Research on vehicle trajectory prediction algorithm based on improved transformer

Trajectory prediction is an important part in autonomous driving technology. The current prediction algorithm model still has some problems in the interaction behavior of traffic participants, the utilization rate of highly refined map information, and the learning of potential traffic rules in the real world. In this paper, a vehicle trajectory prediction algorithm model based on neural network is built, which selects Argoverse 2 Motion Forecasting as the training data set of the model. Relative spatiotemporal position encoding is used to encode the input information of the trajectory prediction model. It can better understand the complex interaction behavior in traffic scenes. The K-Means++ algorithm is used for the trajectory clustering of the Waymo dataset, integrating their results as prior knowledge into the trajectory prediction model to enhance their ability of learning from the potential traffic rules in the real world. The performance of the model was tested according to the selected data set. The average displacement error, the final displacement error, and its minimum value are used as the evaluating indicator of output trajectory and perform quantitative analysis. Visualizing the prediction trajectory of the model to visually judge the error and feasibility of the model. The results prove that the prediction model has good prediction ability for some behaviors such as going straight, turning left, and turning right. It can also well predict the future state information of other traffic participants such as pedestrians and non-motor vehicles.