Vehicle Behavior Prediction Based on GAT and Transformers
Vehicle behavior prediction can assist autonomous driving systems in making decisions,thereby enhancing the safety and ef-ficiency of autonomous driving.However,in different road scenarios,dynamic changes among surrounding traffic participants,such as cars,bicycles,and pedestrians,can lead to significant errors in predicting vehicle position information.This may prevent autonomous vehicles from taking timely evasive or emergency braking actions.This paper aims to construct dynamic spatiotemporal graphs of inter-actions among traffic participants for structured and unstructured road scenarios.It utilizes deep learning techniques to design a GAN model based on GAT(Graph Attention Network)and Transformer for vehicle behavior prediction.GAT is employed to learn the corre-lations and interaction patterns among different participants,while Transformer is used to extract temporal features of traffic partici-pants'motion states.Simulation experiments are conducted on the NGSIM and ApolloScape datasets.The results demonstrate that our model exhibits higher accuracy in long-term predictions while maintaining a more lightweight profile.
NETL
NSTL
万方数据
