A General Approach for Simulating Rain Effects on Sensor Data in Real and Virtual Environments

Abstract

Driving automation systems typically use surround sensors to perceive their local environment. Accidents with automated vehicles have shown that errors in sensor data measurement and interpretation can lead to fatal injuries. It is thus necessary to test the reliability of environmental perception systems before market introduction. Since critical weather conditions are random, rare, and change quickly, relevant data sets are biased towards clear conditions. Consequently, detection algorithms based on these data suffer from limited performance. This article focuses on a two-step approach based on a) an indoor rain facility to test under reproducible, realistic conditions and b) physical-based models to enrich sensor data from clear conditions with virtual rain effects in a post-processing step. We concentrate on data from camera, lidar, and radar sensors. Experimental results show that both approaches simulate critical effects on raw sensor data and, therefore, enable replicable testing and validation at every stage of development.