| description abstract | Trajectory prediction plays a crucial and indispensable role in autonomous driving, enhancing safety and reliability, particularly in complex intersections with variable vehicle trajectories. This study aims to investigate the application of vehicle trajectory prediction in intersections and to improve the prediction accuracy of the model by utilizing multicolumn neural networks and self-attention mechanisms. Trajectory data were extracted using DataFromSky software from unmanned aerial vehicle (UAV)-captured aerial footage at specific intersections in Xi’an, Shaanxi Province, China. The deep learning model developed for trajectory prediction included a multicolumn (MC) neural network and a self-attention (SA) mechanism based on the gated recurrent unit (GRU). The model was abbreviated as MC-SA-GRU in this study. The multicolumn structure captures richer feature representations, whereas SA captures temporal correlations in time-series data. Experimental comparisons on corresponding data sets demonstrate the exceptional performance of these models in accurately predicting vehicle trajectories at intersections. This study contributes valuable insights for improving autonomous driving systems, advancing safety and reliability in real-world scenarios. The results of this case study showed that combining deep learning time series models with plug-in modules effectively improved the accuracy of vehicle trajectory prediction. Vehicle trajectory prediction was a highly significant research problem in the field of autonomous driving, particularly in complex intersection scenarios. Accurate vehicle trajectory prediction had the potential to significantly enhance the safety and reliability of self-driving vehicles. The data utilized in this study were collected by an unmanned aerial vehicle at several specific intersections in Xi’an, Shaanxi Province, China. The proposed method was based on gated recurrent units, incorporating multicolumn neural networks to fully explore the potential features of the data, and included a self-attention (SA) mechanism to consider the temporal correlation of the data. It was compared with various benchmark methods, demonstrating excellent performance across various evaluation metrics. Experimental results validated the effectiveness of this approach. Thus, this research provided valuable insights for improving the safety and reliability of real-world autonomous driving systems. | |
