IDRNet: Intention-Based Deep Reinforcement Learning Network for Multiagent Trajectory Prediction

Multiagent trajectory prediction is a critical challenge in the domains of intelligent traffic management systems and autonomous driving. Comprehensive understanding of the behavior of dynamic agents and their surrounding environment is the foundation for automated decision-making. In the existing research, many methods focus on parsing environmental information. However, in relation to agents, as a result of the high uncertainty of their intentions and that their behavior is constrained by the environment and traffic rules, significant breakthroughs have been slow to emerge. This paper proposes an intention-based deep reinforcement learning network (IDRNet) for multiagent trajectory prediction, which regards each agent as an independent decision-maker and optimizes its intention strategy through the trial-and-error mechanism of reinforcement learning, thereby learning the interaction rules between agents and the dynamic characteristics of the environment. This study presents an evaluation of the IDRNet model against a baseline model on the Argoverse Motion Forecasting V1.1 (Argoverse V1.1) dataset, demonstrating significant improvements in convergence speed, trend of trajectory changes, and prediction accuracy. Specifically, IDRNet achieved a 9.62% reduction in mean average displacement error and a 4.58% reduction in minimum final displacement error compared with the baseline. Notably, in the context of single vehicle prediction across open environments, these improvements are particularly significant, with increases reaching 26.25% and 27.94%, respectively. The research leverages IDRNet for data augmentation and trajectory prediction, effectively addressing the challenge of achieving human-like autonomous driving under extreme conditions such as perception system failures and hardware occlusions, relying solely on localization information. In addition, comparative experiments were conducted, revealing that IDRNet outperforms other models, including graph attention-long short-term memory networks and partially observable Markov decision processes.