GST-former: GCN-based spatial temporal transformer model for interaction-aware vehicle trajectory prediction

Trajectory prediction is an important component in autonomous driving systems as it forecasts the future movements of nearby vehicles. Therefore, accurate prediction of vehicle trajectory is necessary to ensure road safety. The trajectory prediction with accuracy is still facing challenges due to the intricate relationship of spatiotemporal dependencies among vehicles. Therefore, this research proposes a novel framework called GST-Former by combining the Graph Convolutional Network (GCN) based Spatial Temporal with a Transformer model for Trajectory prediction. The main contribution of the study is to apply the GCN for parallel extraction of spatial and temporal features, which are combined using the Fusion technique. The Transformer model in the proposed framework contains an encoder and decoder, and uses a multi-head attention-based mechanism for the trajectory prediction module. The proposed model is trained and validated using the actual vehicle trajectory dataset, Next Generation Simulation (NGSIM). The target and neighboring vehicles are selected for trajectory prediction using Time-Distance criteria. The performance of the proposed model is evaluated using the root mean square error (RMSE) and compared to the traditional baseline and current state-of-the-art models. Furthermore, the performance of the proposed model is evaluated based on several aspects, in the ablation studies using feature extraction modules without spatial GCN and temporal GCN, layer configuration for encoder and decoder in the Transformer model, along with the single and dual encoders. The study also performs quantitative evaluation using single and multiple vehicles. The results show that the proposed model has reduced prediction error compared to the baseline and literature studies. The results also show that using the GCN for spatial and temporal features achieves better accuracy as compared in the ablation study.