On Training Deep-Learning Models for Removing Airborne-Particle Points From SPAD LiDAR Multiecho Point Clouds

A single-photon avalanche diode (SPAD) light detection and ranging (LiDAR) is susceptible to be disrupted by adverse operating conditions, such as fog, haze, or dust. Thereby, the ability of LiDAR-based scene understanding would be significantly affected. In the published literature works, efforts to eliminate these effects are either based on processing histograms or processing point clouds. In this letter, we study the task of removing airborne-particle points from SPAD LiDAR multiecho point clouds. To be more specific, the deep-learning approach is chosen due to its promising performance and expandability. We focus on the issue of training deep-learning models for the task, especially when only datasets from limited scenarios and different types of LiDAR are available. A training strategy mainly based on data preparation and dataset augmentation, together with tweaks made on the neural network model, is shown to be effective in achieving the training task.