Structural Informer Network-Based Vehicle Trajectory Prediction for Motion Planning of Autonomous Driving

Efficient and accurate prediction of surrounding vehicles' trajectories over time is crucial for autonomous vehicle decision-making and planning. While the Transformer method has been widely used for interactive vehicle trajectory prediction due to its ability to consider multi-vehicle trajectories in parallel, this parallel computation mechanism causes computational exponential overload in Long Sequence Time-series Forecasting (LSTF) problems. In response, this paper proposes an improved Transformer method, known as the structural Informer, which can achieve accurate and efficient long-term trajectory prediction of the target vehicle (TV). Specifically, the proposed method considers not only the temporal and spatial features of the interaction trajectory, but also the impact of vehicle state changes on the trajectory. To reduce computational redundancy and complexity while improving memory usage and prediction accuracy, the ProbSparse self-attention mechanisms and attention distillation operations are employed. The method is validated and evaluated using the NGSIM dataset, and the results demonstrate that the proposed structural Informer achieves satisfactory accuracy and time cost in long-term prediction of the TV compared with various interactive trajectory prediction methods.