of our proposed method is substantiated by a comparative analysis with prevailing mainstream detection methods on the CADC dataset.