DEGRADATION EVALUATION FOR AUTONOMOUS DRIVING IN RAINY WEATHER

Abstract

Autonomous Driving (AD) in adverse weather is an open challenge. One of the most significant challenges is sensor performance degradation. Most existing methods attempt to address it through degradation compensation. Degradation compensation may alleviate some degradation, but the remaining degradation can still affect the system's stability. In this work, we work on another approach: degradation evaluation, which aims to understand the amount of degradation in the AD system.

To evaluate degradation, we build our own dataset, SMART-Rain, to specifically study rainy weather conditions. With this dataset, we create three novel degradation tasks: LiDAR scan degradation quantification, laser missing measurement detection, and laser measurement uncertainty estimation. Experiment results show that our degradation estimation is consistent with the natural degradation process in the real environment. Next, we demonstrate using uncertainty estimation to improve perception and localization performance. Finally, we summarize this thesis and conclude with promising future research directions.