Enhanced RF-XGBoost for lane change intention recognition in autonomous driving

Accurate lane change intention recognition is critical for improving road safety and advancing autonomous driving systems. Traditional methods often struggle with high-dimensional feature spaces and dynamic interactions in complex traffic scenarios. This paper proposes a dual-stage ensemble learning framework, RF-XGBoost, to address these challenges. First, a robust feature engineering process extracts 1,036 features from vehicle kinematics, static attributes, and environmental interactions. Random Forest (RF) is employed to select 803 high-impact features, reducing redundancy and enhancing model efficiency. Subsequently, an XGBoost classifier, trained on time-series sliding window data, achieves precise recognition of left/right lane changes and lane-keeping intentions. Evaluated on the NGSIM US101 dataset, the proposed framework achieves 97% accuracy, outperforming SVM (61.1%), AdaBoost (90.6%), and standalone XGBoost (94.4%). The integration of feature selection and dynamic temporal modeling demonstrates significant improvements in robustness and interpretability, offering a reliable solution for real-time intention prediction in intelligent transportation systems.