Exploring the Impact of Social Interaction Factors on Vehicle Trajectory Prediction with Graph Neural Networks

In real-world driving scenarios, human drivers often engage in direct or indirect cooperation or ‘negotiation’ with other road users. This social interaction is crucial for ensuring safe and efficient driving. Fully considering the interactions between vehicles in trajectory prediction significantly enhances the accuracy of these predictions. Thus, this paper introduces a method for driving trajectory prediction that utilizes a Graph Neural Networks-Recurrent Neural Network (GNN-RNN) model. This method involves training on the HighD natural driving dataset, with a focus on extracting and analyzing vehicle interactions. The results show that the model considering multi-vehicle interactions performs better than previous models in terms of root mean square error (RMSE) in lane-changing scenarios. Meanwhile, the precision of trajectory prediction may be affected by different social value orientations, with higher prediction errors for vehicles with egoistic orientations in driving performance. In addition, other factors in vehicle driving performance such as lateral longitudinal speed and acceleration also affect the accuracy of trajectory prediction. Therefore, it is informative to consider the social interactions between autonomous vehicles and other interacting vehicles to improve the precision of vehicle trajectory prediction models in lane-changing scenarios.