PASENet: Snowy Scene 3D Object Detection With Pillar‐Wise Attention and Semantic Enhancement

ABSTRACT LiDAR based 3D object detection plays an essential role in autonomous driving, yet snowy conditions degrade point cloud quality by introducing false returns and causing occlusion of objects, degrading the detection accuracy of existing 3D object detection algorithms. To overcome these challenges, we propose pillar‐wise attention and semantic enhancement network (PASENet), an end‐to‐end network specifically designed for snowy scene. PASENet first employs pillar‐wise attention during feature extraction, utilizing L2‐normalized feature aggregation to compute query‐key differences, thereby enhancing relevant object features while suppressing snow noise. Secondly, we introduce a semantic enhancement branch where a lightweight U‐Net predicts pixel‐wise foreground segmentation in bird's‐eye‐view map and a segmentation head generates feature maps according to classes, supervised by free‐of‐charge semantic segmentation labels. The resulting semantically enhanced features are fused with pillar features to boost discriminability. Extensive experiments on the real‐world snowy STF dataset demonstrate an average performance improvement of 1.3% across varying snowfall intensities, while maintaining robustness in clear weather with 2.2% gain on the KITTI validation set.