Toward High-Accuracy and Low-Latency Group Vehicle Trajectory Prediction With Linear UNet-Enhanced Fully Connected Spatial-Temporal Graph Neural Network

Group vehicle trajectory prediction (GVTP) is important for analyzing the traffic states and optimizing the traffic management. However, existing studies have performance bottlenecks in the prediction accuracy and inference latency. To tackle the problems, we propose a linear UNet-enhanced fully connected spatial-temporal graph neural network (LUFC-STGNN) for GVTP. First, a spatial graph is constructed by integrating prior-based and data-driven methods to capture both explicit and implicit spatial interactions between the vehicles. After that, a comprehensive temporal graph is created to capture the varying strengths of temporal interactions between all vehicles throughout historical timestamps. Furthermore, a fully connected spatial-temporal graph combining the spatial and temporal graphs is introduced to extract the effective spatial-temporal interaction features of the vehicles through the graph convolution operation. Finally, a linear UNet-based temporal dependency encoder (LU-TDE) is designed to further enhance the model’s ability of capturing the potential temporal patterns in the vehicle interactions. The encoder with linear complexity explores the multi-scale temporal dependencies from the spatial-temporal interaction features but also reducing the inference latency. Experiments results based on real-world datasets show that compared to state-of-the-art models, our model reduces the average root mean square error over the 5-second prediction horizon by 31% and 10% on the NGSIM and HighD datasets, while reducing the inference latency by at least 1.26 times.