MSNR: A Multi-Stage Noise Reduction Method on Point Cloud+

In autonomous driving, direct time-of-flight (d-ToF) light detection and ranging (LiDAR) is one of the most emerging techniques due to its high depth resolution and detection ability in specific environments. Tasks such as object recognition and scene segmentation use distance information in LiDAR point clouds as ground-truth to improve performance. However, in harsh measurement environments, the amount of noise in a LiDAR point cloud increases significantly. In this case, an algorithm using LiDAR data will be misleading, resulting in performance degradation. Although many noise filtering algorithms at point cloud-level were proposed, their effectiveness is limited, as traditional point clouds contain only distance information. In this paper, a method with multi-stage noise reduction (MSNR) is proposed. It takes point cloud+ as input. Comparing to the method using conventional point cloud as input, MSNR improves accuracy by 29.50 % under high background light.