VIF-GNN: A Novel Agent Trajectory Prediction Model based on Virtual Interaction Force and GNN

Abstract

Agent trajectory prediction of traffic scenarios is a significant module of environment reasoning and autonomous vehicle decision, and the core challenge is the ability to interaction reasoning under complex scenes. Previous prediction models are either not precise enough or require massive computational costs. In this paper, we propose VIF-GNN, a novel traffic agent trajectory prediction framework based on the Virtual Interaction Force (VIF) concept and Graph Neural Network, which consists of semantic feature engineering, a subgraph encoder, a global graph, and the trajectory decoder. In particular, this method extracts vectorized features including VIF adjacent matrix from raw inputs and transfers them into graph nodes through the subgraph encoder. The global graph module obtains spatiotemporal reasoning information from four various interaction layers combined with the VIF prior knowledge. And the decoder translates the graph into trajectories of the target agent. Experiments prove that VIF-GNN could achieve precise forecasting on both single and multi-modal prediction task compared with the baselines while maintaining a relatively light parameter size scale, ensuring the real-time performance of vehicle platform applications.