Trajectory Prediction Based on Spatio-Temporal Fusion Graph Neural Networks

Abstract

The development of artificial intelligence has brought new opportunities to the field of autonomous driving trajectory prediction. Most existing research considers pairwise interaction between individual vehicle behaviors, while overlooking the impact of different information processing factors between static map information and traffic participants on predictions. This paper proposes a spatio-temporal fusion convolution trajectory prediction method based on Graph Neural Networks (STGCN). First, a novel dual-channel spatio-temporal graph mechanism is constructed to capture global map and local interaction information. Next, historical information between interacting agents is processed in the temporal dimension, introducing the temporal convolutional network to extract temporal features of historical trajectories, FusionNet is introduced to handle the spatio-temporal information. Finally, the encoder-decoder structure of GRIP++ is employed to decode the graph features and generate predicted trajectories. Experiments are conducted on the nuScenes dataset. Quantitative experiments demonstrate a significant improvement in ADE and FDE performance on the nuScenes dataset. Qualitative analysis in typical scenarios indicates that the proposed model can successfully complete prediction tasks for left turns, straight driving, and right turns.