Multimodal Vehicle Trajectory Prediction with Stacked Attention

Predicting behavior of autonomous vehicles in the near future is very important for practical applications. This is a challenge because the driver’s intentions and behaviors are unheard, and it is inherently multimodal. Predicting the motion of road vehicles over a period of time in the future is a core challenge for autonomous vehicle safety, and viable solutions must consider static environments (such as road lanes) and dynamic environments that interact with multiple actors. This paper presents a multimodal trajectory prediction method based on graph theory and attention mechanism. Feature extraction of map lane environment and vehicle historical trajectory; the stacked attention module in the feature fusion layer is used to obtain the fusion information coding feature; finally, the predicted trajectory and the corresponding confidence score are obtained by the decoder. The experimental results show that compared with the comparative models, the present model has outstanding excellent results on the Argoverse dataset, and the prediction accuracy is better than the current mainstream methods.