EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Multi-Vehicle Interaction and State-Aware Temporal-Enhanced Graph Attention Network for Lane-Changing Intention Prediction

Yonghai He, Huajie Wan, Changchang Li, Yanli Jiao, Mingmin Han

IEEE Access · 2026 · 14 · 89612-89642

DOI: 10.1109/access.2026.3699732 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Yonghai He und 4 weiteren Autor:innen. Die Arbeit nennt 35 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Early and reliable recognition of lane changing intention is essential for improving road safety and sustaining efficient freeway operations in intelligent transportation systems. This study develops a state aware processing pipeline for high resolution vehicle trajectories and proposes a Temporal enhanced Graph Attention Network (TE-GAT) for early lane changing intention prediction. Using the real-world drone-based highD freeway dataset as the primary data source, this study first extracts nine longitudinal, lateral, and interaction-related microscopic variables. After entropy-weight analysis and Pearson-correlation screening, six representative variables are retained for traffic-state clustering, leading to three interpretable traffic states: stable free flow, high-speed free flow, and transitional/congested flow. These traffic-state descriptors are then fused as global contextual features in TE-GAT, which augments interaction-aware graph attention with temporal positional encoding and lightweight one-dimensional convolutions. To address class imbalance in lane-changing events, the Synthetic Minority Oversampling Technique is applied only to the training set after the train-test split. The proposed framework is evaluated against representative baselines, including LSTM, GCN, and vanilla GAT-based models, across multiple modeling horizons. Results show that a modeling window of approximately 1.0 s before lane-change onset provides the best empirical trade-off between anticipation time and classification robustness. Quantitative evaluations demonstrate that the proposed framework achieves a peak accuracy of 97.80% at this optimal 1.0 s prediction horizon with traffic-state fusion. The model also obtains F1 scores above 98% for lane-changing maneuvers, outperforming the baseline methods under the same experimental setting. By enabling earlier and smoother responses of advanced driver assistance and connected automated driving systems, the proposed method can help reduce critical lane changing conflicts, mitigate stop and go oscillations, and support safer and more sustainable freeway traffic operations.

Zitieren

APA

He, Y., Wan, H., Li, C., Jiao, Y., & Han, M. (2026). Multi-Vehicle Interaction and State-Aware Temporal-Enhanced Graph Attention Network for Lane-Changing Intention Prediction. IEEE Access, 14, 89612-89642. https://doi.org/10.1109/access.2026.3699732

Vancouver

He Y, Wan H, Li C, Jiao Y, Han M. Multi-Vehicle Interaction and State-Aware Temporal-Enhanced Graph Attention Network for Lane-Changing Intention Prediction. IEEE Access. 2026;14:89612-89642. doi:10.1109/access.2026.3699732

BibTeX
@article{He2026,
  author = {He, Yonghai and Wan, Huajie and Li, Changchang and Jiao, Yanli and Han, Mingmin},
  title = {Multi-Vehicle Interaction and State-Aware Temporal-Enhanced Graph Attention Network for Lane-Changing Intention Prediction},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {89612--89642},
  doi = {10.1109/access.2026.3699732}
}
Kostenlos starten →