A High-Precision Vehicle Trajectory Prediction Method based on Multi-Scale Features and Attention

Existing Long Short-Term Memory (LSTM)-based vehicle trajectory prediction methods often suffer from information forgetting when processing long-span trajectory data. To address these issues, this paper proposes a trajectory prediction framework that integrates a multi-scale convolution module and an LSTM encoder, along with a cross-attention mechanism, to enhance the modeling capability of both local dynamic features and global temporal relationships. Specifically, a multi-scale convolution module is introduced in the encoder to extract shortterm motion patterns of vehicle trajectories with varying receptive fields, compensating for LSTM’s limitations in local dynamic feature extraction. Simultaneously, the cross-attention mechanism adaptively adjusts the focus on different temporal information in the encoder’s output, enabling the model to fully leverage historical trajectory data. Comprehensive experiments are conducted on real-world road datasets; it outperforms existing mainstream methods in both short-term and long-term prediction tasks.