An improved hybrid model for vehicle trajectory prediction considering spatiotemporal interactions

The inherent uncertainty of surrounding drivers' intentions and spatiotemporal interactions between vehicles pose a formidable challenge in precisely forecasting trajectories. To address these issues, this paper proposes a hybrid deep learning model, termed CGAN, to predict vehicle trajectories with spatiotemporal interactions between the target vehicle and its six surrounding vehicles. It combines the Convolutional Neural Network with dilated convolution, the Gated Recurrent Unit, and the Attention mechanism, effectively capturing essential information from spatiotemporal interactions in extended sequences. The hyperparameters of the CGAN model were optimized by a Tree-structured Parzen Estimator and a novel feature scaling method was also designed to reduce dimensionality and enhance optimization efficiency. Comparative analysis utilizing the NGSIM dataset with other existing models reveals that the proposed model's performance improvement ranges from 11.32% to 18.13%, achieving prediction scores of 0.47 to 0.92 within the 2s to 5s prediction horizon, respectively.