Abstract: Accurately predicting the trajectory of vehicle is crucial to ensure the safety of autonomous vehicles. However, traditional methods have limited modeling and predictive capabilities when dealing with long sequence trajectories. To address this issue, a vehicle trajectory prediction model was proposed based on the Transformer network. The approach involves inputting the motion and interaction data of the vehicle into a driving intention prediction module to generate a probability intention vector. The trajectory prediction encoder is obtained after the Concatenate function is spliced with the trajectory information, and the trajectory features are fully extracted by using the multi-head attention mechanism. Through the decoder, a distribution of future vehicle trajectories is obtained. Validation on the NGSIM real vehicle trajectory dataset indicates that the accuracy of the driving intention prediction module can reach more than 85% under a prediction time of 2 seconds. Furthermore, the RMSE of the trajectory prediction model is reduced by more than 10% compared with the existing models under a prediction time of 4 seconds. The method provides technical support for accurately predicting the trajectory of autonomous vehicles.

Key words: autonomous driving, vehicle trajectory prediction, driving intention, feature extraction, multi-head attention mechanism