Simulating Realistic Rain, Snow, and Fog Variations For Comprehensive Performance Characterization of LiDAR Perception

Abstract

For robust object detection on LiDAR data, neural networks have to be trained on diverse datasets that contain many different environmental influences like rain, snow, or fog. To this date, few datasets, with those features, are available while there exist many datasets recorded under perfect weather conditions. Repurposing those datasets by simulating adverse environmental conditions on top of them and training networks with the resulting enhanced datasets, is intended to lead to more robust neural networks. In the following we propose models to realistically simulate the effects of rain, snow, and fog on LiDAR datasets based on physical and empirical fundamentals. Then we parameterize our simulation to best fit real LiDAR data that was captured in those environments, in order to achieve a highly accurate simulation. Finally, the impact of adverse weather on neural network detection performance is demonstrated.