GAN-Based LiDAR Translation between Sunny and Adverse Weather for Autonomous Driving and Driving Simulation

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Autonomous driving relies on accurate perception, with LiDAR being crucial alongside cameras.
• A significant data scarcity exists for deep learning algorithms utilizing LiDAR, hindering development.
• Existing data augmentation methods primarily focus on image-to-image translation, neglecting LiDAR data needs.

Purpose of the Study:

• To address the lack of LiDAR data for autonomous driving and simulation.
• To develop a Generative Adversarial Network (GAN)-based LiDAR translation algorithm.
• To enable realistic LiDAR data generation across various adverse weather conditions.

Main Methods:

• Proposed a novel GAN-based LiDAR translation algorithm specifically for autonomous driving applications.
• Developed an empirical approach to handle diverse weather conditions, including precipitation and varying visibility.
• Validated the algorithm using the JARI dataset (adverse weather) and the real-world Spain dataset.

Main Results:

• The proposed method successfully generated realistic LiDAR data under various adverse weather scenarios.
• Experimental results confirmed the algorithm's effectiveness in simulating challenging driving environments.
• The approach demonstrated its capability to produce high-fidelity LiDAR point clouds for diverse conditions.

Conclusions:

• The developed GAN-based LiDAR translation technique is effective for autonomous driving and simulation.
• This method provides a viable solution for augmenting LiDAR datasets, particularly in adverse weather.
• The algorithm contributes to more robust perception system development and verification for autonomous vehicles.