Smart Navigation: Enhancing Vehicle Trajectory Forecasting with Deep Learning

Methods for accurately predicting the trajectories of vehicles have always been sought after by Intelligent Transportation Systems (ITS) to improve travel efficiency and traffic management. Advances in deep learning and the availability of extensive trajectory datasets have made such predictions feasible with low error rates. The paper presents a vehicle trajectory prediction model by leveraging the power of stacked Long Short-Term Memory (LSTM) networks and a highly refined dataset. A novel approach for feature extraction and selection is introduced focusing on trajectory properties such as entropy, elongation and elongation angle. The most relevant predictors are identified through a combination of similarity measures like cosine similarity, Tversky index and Pearson correlation coefficient. The model demonstrates superior performance compared to existing methods achieving a low Mean Squared Error (MSE) of 0.004 and Mean Absolute Error (MAE) of 0.01. The model has potential applications in manual, hybrid and autonomous driving environments.