Vehicle trajectory prediction based on attention optimized with real-scene sampling

Advancements in autonomous vehicles and deep learning have notably improved vehicle trajectory prediction accuracy. However, extracting interaction features in complex driving scenarios, such as vehicle-to-vehicle interactions and lane constraints, presents challenges. Deep learning-based methods struggle to achieve optimal predictive performance under limited computational resources. This study introduces a global attention mechanism to enhance feature extraction from driving scene encodings, focusing the decoder on interactive behaviours and boosting long-term prediction performance. An adaptive scheduled sampling model is employed, using actual driving scenarios probabilistically for training, addressing slow learning of actual driving behaviours and lack of initial feature correction. This method increases attention to actual interactions, reducing reliance on natural scenes and improving model generalizability. On the NGSIM dataset, sampling attention encoder-decoder (SAED) achieves a 1–5 s average displacement error (ADE) of 1.34 m, with 4 s and 5 s final displacement errors (FDEs) of 1.64 and 2.06 m, respectively. Compared to methods based on long short-term memory (LSTM), SAED reduces the model's storage space by 24.68% under the same network layer count. That demonstrates its effectiveness in extracting interactive behaviours in complex scenarios and enhances the accuracy of long-term predictions.