A Benchmark for Spray from Nearby Cutting Vehicles
This work addresses the need for reproducible benchmarking to improve driving safety in adverse weather conditions for autonomous vehicles, though it is incremental in focusing on a specific environmental factor.
The paper tackles the problem of autonomous driving perception degradation due to spray from nearby vehicles by introducing a novel testing methodology and setup, showing that distortions severely affect perception stacks for up to four seconds in common scenarios.
Current driver assistance systems and autonomous driving stacks are limited to well-defined environment conditions and geo fenced areas. To increase driving safety in adverse weather conditions, broadening the application spectrum of autonomous driving and driver assistance systems is necessary. In order to enable this development, reproducible benchmarking methods are required to quantify the expected distortions. In this publication, a testing methodology for disturbances from spray is presented. It introduces a novel lightweight and configurable spray setup alongside an evaluation scheme to assess the disturbances caused by spray. The analysis covers an automotive RGB camera and two different LiDAR systems, as well as downstream detection algorithms based on YOLOv3 and PV-RCNN. In a common scenario of a closely cutting vehicle, it is visible that the distortions are severely affecting the perception stack up to four seconds showing the necessity of benchmarking the influences of spray.