GTC-DAN: A graph-temporal convolutional model with dynamic adjacency for vehicle trajectory prediction

Abstract

Autonomous driving is currently an issue of heated debate in automotive engineering. Accurate prediction of the future trajectory of self-driving cars can significantly reduce the occurrence of traffic accidents. However, predicting the future trajectories of vehicles is a challenging task since it is influenced by the interaction behaviours of neighbouring vehicles. This paper proposes a framework that allows for parameter sharing and cross-layer independence, based on a dynamic graph convolutional spatiotemporal network, to study the interactions between vehicles and the temporal dynamics in historical trajectories. By extracting dynamic adjacency matrices from different vehicle interaction features, the model can describe dynamic spatiotemporal relationships and facilitate addressing changes in traffic scenarios. Finally, the proposed model is experimentally compared with existing mainstream trajectory prediction methods using the NGSIM dataset. The results demonstrate that our trajectory prediction model achieved excellent performance in terms of model parameters and prediction accuracy. Compared to the four mainstream models, our model improved accuracy by 35.73%. In addition, we also analyze the relationship between model complexity and efficiency.