REHEARSE-3D: A Multi-Modal Emulated Rain Dataset for 3D Point Cloud De-Raining.

Area of Science:

• Autonomous Driving Systems
• Sensor Technology
• Computer Vision

Background:

• Sensor degradation, particularly from raindrops, significantly impacts LiDAR point cloud quality in autonomous driving.
• Inaccurate measurements due to weather interference pose safety risks if systems lack weather awareness.

Purpose of the Study:

• Introduce REHEARSE-3D, a novel, large-scale, multi-modal emulated rain dataset for 3D point cloud de-raining research.
• Provide a valuable resource for advancing weather-aware autonomous driving technologies.

Main Methods:

• Developed REHEARSE-3D, featuring 9.2 billion point-wise annotations, high-resolution LiDAR (LiDAR-256), and 4D RADAR point clouds.
• Collected data under controlled weather, day/night conditions, including rain-characteristic information.
• Benchmarked raindrop detection and removal using fused LiDAR and 4D RADAR data.

Main Results:

• Evaluated various statistical and deep learning models for raindrop detection and removal.
• SalsaNext and 3D-OutDet models achieved over 94% Intersection over Union (IoU) for raindrop detection.

Conclusions:

• REHEARSE-3D is the largest point-wise annotated dataset for 3D point cloud de-raining.
• The dataset facilitates sensor noise modeling and point-level weather impact analysis.
• The findings demonstrate effective raindrop detection in fused sensor data, enhancing autonomous driving safety.