DERGCN: Dynamic-Evolving graph convolutional networks for human trajectory prediction

Pedestrian trajectory prediction is an increasingly important research area in applied autonomous driving and social robotics. Effectively modeling the intricate interactions between pedestrians is paramount for improving trajectory prediction accuracy. However, when using Graph Neural Networks(GNNs) to model these interactions, fixed interactions tend to remain, preventing the graph model from making adaptive adjustments and thus resulting in significant discrepancies between the predicted and true trajectories. In this study, we propose a Dynamic-Evolving Relative Graph Convolutional Network(DERGCN) to predict the future trajectories of pedestrians. The network model captures the dynamically evolving pedestrian interactions and incorporates an evolving mechanism to simulate them. In addition, with a relative temporal encoding strategy employed to improve the dynamics of the graph further, our policy network yielded an improved predictive performance when tested on two challenging datasets.