A Graph Neural Network-Based Multi-agent Joint Motion Prediction Method for Motion Trajectory Prediction

Autonomous driving is a revolutionary automotive technology that can greatly improve driving safety and reduce traffic congestion. Multi-agent trajectory prediction plays an important role in autonomous driving technology. Considering that the existing methods are limited to local information areas because only local information is considered in the computation process, a multi-agent joint trajectory prediction model based on a GNN is proposed in this paper to help self-driving cars better plan their paths and avoid collisions. In the joint multi-agent prediction, this paper adopts a multitask fusion training strategy, which implements the selection and setting of multiple tasks through masks, and fuses the model for training. In the decoder and loss function design, the model is based on learnable anchors. A multimodal joint trajectory prediction decoder considering multi-agent and multimodal interactions is designed, and triple loss is introduced. The proposed method is validated on the Argoverse trajectory prediction dataset, and the experimental results show that the method can visualize the prediction results for typical traffic scenarios, intuitively demonstrating the accuracy of the prediction. The proposed model achieves better mean displacement errors and minimum end-point displacement errors with lower computational complexity, outperforming existing methods.