A dual-layer network architecture for long-term trajectory prediction in unstructured scenarios

Accurate trajectory prediction is essential for ensuring the safety and reliability of autonomous driving, particularly in unstructured environments. This study proposes a novel dual-layer neural network architecture for high-precision, long-term vehicle trajectory prediction. The architecture integrates ModeNet, which classifies high-level maneuver modes, with TrajNet, an attention-enhanced LSTM encoder-decoder that generates future trajectories using a hybrid teaching strategy. To ensure the physical plausibility and rule compliance of the predicted trajectories, ModeNet associates each maneuver mode with a corresponding signal temporal logic (STL) formula, which is subsequently used to guide TrajNet’s trajectory generation. We further construct an optimized trajectory dataset tailored to diverse unstructured scenarios. Extensive experimental results demonstrate that our approach significantly outperforms existing baselines in both prediction accuracy and motion consistency.