Recognizing overtaking intentions on two-lane bidirectional highways considering multi-scenario coupling: An HHO-optimized dual-attention Bi-LSTM approach

Four typical overtaking scenarios were constructed using a high-fidelity driving simulator to collect high-frequency (50 Hz) trajectory data. A 10-dimensional input framework was established, encompassing ego-vehicle dynamics, inter-vehicle interaction risks, and scenario-coupled features. A four-state adaptive state machine was used for behavioral label calibration, and trajectories were truncated at the initial centerline crossing so that only normal car-following and overtaking-preparation samples were retained for binary intention recognition.

A Bi-LSTM network extracted bidirectional temporal dependencies, while feature and temporal attention mechanisms captured critical features and time windows of intention onset. The Harris Hawks Optimization (HHO) algorithm was used for global hyperparameter optimization, reducing reliance on manual tuning. Results: Integrating scenario-coupled features improved recognition performance across all candidate models, with SVM and Transformer accuracies increasing by 1.53 and 0.30 percentage points, respectively.

The proposed HHO-DualAtt-BiLSTM achieved an accuracy of 90.54% ± 0.56% and a macro F1-score of 90.10% ± 0.59%, outperforming the baseline and ablation models. At the 2.0s pre-warning horizon before centerline crossing, scenario-coupled features improved prediction accuracy by 2.85–5.00 percentage points, with the proposed model reaching 85.00%. Misclassifications were mainly concentrated in transitional micro-adjustments between car-following and early preparation states.

Conclusions: Incorporating joint scenario-coupled context improves early discrimination between normal car-following and overtaking preparation, while HHO reduces reliance on manual hyperparameter tuning and supports lateral active-safety pre-warning applications on two-lane bidirectional highways.</p>