Research on Multi-Level Trajectory Prediction Model for Personalized Driving Behavior

To enhance the accuracy of vehicle trajectory prediction and ensure driving safety, this paper proposes a Driving Style-Aware Trajectory prediction model based on a Two-Level Classification and Trajectory Module (DS-TCTM). The proposed framework comprises three key stages: data preprocessing, driving style recognition, and personalized trajectory prediction. Driving style features are extracted using the vehicle acceleration change rate and average headway, while traffic flow density is clustered via K-Means++ and subsequently fed into a Gaussian Mixture Model (K-GMM) guided by the K-Nearest Neighbor algorithm. Based on these analyses, driving behaviors are categorized into conservative, moderate, and aggressive styles. A multi-level personalized trajectory prediction network (TCTM) is then developed, in which trajectory data corresponding to each driving style is fed into separate sub-networks to enhance prediction accuracy. The proposed K-GMM model is compared with the traditional GMM in terms of clustering performance, and DS-TCTM is evaluated against several baseline models, including LSTM, Social-Lstm, Social-Velocity-Lstm, and Convolutional-Social-Lstm, focusing on prediction error and computational efficiency. Experimental results demonstrate that DS-TCTM achieves average Root Mean Square Error (RMSE) and Negative Log-Likelihood (NLL) values below 2.03 and 1.95, respectively, across various time horizons. After more than 100 iterations of hyperparameter optimization, DS-TCTM exhibits a 17.4% reduction in prediction error, with notably improved performance in long-term trajectory forecasting. These results indicate that DSTCTM effectively captures the influence of driving style on trajectory patterns, significantly improves predictive accuracy, and provides robust support for vehicle early warning systems.