Octrees based Adaptive Radius Outlier Removal for point clouds in harsh weather condition

Abstract

The LiDAR (Light Detection and Ranging) sensor is used in numerous applications such as SLAM (Simultaneous Localization And Mapping) and object detection. Especially, in autonomous driving, the LiDAR takes an important role in object detection. The LiDAR sensor emits light and measures the time of flight (ToF) until the emitted light returns. However, in harsh weather conditions, for example, snow or rainy day, since the light is blocked by those environmental elements, the performance of LiDAR will be degraded. Those environmental elements considered as noise and they make the accuracy of object detection lower. To overcome this problem, a proper noise filtering method is required. However, because of the sparsity and angular resolution of the LiDAR sensor, the conventional filtering methods such as Statistical Outlier Removal (SOR) showed poor performance. Thus, a new filtering method considering the LiDAR scanning mechanism and the property of the LiDAR point cloud is required. In this work, we focused to overcome this problem and eliminate environmental noises. We properly revised the Radius Outlier Removal (ROR) filtering method and show better performance than conventional methods. We divided point clouds into octrees to reduce computation costs. Also, depending on the condition of the weather, we focus on adaptively eliminate noises.