Learning naturalistic driving environment with statistical realism

To enhance the effectiveness of simulation in autonomous vehicle development and testing, simulators must accurately reproduce real-world safety-critical scenarios. This paper introduces NeuralNDE, a deep learning framework that learns multi-agent interaction behavior from vehicle trajectory data. It uses a conflict critic model and a safety mapping network to refine the generation of safety-critical events, aligning with real-world frequencies and patterns. NeuralNDE achieves accurate safety-critical (crash rate, type, severity, near-misses) and normal driving statistics (speed, distance, yielding behavior). This is the first simulation model to reproduce real-world driving environments with statistical realism, especially for safety-critical situations.

Xintao Yan, Zhengxia Zou, Shuo Feng, Haojie Zhu, Haowei Sun, Henry X. Liu