Evaluating LiDAR Perception Algorithms for All-Weather Autonomy.

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• LiDAR is essential for autonomous driving tasks like navigation and obstacle avoidance.
• Adverse weather (snow, rain, fog) introduces noise into LiDAR data, impacting perception system reliability and safety.

Purpose of the Study:

• Investigate LiDAR perception algorithm limitations in adverse weather.
• Explore noise mitigation techniques for LiDAR data.
• Propose future research for all-weather autonomous driving.

Main Methods:

• Characterized LiDAR noise in snow, rain, and fog using real-world and synthetic datasets.
• Evaluated point cloud filtering methods for denoising (processing time, accuracy, limitations).
• Assessed the impact of weather and filtering on 3D object detection, localization, and SLAM.

Main Results:

• Point cloud filtering partially reduces noise but requires specific tuning.
• Adverse weather negatively impacts 3D object detection; dynamic filtering improves performance.
• Localization is robust in snowfall but fails in dense fog; SLAM performs well in snowfall but fails in fog.

Conclusions:

• LiDAR perception is challenged by adverse weather, necessitating adaptive filtering strategies.
• Fine-tuning filtering methods is critical for specific LiDAR sensors, scenarios, and weather types.
• Further research is needed to achieve robust all-weather LiDAR autonomy.