Wachtel Granado, Diogo (Technische Hochschule Ingolstadt), Derder Trevisol, Heitor (The Federal University of Technology – Parana (UTFPR)), Rothmeier, Thomas (University of Applied Sciences Ingolstadt), Nassu, Bogdan Tomoyuki (Federal University of Technology - Parana), Werner Huber, Werner (Technische Hochschule Ingolstadt)

Navigating on Adverse Weather: Enhancing LiDAR-Based Detection with the DBSPRY Dataset

Scheduled for presentation during the Regular Session "LiDAR-based perception" (FrBT6), Friday, September 27, 2024, 14:10−14:30, Salon 14

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on December 26, 2024

Abstract

Autonomous vehicles depend on camera sensors, LiDAR, and RADAR sensors, among others, to navigate their surroundings. Adverse weather conditions, such as snow, rain, and fog, pose significant challenges for LiDAR-based perception systems. Unfortunately, the scarcity of datasets reflecting these conditions hinders the development, testing, and improvement of detection systems designed to handle such adverse weather scenarios effectively. To address this, we present the DBSPRY dataset, which is focused on scenarios containing spray whirled up from vehicles driving over wet surfaces. This dataset offers multiple scenarios repeated in different settings, providing data on dry and wet road conditions, various speeds, vehicles, and situations.

Three different Point Cloud detectors are evaluated across the ONCE and DBSPRY datasets. Furthermore, we present an analysis of the spray's impact on the LiDAR detections. We show that including our dataset significantly improved the detection accuracy of all models compared to the baseline, highlighting the importance of diverse training scenarios and objects. All models performed optimally under ideal conditions, but rain and spray reduced their accuracy by about 5%. Among the models, CenterPoint performed best, maintaining high accuracy even in adverse weather, followed by PV-RCNN and PointPillar, which still achieved over 80% accuracy.