Integrating Expert Knowledge and Traffic Data for Lane-Changing Intention Prediction in Autonomous Vehicles

Abstract

Accurate vehicle intention prediction is critical for autonomous driving safety in complex traffic environments. To address the interpretability limitations of data-driven methods while maintaining high accuracy, this letter proposes a knowledge-data co-learning framework featuring: (1) a knowledge-guided spatio-temporal attention network (K-STAN) that integrates expert rules with attention mechanisms for explainable feature extraction, and (2) an enhanced explanation generator producing human-intuitive textual rationales. Our K-STAN architecture combines spatio-temporal attention with domain knowledge to model social interactions under the knowledge guidance, while the explanation framework bridges the gap between model decisions and human understanding. Experiment results demonstrate that accuracy is improved by 0.79%–6.41% over baseline methods without expert knowledge, with the generator successfully explaining 97.48% of lane-changing and 98.3% of lane-keeping scenarios. The proposed approach provides valuable insights that can enhance the development of advanced decision-making systems in autonomous vehicles.