STF-GT: spatio-temporal fusion graph transformer for vehicle trajectory prediction

Abstract

To enhance the understanding of multi-vehicle interaction scenarios in autonomous driving and achieve accurate vehicle trajectory prediction, we proposed a spatio-temporal fusion graph transformer network for vehicle trajectory prediction. The multi-head attention mechanism is employed to capture the historical motion states, while the graph attention mechanism is utilized to model dynamic spatial interactions among vehicles. A gated fusion module is designed to integrate spatio-temporal features, and the future trajectory of the ego vehicle is generated by optimizing a diversity loss function. To validate the effectiveness of the proposed model, extensive experiments are conducted on open-source datasets, and the results are compared with advanced prediction models. Experimental results demonstrate that the error of the proposed model is close to the optimal level of the existing models in the prediction horizon of 1 s, and the root-mean-square error at 2, 3, 4, and 5 s reaches the optimal or suboptimal level. When the prediction time is 5 s, the average root mean square errors of the two datasets are 1.42 and 0.44 m, respectively. The overall prediction accuracy across all time domains is significantly improved, and the model exhibits high accuracy and robustness in complex scenarios.