DS-SGAN: Driving Style-Guided Spatial-Graph Attention Network for Trajectory Prediction of Vehicle

Vehicle trajectory prediction is a crucial and challenging task, especially in complex traffic scenarios involving multiple vehicle interactions, due to the complexity of the road environment and the variety of driving styles exhibited by different drivers. Prediction models ignoring influences of surrounding vehicles/roads and driving styles may produce unrealistic results and have low prediction accuracy. To address this issue, this paper introduces a novel driving style-guided trajectory prediction method based on a generative adversarial network. The generator utilizes LSTM and graph neural networks to encode relevant information about the target vehicle and the surrounding scene. To enhance the model's predictive capabilities further, it introduces a spatial-graph attention mechanism, which can pay attention to more important information in the spatial domain, to capture influences of surrounding vehicles/roads and generate reasonable predicted trajectories. We introduce unsupervised clustering to distinguish different driving styles and use discriminator to ensure that the generator generates trajectories conforming to the corresponding driving styles. Therefore, the entire framework is capable of not only generating accurate predictions but also aligning with the potential driving behaviors exhibited by drivers. To validate the efficacy of the proposed method, extensive evaluations are conducted on the Argoverse motion forecasting benchmark. The results demonstrate that our method outperforms other competitive approaches.