Turn-Aware LSTM model for vehicle trajectory forecasting

Accurate trajectory prediction is essential for autonomous driving safety at intersections. Existing deep learning models often overlook turning behaviors, leading to curvature misestimation. This study proposes a Turn-Aware LSTM network that explicitly encodes maneuver intentions—left, right, or straight—using vehicle trajectories extracted from UAV footage via YOLOv8 and DeepSORT. To mitigate tracking noise, a cumulative turning-angle strategy is introduced for robust maneuver classification. Experiments demonstrate that the proposed model significantly improves prediction accuracy for turning maneuvers, reducing Final Displacement Error (FDE) by 15–20% at a 3-second horizon compared to vanilla LSTM and physics-based baselines. The findings validate the integration of maneuver-aware encoding for enhanced intersection-level forecasting in real-time applications.