Bayesian inference of visibility in fog and smoke artificial conditions from 3D-LiDAR point clouds

3D-LiDARs are heavily impacted by a degraded visual environment (DVE) like rain, fog and smoke which limits their use for perception algorithms. The capacity to retrieve information about the environmental conditions from an embedded sensor can be an asset to improve autonomous driving performances. False positive artifacts in the point clouds caused by aerosols and hydrometeors particles tend to cause perception issues and thus need filtered out. However, those artifacts can also be used as valuable information to infer weather properties and maybe improve filters. This article proposes a Bayesian inference model which can classify discrete values of visibility using 3D-LiDAR point clouds. Gamma and Log-normal distributions are used to model the distance distributions of the noise points and the model is extended using the Random Finite Set (RFS) formalism with the Poisson and Binomial RFS models. Experiments in artificial fog and smoke conditions are presented and the classification model is trained and tested independently for each experiment. The used point clouds are extracted from specific parts of the field-of-view that can be used to generalize the proposed method to any outdoor scenario. The model shows good classification results with increased performances when the RFS extension is used.