Modeling Naturalistic Driving Environment with High-Resolution Trajectory Data

Feng, Shuo; Yan, Xintao; Liu, Henry · 2023 · ROSA P / University of Michigan. Center for Connected and Automated Transportation

archive: archived pipeline: cataloged verified

Get this paper ↗ (full text — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of achieving statistical realism in autonomous vehicle (AV) simulation, specifically regarding the accurate reproduction of rare, safety-critical events. Current simulators often fail to model the high-dimensional, interactive nature of real-world driving environments, leading to distribution shifts and inaccurate safety metrics. The authors developed NeuralNDE, a deep learning framework designed to learn multi-agent interaction behaviors from high-resolution trajectory data, ensuring that simulated environments match real-world distributions for both normal driving and long-tail safety-critical scenarios. The methodology employs an imitation learning paradigm using a Transformer-based behavior modeling network to predict the joint distribution of future actions for multiple agents. To address the rarity of crashes and near-misses in training data, the framework integrates a conflict critic module and a safety mapping network. The conflict critic monitors generated trajectories for potential conflicts and applies calibrated acceptance probabilities to determine whether to allow dangerous behaviors or rectify them via the safety mapping network, which maps unsafe actions to feasible safe domains. This process ensures that the frequency and patterns of safety-critical events align with ground-truth statistics. The model was trained and validated using high-resolution trajectory data collected from roadside sensors at a multi-lane roundabout in Ann Arbor, Michigan. The results demonstrate that NeuralNDE achieves distribution-level accuracy for both normal and safety-critical driving statistics. The simulated environment accurately reproduced real-world distributions for vehicle instantaneous speed, distance, and yielding behaviors. Crucially, the model matched real-world crash rates, crash types, crash severity, and near-miss measurements, such as post-encroachment time. The fidelity of generated crash events was further validated against real-world crash videos and police reports. Additionally, the framework supports hour-level simulations, allowing for continuous interaction between AVs and background traffic without the distribution collapse common in other learning-based simulators. The significance of this work lies in providing a statistically realistic simulation environment for the development and testing of autonomous vehicles. By accurately modeling long-tail safety-critical events, NeuralNDE reduces the sim-to-real gap, preventing optimistic safety estimates that could mislead AV development. The framework is scalable and can be integrated with existing high-fidelity simulators like CARLA to provide realistic traffic environments. Beyond AV testing, the model offers a tool for estimating the safety performance of traffic facilities under various flow conditions, addressing a longstanding problem in transportation engineering regarding the statistical realism of microscopic driving behavior simulations.

Key finding

The NeuralNDE framework achieves statistical realism in simulated driving environments by accurately reproducing both normal driving behavior distributions and safety-critical event statistics, such as crash rates and types, when compared to real-world trajectory data.

Methodology

modeling

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via bulk_ingest_rosap on 2026-05-23 (6 acquisition events logged).

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• naturalistic crash near crash
• situational awareness
• telematics crash prediction

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: crash risk outcomes
• Methodological Resource: dataset resource
• Theoretical Contribution: computational model