Vehicle trajectory-based prediction of traffic conflicts on sharp horizontal curves

The traffic conflict situations at sharp curve sections are evaluated by analyzing vehicle trajectory data during navigation through these hazardous road segments.

This study develops a methodology for quantifying traffic conflict probabilities in curve scenarios based on multi-source trajectory data acquisition. Vehicle movement trajectories through curves are captured via integrated UAV aerial photography systems and onboard vehicle recorders. High-precision spatiotemporal coordinates with dynamic parameters (instantaneous velocity and acceleration) are extracted using the professional trajectory analysis software. To address noise interference in raw trajectory data, a Kalman filtering algorithm is implemented for optimal motion state estimation and data smoothing. At the model architecture level, we propose a CNN-LSTM hybrid predictive model that synergistic-ally combines the spatial-temporal feature extraction capabilities of convolutional neural networks with the temporal dependency modeling advantages of long short-term memory networks, enabling end-to-end learning for quantitative trajectory conflict prediction. To validate model generalizability, this study concurrently constructed multiple benchmark models-including Support Vector Machine (SVM), Gradient Boosted Trees (XGBoost), GNN-LSTM, Vanilla LSTM, and Bi-LSTM-for comparative experiments. The evaluation framework employed a rigorous multi-dimensional validation protocol from machine learning, assessing all models not only by fundamental classification accuracy but also through fine-grained efficacy metrics (Precision, Recall, F1-score). Results demonstrated the superior performance of the hybrid CNN-LSTM model in predicting traffic conflicts at curves. Ultimately, curve-specific conflict probabilities were derived by applying the CNN-LSTM model to experimental data analysis. The generalization performance under class-imbalanced conditions was quantified using the Area Under the Receiver Operating Characteristic Curve (AUC-ROC), while prediction accuracy was validated through metrics including classification accuracy. This establishes a comprehensive multi-capability evaluation framework covering model stability, sensitivity, and generalization capability.

Empirical results confirm the CNN-LSTM model's superior performance in sharp-curve conflict prediction, achieving a mean accuracy exceeding 85%, precision above 82.7%, recall over 89.9%, and F1-score surpassing 86.1%, complemented by a 93.5% or higher average AUC-ROC that demonstrates robust generalization in class-imbalanced scenarios. These metrics collectively substantiate its exceptional spatiotemporal feature extraction capability and precise risk evolution pattern fitting, enabling enhanced representation of interactive vehicle conflicts in complex environments.

The research outcomes provide intelligent decision support for geometric optimization design of sharp curve sections and establish a reliable theoretical foundation for developing real-time dynamic risk warning systems. This work holds significant practical value for advancing the transformation of intelligent transportation management systems toward data-driven paradigms.