Vehicle trajectory prediction based on spatial-temporal graph transformer networks with LSTM

Abstract

Vehicle trajectory prediction plays a vital role in autonomous driving and Intelligent Transportation Systems (ITS), significantly enhancing road safety and traffic efficiency. Over the past few years, deep learning techniques have achieved remarkable progress in this field. However, many existing methods may not fully capture the spatial-temporal interaction features between vehicles, which could potentially limit the prediction accuracy. This article presents a novel vehicle trajectory prediction model based on a Spatial-Temporal Graph Transformer and Long Short-Term Memory network (STGTL). The model adopts a sequential coupling design of multi-layer GCN and Transformer modules to effectively capture spatial interactions and temporal dependencies. Initially, a Long Short-Term Memory (LSTM) encoder separately processes the historical trajectories of the target vehicle and the neighboring vehicles. Following that, a Graph Convolutional Network (GCN) module is employed to model the traffic scene, capturing the spatial interactions among vehicles. Subsequently, a Transformer module enhances the extraction of temporal features, resulting in rich spatial-temporal interaction representations. In the final stage, an LSTM decoder integrates these spatial-temporal interaction features with the historical trajectory of the target vehicle for decoding, thereby generating the predicted trajectory sequence. To verify the effectiveness of the proposed model, comprehensive experiments were carried out using the NGSIM dataset. The comparative results indicate that the proposed model achieves substantial improvements in prediction accuracy compared to existing methods. Furthermore, ablation studies were conducted to examine the impact of different modules on the model’s performance, which further validates the rationality of the model design.