STGNet: A spatio-temporal graph neural network for motion prediction in autonomous driving

Abstract

Predicting the future motion of agents in a scene is vital for safe autonomous driving. Current motion prediction methods typically utilize graph neural networks (GNNs) to model the interactions between scene entities retrieved from a fixed radius, but fail to consider potential correlations that extend beyond this radius. Moreover, existing methods only obtain a relative positional encoding between scene entities from a single viewpoint, ignoring spatio-temporal features from mutual viewpoints. Targeting these problems, this paper proposes a GNN-based motion prediction method referred to as STGNet. STGNet first exploits the spatio-temporal correlations in a scene to construct an interaction graph between entities based on their similarities in a feature space. Then, it adopts a dynamic neighboring strategy to adaptively select the radius of the neighborhood in the graph according to the scene context. Moreover, STGNet employs an expressive relative positional encoding to represent the pairwise relationships in the graph, which enhances the quality of the positional encoding by using the spatio-temporal features in mutual viewpoints. Experiments on two public motion prediction datasets, along with an extensive analysis, validate the effectiveness of STGNet and its superiority to other motion prediction methods in the literature.