Lane-Change Intention Prediction with Physics-Informed AI
Research Paper#Autonomous Driving, Lane-Change Prediction, Deep Learning🔬 Research|Analyzed: Jan 3, 2026 16:50•
Published: Dec 30, 2025 08:36
•1 min read
•ArXivAnalysis
This paper addresses a critical challenge in autonomous driving: accurately predicting lane-change intentions. The proposed TPI-AI framework combines deep learning with physics-based features to improve prediction accuracy, especially in scenarios with class imbalance and across different highway environments. The use of a hybrid approach, incorporating both learned temporal representations and physics-informed features, is a key contribution. The evaluation on two large-scale datasets and the focus on practical prediction horizons (1-3 seconds) further strengthen the paper's relevance.
Key Takeaways
- •Proposes TPI-AI, a hybrid framework for lane-change intention prediction.
- •Combines deep temporal representations (Bi-LSTM) with physics-inspired features.
- •Addresses class imbalance and generalization across highway scenarios.
- •Achieves high macro-F1 scores on two large-scale datasets (highD and exiD).
- •Demonstrates improved performance compared to baseline models.
Reference / Citation
View Original"TPI-AI outperforms standalone LightGBM and Bi-LSTM baselines, achieving macro-F1 of 0.9562, 0.9124, 0.8345 on highD and 0.9247, 0.8197, 0.7605 on exiD at T = 1, 2, 3 s, respectively."