Learning naturalistic driving environment with statistical realism.

Area of Science:

• Autonomous Vehicle Simulation
• Artificial Intelligence in Transportation
• Deep Learning for Behavior Modeling

Background:

• Achieving statistical realism in autonomous vehicle (AV) simulation, especially for rare safety-critical events, is challenging due to high-dimensional real-world driving data.
• Existing simulation methods struggle to accurately represent the distribution and patterns of safety-critical scenarios.

Purpose of the Study:

• To develop a deep learning framework, NeuralNDE, capable of learning multi-agent interaction behavior from vehicle trajectory data.
• To refine the generation of safety-critical events in AV simulation to match real-world frequencies and patterns.
• To enable statistically realistic simulations for AV development and testing.

Main Methods:

• Developed NeuralNDE, a deep learning framework utilizing vehicle trajectory data.
• Proposed a conflict critic model and a safety mapping network to enhance safety-critical event generation.
• Trained and validated the framework using urban driving environment data.

Main Results:

• NeuralNDE successfully reproduces accurate safety-critical driving statistics, including crash rates, types, and severity.
• The framework also achieves accurate distributions for normal driving behaviors like speed, distance, and yielding.
• Demonstrated statistical realism in simulating urban driving environments, particularly for safety-critical situations.

Conclusions:

• NeuralNDE represents a significant advancement in creating statistically realistic AV simulations.
• This framework can accurately model both normal and safety-critical driving scenarios, crucial for effective AV testing.
• This is the first simulation model to achieve statistical realism for safety-critical situations in AV environments.