RGOR: De-noising of LiDAR point clouds with reflectance restoration in adverse weather

Abstract

Recently, LiDAR sensors have become indispensable in autonomous driving research. Despite continuous improvements in performance and price reductions, noise generated under adverse weather conditions remains a serious challenge. Most of the noise generated under such conditions is due to particles such as fog, rain, and snow. These particles are extremely fine; therefore, they have a very low reflectance compared to the targets that the laser should detect. In this study, we propose a method to distinguish particles by restoring the reflectance from LiDAR sensing data based on the reflectance characteristics of the particles. In addition, we propose a method to make additional judgments based on the geometrical shapes of adjacent particles to distinguish the particles more accurately. The proposed method is accurate enough to be compared to state-of-the-art deep learning methods. Moreover, the execution time is less than 2 ms on a single-core CPU, demonstrating a remarkable efficiency, being more than three times faster than that of methods performed on a GPU. Because noise removal is a preprocessing step, the proposed method is expected to allow more resources to be allocated to other, more important processes for autonomous driving.