A De-Stationary and Cross-Attention LSTM Model for Highway Vehicle Trajectory Prediction

Abstract

Accurate vehicle trajectory prediction is important for autonomous driving, especially in highway scenarios with high-speed motion and complex vehicle interactions. To address these challenges, this paper proposes a de-stationary and cross-attention long short-term memory (DCA-LSTM) model for highway vehicle trajectory prediction. The model combines de-stationary temporal feature enhancement, multi-scale temporal convolution, and ego-centric spatial cross-attention to improve temporal modeling and interaction representation. The de-stationary temporal module is designed to improve the representation of motion sequences with changing statistical properties, while the multi-scale convolution structure helps capture both short-term fluctuations and long-term motion trends. In addition, the spatial cross-attention mechanism dynamically aggregates interaction information from neighboring vehicles according to their influence on the target vehicle. Experiments on the NGSIM US-101 and I-80 datasets show that the proposed method outperforms the compared baseline models across multiple prediction horizons and achieves lower RMSE, especially for 5 s prediction. The ablation and qualitative results further demonstrate the effectiveness of the proposed temporal enhancement and spatial interaction modeling modules. The results indicate that the proposed model is effective for long-horizon highway trajectory prediction in complex traffic environments.