Multidimensional graph transformer networks for trajectory prediction in urban road intersections

With the rapid development of autonomous driving, accurately predicting the future movement trajectories of various agents in complex scenarios, such as urban road intersections, will directly affect whether planning and decision-making levels can plan reasonable and effective control strategies, which is crucial for autonomous driving and road safety. Currently, trajectory prediction methods in urban road intersections do not fully consider the influence of multiple types of peripheral agents and multidimensional global interaction information, and cannot effectively model the spatial interaction and temporal correlation of agents in complex traffic scenes. To solve these problems and improve the prediction accuracy, a new multidimensional graph transformer network (MDGTN) method for urban road intersections is proposed. In this model, a multidimensional spatio-temporal class heterogeneous graph network is considered, and an encoding-decoding architecture is used to extract agent interaction features via the spatio-temporal class multi-layer transformer architecture. Compared with the current optimal method, the average deviation error and final deviation error of the MDGTN are improved by 10% and 8% in a public complex mixed-traffic data set, which fully proves the superiority of the MDGTN model and the accuracy of multi-agent future trajectory prediction.