Navigating on adverse weather: enhancing LiDAR-based detection with the DBSPRY dataset

Abstract

Autonomous vehicles rely on camera sensors, LiDAR, and RADAR sensors to navigate their surroundings. However, adverse weather conditions like snow, rain, and fog present significant challenges for LiDAR-based perception systems. The lack of datasets representing these conditions hinders the development, testing, and enhancement of detection systems capable of handling adverse weather scenarios effectively. To tackle this issue, we introduce the DBSPRY dataset, which is specifically designed for scenarios involving spray kicked up from vehicles traveling on wet surfaces. This dataset contains multiple scenarios repeated in different settings, offering data on dry and wet road conditions, various speeds, types of vehicles, and different situations. Three different point cloud detectors were assessed using the ONCE and DBSPRY datasets. Additionally, we conducted an analysis of the impact of spray on LiDAR detections. Our findings indicate that incorporating our dataset led to a significant improvement in the detection accuracy of all models compared to the baseline, underscoring the importance of diverse training scenarios and objects. While all models performed well under optimal conditions, their accuracy decreased by approximately 5% in rainy and spray conditions. Among the models, CenterPoint demonstrated the best performance, maintaining high accuracy even in adverse weather, followed by PV-RCNN and PointPillar, both of which achieved over 80% accuracy.