Deep Learning-Based Vehicle Trajectory Prediction in the Internet of Vehicles with LSTM Models

Predicting vehicle trajectories utilizing deep learning techniques holds paramount significance in advancing the development of vehicular networks and the realization of autonomous driving. However, the complexity and diversity of real-world traffic environments present significant challenges for accurate vehicle trajectory forecasting. In this paper, we initially devise a vehicle trajectory forecasting model, grounded on an encoder-decoder LSTM framework. This model incorporates both the target vehicle's intrinsic parameters and its interaction data with the preceding vehicle as input. The XY-LSTM model anticipates the path of the focal vehicle, whereas the XV-LSTM model estimates the lateral positional values and the longitudinal instantaneous velocities of the same vehicle. Performance metrics such as Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Average Displacement Error (ADE) are calculated and analyzed. Additionally, comparative experiments are conducted with CTRA (Constant Turn Rate and Acceleration), Linear model, and R_LSTM models. The simulation outcomes showcase that the introduced XV-LSTM model surpasses alternative models in long-range trajectory forecasting, while maintaining a commendable performance in short-range predictions as well.