An LSTM-Based Lane-Change Classification Inference Method Based on Coupling Relationship of Drivers’ Car-Following and Lane-Change Behaviors

To provide personalized lane-change (LC) assistance to drivers with different driving styles, leveraging the intertwined relationship between car-following (CF) and LC behaviors, an LSTM-based LC classification inference (LCCI) method is implemented in this study. Specifically, the minimum acceptance gaps and the maximum rejection gaps during the LC maneuvers are obtained first. Through the k-means clustering algorithm, the drivers’ LC behaviors are then classified into three categories: aggressive, moderate, and cautious. The LSTM-based LCCI model with a three-layer neural network is then constructed, using the subject vehicle and its leading vehicles’ speed, acceleration, and longitudinal distance during CF behaviors as inputs and the drivers’ LC categories as outputs. The model was trained using the NGSIM data set, achieving an accuracy of 89.7% on the test data set, and the drivers’ LC categories prediction accuracy is 92.4%. The experimental results demonstrate the effectiveness of inferring the LC categories using CF behavior data.