Vehicle Trajectory Prediction Using Deep Learning for Advanced Driver Assistance Systems and Autonomous Vehicles

Vehicle trajectory prediction is essential for advanced driver assistance systems (ADAS) and autonomous vehicles (AVs), playing a crucial role in collision avoidance, path planning, and traffic control. Traditional models often overlook the variability in driver behavior, particularly in braking patterns, which significantly impacts trajectory predictions. Our study introduces an improved trajectory prediction model that uses Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks while considering driver braking patterns along with vehicle dynamic information. The models' accuracy is evaluated in a simulated environment that replicates real-world driving conditions. This environment captures comprehensive vehicle dynamics data, encompassing critical parameters such as position, rotation, acceleration, speed, and braking patterns. To ensure a realistic and varied dataset, data were meticulously gathered from 17 drivers, each utilizing a driving simulator equipped with the Euro Truck Simulator software. The model was implemented and validated using Python 3.9, Google Colab, and Scikit-learn, selected for their robustness in deep learning applications. Our results indicate that incorporating braking patterns significantly improves position predictions, outperforming the existing models based solely on vehicle dynamic data. This was evident by a notable decrease in Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) when braking patterns were incorporated. This advancement strengthens trajectory prediction systems for ADAS and AVs, enhancing operational safety.