Direct vehicle warnings offer great potential toward improved safety by preventing or warning highway vehicles entering crossings at the time of train arrivals. However, this requires high ontime locational accuracy and timely communications. This project aims to work towards improved accuracy by developing an advanced, uncertainty-aware prediction model for grade crossing Signal Phase and Timing (SPaT) using train trajectory prediction. Accurate prediction of train arrival and departure times at highway-rail grade crossings remains challenging due to operational uncertainties such as varying speeds, dwell times, train length, and network interactions. Existing systems rely on Kalman filtering using real-time train location data from the head-of-train (HOT) and end-of-train (EOT), but these approaches have limited capability in
capturing nonlinear dynamics and uncertainty.
Building on prior work, the research team has already developed a Kalman-filter-based baseline model and collected train trajectory datasets from the Transportation Technology Center (TTC), Pueblo, CO. This project will leverage these datasets along with simulation data to develop and evaluate a hybrid probabilistic prediction framework.
The proposed framework integrates track-constrained state-space modeling with a Temporal Fusion Transformer (TFT) for multi-horizon forecasting. To explicitly address uncertainty, the model incorporates quantile prediction and conformal calibration to generate probabilistic predictions with calibrated confidence levels.
Rather than producing a single deterministic estimate, the system will generate multiple likely train trajectories and associated confidence levels. These predictions are used to derive SPaT at grade crossings. All evaluations will be conducted in a simulation environment using TTC datasets to assess prediction accuracy, robustness, and uncertainty calibration.
By leveraging existing datasets and focusing on simulation-based validation, this project reduces development risk while advancing beyond current methods. The outcomes will improve SPaT reliability, enhance situational awareness, and reduce safety risks and traffic congestion at highway-rail grade crossings.