AADS: Augmented autonomous driving simulation using data-driven algorithms

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Current autonomous driving (AD) simulation relies on costly, time-consuming computer graphics (CG) models.
• CG images lack real-world authenticity, leading to performance degradation in trained AD systems.
• Existing virtual environments struggle to capture real-world complexity and diversity.

Purpose of the Study:

• To develop a novel simulation method for autonomous driving (AD) technologies.
• To enhance the realism and scalability of AD simulation environments.
• To create photorealistic, fully annotated training data for AD systems.

Main Methods:

• Augmented real-world images with simulated traffic flow using LiDAR and camera data.
• Generated plausible traffic flows for vehicles and pedestrians from acquired trajectory data.
• Composed simulated traffic into real-world backgrounds and resynthesized images from various viewpoints and sensor models.

Main Results:

• Developed an Augmented Autonomous Driving Simulation (AADS) producing photorealistic simulation images.
• Generated fully annotated images suitable for training and testing AD systems across perception and planning tasks.
• Validated the system's effectiveness on various AD tasks, including detection, segmentation, and prediction.

Conclusions:

• The AADS method offers superior scalability and realism compared to traditional CG-based simulations.
• This approach effectively combines virtual environment flexibility with real-world data richness for advanced AD simulation.
• The AADS system provides a robust solution for the development and validation of autonomous driving technologies.