Publications

This study proposes a CV data-based arterial traffic prediction framework with two-stage traffic-state extraction and an AASTGCN for arterial delay and queue predictions.

A Macro-Micro Cross-Attention Transformer (MMCAformer) model that integrates CV data-based micro driving behavior features with macro traffic features for freeway speed prediction.

A hybrid model for real-time secondary crash likelihood prediction that excludes post-crash features, using dynamic spatial-temporal windows and ensemble machine learning methods to achieve 91% identification accuracy with AUC of 0.952.

This study develops an ensemble deep learning model integrating MVDS and connected vehicle data for crash prediction, demonstrating superior performance over single-source models, with CV-only models achieving sensitivity exceeding 0.79 at 4% penetration or higher.

This study analyzes segment-level safety using lane-changing and driving volatility features from CV trajectories, revealing critical relationships between risky driving behaviors and crash types.

This study proposes a Multiscale Geographical Random Forest (MGRF) integrating street-view semantic visual features for traffic safety modeling by capturing multiscale spatial heterogeneity.

This study develops a grouped random-parameters Poisson-Lindley model with spatial effects to analyze intersection crashes with visual environment features and spatiotemporal variations.

This study proposes a spatial-ML framework integrating longitudinal and lateral risky driving behaviors from connected vehicle data for intersection crash analysis.

This study incorporates visual environment features from Google Street View images into a random parameter negative binomial-Lindley model for intersection crash frequency analysis.

This study proposes a joint cross-stitch network model to simultaneously evaluate tunnel crash severity and congestion duration, achieving improved performance over separate models by capturing coupling relationships between the two outcomes.

This study investigates the impact of speed on injury severity in single-vehicle run-off-road crashes using a partially temporal constrained modeling approach.

This study examines the association between helmet usage and injury severity in moped-vehicle crashes using partially temporally constrained random parameters bivariate probit models.

This study analyzes the impact of speed differences on joint injury severities in freeway rear-end crashes using joint statistical and machine learning models.

This study develops an LSTM + Transformer approach integrating traffic flow data from millimeter-wave radars and risky driving behavior data from connected vehicles, achieving 77.7% accuracy and 78.5% AUC for real-time crash risk evaluation.

This study proposes a Transformer-based model for freeway crash risk evaluation using abnormal driving events, achieving 84.1% accuracy and 77.7% AUC and outperforming traditional methods. It also reveals the temporal-spatial decay and collective effect of abnormal driving events.

This study investigates the impact of speed differences on rear-end crash severity using joint random parameters bivariate probit models, accounting for temporal instability during COVID-19.

This study introduces an unsupervised Angle-Based Outlier Detection approach with sliding window mechanism for real-time expressway crash risk identification, achieving 80.4% sensitivity and addressing data imbalance issues.

This study addresses robustness issues in deep learning-based crash risk evaluation models. By proposing a coverage-oriented adversarial training method, it generates traffic flow adversarial examples (TF-AEs) to prevent 76.5% accuracy drops and 98.9% sensitivity drops.

This study proposes a CNN-LSTM hybrid model for driving risk assessment using connected vehicle data. It takes time-series top views of CV driving environment as inputs, achieving an AUC of 0.997 by capturing spatial interactions and temporal variability.

This study investigates optimal spacing for millimeter-wave radar detectors on highways using a deep forest model. It achieves AUC of 0.849 with 80.9% recall, which is higher than CNN and RF models. It also finds that 750 m spacing balances cost and crash risk evaluation accuracy.

This study uses High-D realistic vehicle trajectory data to predict high-risk events on freeways using random-parameter logistic regression models, achieving 97% prediction accuracy 2 seconds ahead of the occurrence of events.