IPF-GCN: A graph convolutional network based on the interaction potential field for multi-vehicle trajectory prediction

Vehicle trajectory prediction is a key task to ensure the safety of autonomous driving, especially in dense traffic scenarios, where the future trajectory of a vehicle is jointly influenced by the historical trajectory of the self-vehicle and the interaction of the surrounding vehicles, and the complex and stochastic interactions among the vehicles bring challenges to the prediction of vehicle trajectories. In this paper, we analyze the temporal and interaction characteristics of the vehicles and propose a trajectory prediction model based on the Interaction Potential Field Graph Convolutional Network (IPF-GCN). A Bi-LSTM attention network is used to extract the bidirectional temporal features of historical trajectories so that the model focuses on the important information in the trajectories. An artificial potential field that captures the longitudinal and lateral interactions between vehicles is constructed, and the vehicle interaction features are extracted based on a bi-layer graph convolution network (GCN). Furthermore, the future trajectory prediction of the vehicles is achieved based on the LSTM decoder and considering the driving intention. Finally, the model is experimentally validated on HighD and ExiD datasets. Compared to the baseline models, our model has higher trajectory prediction accuracy and provides good trajectory prediction in dense traffic situations.