LiDAR-to-Radar Translation Based on Voxel Feature Extraction Module for Radar Data Augmentation

Area of Science:

• Computer Vision
• Robotics
• Sensor Fusion

Background:

• Autonomous vehicles rely on LiDAR and radar sensors for object distance measurement.
• Deep learning for LiDAR is advanced, but radar sensor deep learning lags.
• LiDAR offers precision but is weather-vulnerable; radar is robust but noisy.

Purpose of the Study:

• To address noise and data augmentation challenges in radar sensors for autonomous driving.
• To propose a novel LiDAR-to-Radar data translation method.

Main Methods:

• Developed a LiDAR-to-Radar translation technique.
• Incorporated a voxel feature extraction module.
• Leveraged the point-based data acquisition similarity between LiDAR and radar.

Main Results:

• Successfully translated high-quality LiDAR data into radar data.
• Demonstrated the method's superiority using real-world, co-located LiDAR and radar data.
• Showcased enhanced radar data quality and augmentation capabilities.

Conclusions:

• The proposed LiDAR-to-Radar translation method effectively reduces noise in radar data.
• This approach significantly enhances radar data augmentation for autonomous vehicle perception.
• The method offers a viable solution for improving radar sensor performance in diverse environmental conditions.