Self-Interaction Dynamic Graph Convolutional Network With Multiscale Time-Frequency Fusion for Vehicle Trajectory Prediction

The ability of autonomous vehicles to accurately predict future trajectories is crucial for ensuring safe driving in complex traffic environments. However, current methods exhibit limitations in capturing the dynamic dependencies inherent in vehicle social interactions and lack effective integration of multiscale time-frequency features. To address these issues, we propose a self-interactive dynamic graph convolutional network with multiscale time-frequency fusion. It introduces a self-interaction graph convolution module designed to capture dynamic high order features of social interactions and employs a hybrid attention mechanism to fuse multiscale time-frequency features, thereby enhancing the representation of vehicle trajectories. In addition, the spatiotemporal attention module is adopted to achieve dynamic weight update of key information focusing on the interaction of the target vehicle with surrounding vehicles. Finally, a comparison and analysis of the proposed method was conducted based on datasets of NGSIM and HighD. The findings suggest that the suggested model demonstrates enhanced accuracy in trajectory prediction in comparison to leading techniques, achieving average root mean square errors of 1.54 m and 0.69 m across the two datasets when predicting for a duration of 5 seconds,and the average ADE and FDE indicators were 17.33% and 26.61% higher than the best performance.