This study focuses on vehicle trajectory prediction on highways and urban expressways with the aim of enhancing traffic safety. Leveraging the CitySim dataset, three critical scenarios are examined: basic highway segments, highway merge/diverge areas, and urban expressway weaving zones. For each scenario, Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Graph Neural Network (GNN), and Transformer models are employed to predict trajectories 5 s ahead using 5 s of historical data. To improve prediction performance, this study proposes key enhancements centered on kinematic feature optimization, culminating in the development of the Kinematic Feature Enhanced-Transformer (KF-Transformer) model. The first enhancement modifies the attention mechanism to a relative attention framework incorporating position bias terms. The second replaces the feedforward network with a convolution-enhanced variant that integrates depthwise separable and dilated convolutions. Comparative experimental results demonstrate that the KF-Transformer model achieves superior performance, with Root Mean Square Error (RMSE) values of 1.93, 2.07, and 2.33 m recorded for basic highway segments, highway merge/diverge areas, and urban expressway weaving zones, respectively. The proposed model exhibits promising potential for application in Advanced Driver-Assistance Systems (ADAS), enabling proactive risk anticipation and effective accident mitigation.