Driving Behavioral Learning Leveraging Sensing Information from Innovation Hub

Di, Xuan; Jin, Peter; Huang, Yufei; Mo, Zhaobin · 2022 · ROSA P / Rutgers University. Center for Advanced Infrastructure and 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 research addresses the challenge of characterizing human driving behavior and assessing roadway safety in the context of connected and automated vehicle (CAV) deployment. The authors identify two primary gaps in existing literature: the difficulty of quantifying uncertainty in stochastic human car-following behaviors and the limitations of traditional safety assessments that rely solely on historical crash data rather than real-time driver and traffic conditions. To address these issues, the study develops physics-informed deep learning models for behavior prediction and a multi-source data framework for real-time safety index calculation. The methodology comprises two distinct components. First, the authors propose a "DoubleGAN" model to quantify uncertainty in car-following behavior. This physics-informed generative adversarial network integrates stochastic physics equations with neural networks, utilizing a moment-matching technique to accelerate convergence and prevent model collapse. The model was trained and validated using the Next Generation SIMulation (NGSIM) dataset, specifically focusing on US Highway 101 trajectories. Additionally, the authors tested a physics-informed LSTM variant to evaluate sequential prediction capabilities. Second, the study developed a system to predict driving safety indices by combining in-vehicle and roadside data. An Android application captured driver facial data, which was processed using the OpenFace toolkit to extract fatigue levels (via Eye Aspect Ratio), gaze direction, and emotional states. Simultaneously, roadside cameras captured vehicle trajectories to calculate traffic conflict indicators, including time to collision and deceleration rate to avoid a crash. These inputs were fed into a LightGBM model to predict near-miss events, which were then aggregated using fuzzy logic to generate risk scores. The results demonstrate that the DoubleGAN model effectively captures the uncertainty of real-world car-following data, with prediction distributions overlapping significantly with ground-truth data. The moment-matching technique proved superior to standard adversarial losses, enabling faster convergence during training. In comparative tests, the physics-informed LSTM model outperformed pure data-driven LSTM and ANN-based models when sufficient training data was available, though it required more data to achieve optimal performance. For the safety index prediction, the system successfully generated dynamic risk heat maps for road segments in the New Brunswick Smart Intersection Mobility Testbed. Confusion matrices indicated the model's ability to predict risk score levels for the next one to two seconds, allowing for the visualization of safety corridors. The significance of this work lies in its integration of human factors and real-time traffic data into safety assessment frameworks, moving beyond static historical crash analysis. The proposed DoubleGAN offers a robust method for uncertainty quantification in autonomous driving simulations, aiding in the development of human-aware motion planning. Furthermore, the digital twin environment built on the COSMOS testbed provides a scalable platform for validating these algorithms. These findings support the deployment of AVs by providing tools to predict driver behavior and roadway safety with greater precision, ultimately enhancing traffic mobility and safety management.

Key finding

The DoubleGAN model successfully quantified uncertainty in human car-following behavior with improved convergence, and the integrated multi-source data approach enabled the prediction of dynamic driving safety indices for road segments.

Methodology

mixed_methods

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.

• telematics crash prediction
• situational awareness
• in vehicle coaching
• exposure measurement
• distraction detection algorithms

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).

• Methodological Resource: tool software, dataset resource
• Theoretical Contribution: computational model