A Social-Aware Vehicle Path Forecasting Method using Graph Neural Networks

Mozhgan Nasr Azadani, Azzedine Boukerche

Abstract: Situational awareness can help the safety of automated vehicles, which involves understanding and forecasting the motions of nearby road users. Accurate motion forecasting enhances vehicular commutations, road safety, and mobility management. Early approaches merely model vehicle kinematics and ignore the impacts of nearby agents on each other, leading to inefficient results, especially for long predictions. Various types of agents use the same paths in a driving scenario. However, not all of these agents interact with each other. In fact, the actions of one agent in a road section do not impact all agents that use the same section. Accordingly, in this work, we argue that although modeling social interactions among road users is critical to have a safe path forecasting model, it should not be assumed that there is a connection between an agent and all of its nearby agents. We introduce a novel path forecasting model which benefits from graph neural networks to reason about these connections in terms of both time and distance. We produce the final predictions with temporal convolutions. We validate the path forecasting performance of our model using two large motion prediction benchmarks with different scenes and achieve state-of-the-art results in terms of displacement errors.