LAWA: LiDAR Adverse Weather Augmentation for Robust Point Cloud Semantic Segmentation

Abstract

This article presents a novel data augmentation approach aimed at enhancing light detection and ranging (LiDAR) point cloud semantic segmentation (PCSS) performance under adverse weather conditions, specifically focusing on rainfall and snowfall. The proposed augmentation approach takes into account the inherent characteristics of laser-based sensing of LiDAR, incorporating simulations for point intensity reduction, range noise, and LiDAR occlusion specific to adverse weather conditions. Moreover, the proposed simulation allows for precise control by varying the number of scattering points according to different levels of precipitation and introduces a method to realistically simulate noise from wet ground. The augmented dataset resulting from these simulation strategies is then utilized to assess the influence of adverse weather on various PCSS models and validate performance enhancements.