Toward Realistic Autonomous Driving Dataset Augmentation: A Real-Virtual Fusion Approach with Inconsistency...

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Autonomous driving systems require extensive datasets for reliable object recognition.
• Acquiring real-world data for rare or hazardous scenarios is costly and dangerous.
• Purely synthetic data often exhibits a reality gap due to visual and physical discrepancies.

Purpose of the Study:

• To propose a novel real-virtual fusion framework for generating realistic augmented image datasets for autonomous driving.
• To address the limitations of real-world data acquisition and the reality gap in synthetic data.
• To enhance the generalization capabilities of autonomous driving perception models.

Main Methods:

• Leveraging real-world driving data from K-City, South Korea.
• Synchronizing real data with a digital twin environment (Morai Sim) using a look-up table and fine-tuned localization.
• Injecting diverse virtual objects into real image backgrounds with inconsistency mitigation techniques like illumination matching.

Main Results:

• The real-virtual fusion strategy effectively bridges the reality gap between synthetic and real data.
• The framework provides a cost-effective and safe method for dataset augmentation.
• Experimental results demonstrate improved generalization for perception models.

Conclusions:

• The proposed real-virtual fusion framework offers a viable solution for creating high-fidelity datasets for autonomous driving.
• This approach enhances the robustness and safety of autonomous driving systems.
• The method contributes to overcoming data acquisition challenges in autonomous driving research.