Dynamic Spatio-temporal Graph Neural Network for Surrounding-aware Trajectory Prediction of Autonomous Vehicles

Trajectory prediction is a critical aspect of understanding and estimating the motion of dynamic systems, including robotics and autonomous vehicles (AVs). For safe and efficient driving behavior, an AV should predict its own motion and the motions of surrounding vehicles in the upcoming time steps. To achieve this, understanding the interaction among vehicles is crucial for accurate trajectory prediction. In this research, we implement a dynamic Spatio-temporal graph convolutional network to predict the trajectory distribution of vehicles in a traffic scene. We perform the graph convolutional network (GCN) operation on directed graphs to capture the spatial dependencies among vehicles in each traffic scene. To accurately replicate the interaction among vehicles, we propose a novel weighted adjacency matrix derived by the strategic positions of vehicles (angular encoding) and the reciprocal of distances among vehicles in a traffic scene. Additionally, we employ the temporal convolution network (TCN) to learn the temporal dependencies of a trajectory sequence and decode the future driving status using historic trajectories. We test the model with a naturalistic trajectory dataset (HighD) and conduct performance evaluation. The findings reveal that the proposed model could significantly improve accuracy compared to existing state-of-the-art models. Meanwhile, we conduct transfer learning to test the generalizability of our model on low data availability scenario using NGSIM (US-101) dataset. The results show that the relearned model perform comparability well and depicts competing performance in comparison to the state-of-the-art methods.