LiDAR Point-Wise Segmentation-Based De-Noising for Autonomous Driving Perception System in Fog Weather
Jong‐Min Lee · IEEE Access · 2026
LiDAR sensors provide precise range measurements that complement cameras, radar, and ultrasonic sensors. However, adverse weather conditions such as rain, fog, and snow can severely degrade LiDAR returns and reduce perception reliability. This paper proposes a point-wise semantic segmentation–based de-noising method to suppress fog-induced artifacts in LiDAR point clouds. The proposed network directly operates on 3D points and incorporates feature aggregation to preserve local geometric structure while progressively expanding the receptive field, thereby reducing information loss and misclassification. In particular, the sampling and aggregation stages are designed to exploit the geometric characteristics of fog-affected point clouds and the distinctive intensity patterns of fog returns. The proposed method achieves real-time on-vehicle inference while maintaining robust point-wise segmentation accuracy in real fog. It is validated on three datasets, including controlled-environment chamber data, fog-model–based data, and real-driving dense-fog point clouds collected by our research team. Under an identical hardware/software platform, the per-frame runtime is reported as 60 ms, 45 ms, and 52 ms across the three validation datasets, respectively. The results demonstrate that the proposed model effectively segments and de-noises fog-degraded point clouds, improving perception robustness for autonomous driving in adverse weather and supporting safer operation under degraded visibility.