Generating Rainy Scenarios With LiDAR for Autonomous Driving

Abstract

Rainy datasets play a critical role to train the neural networks for autonomous driving, and yet they are scarce when compared with datasets captured in the clean weather. Simulation works consisting of generating the rainy scenarios completely on engine and inserting raindrops into the real sunny scenarios have been proposed. However, those works cannot construct realistic scenarios in any rainy day owing to their limitations of relying on the empirical ingredients and not comprehensively modelling the random process. We therefore reconsider the whole procedure of the scenario generation by re-reducing the rainfall rate equation from the perspective of raindrops while taking the effect of wind into consideration, and utilizing a joint probability to model the raindrop generation in terms of eventual points in the cloud. The experimental results represented by Kullback-Leibler divergence and Bhattacharyya distance between generated and real scenarios from the viewpoints of raindrop ranges, point angles, and point intensities respectively show that those rainy scenarios are close enough, and thus generated scenarios can replace real captured counterparts.