TRIP-AV: Transformer-Powered Vehicle Prediction for Autonomous Vehicles

Accurate path prediction has become an integral part of autonomous vehicle (AV) navigation on complex and dynamic roads today. Existing approaches often face challenges in effectively capturing complex spatio-temporal dependencies and the latent intentions of drivers, which can significantly limit their predictive accuracy and reliability. This paper introduces TRIP-AV, an advanced Transformer-based model that relies on sophisticated attention mechanisms to comprehensively model historical vehicle trajectories and their interactions with neighboring vehicles. The model’s main innovation lies in its ability to implicitly infer driver intention through XGBoost, along with attention-driven analysis of historical movement patterns and real-time interaction data, thus enhancing contextual awareness and prediction accuracy. Validated on the NGSIM US-101 dataset, TRIP-AV achieves Root Mean Squared Error (RMSE) of 2.50 at a 5-second prediction horizon, representing a 39.5% reduction compared to the best-performing baseline.