Multimodal vehicle trajectory prediction based on intention inference with lane graph representation

Accurately predicting the trajectories of nearby vehicles is a crucial and complex task in autonomous driving due to the inherent uncertainty in driving behavior. Multimodal trajectory prediction methods have emerged as promising approaches to reduce uncertainty. However, these methods frequently face the "mode collapse" issue, where the generated trajectories are limited to one or a few modes, or most generated trajectories fail to comply with road constraints. To address this problem, we propose a multimodal trajectory prediction model named Intention Inference with Lane Graph representation (IILG). This model divides the problem into three subtasks: encoding traffic agents (i.e., road users) and scenes with interaction considerations, predicting the goal set through learning and optimization, and decoding multimodal trajectory using multi-head attention. The Agentnode attention method is implemented to capture complex interactions among the target vehicle, surrounding agents, and scenes. To capture all potential reasonable intentions, we integrate the maximum entropy principle into the optimization function for multi-goal selection. This methodology eliminates the necessity for intricate manual anchor settings and overcomes the limitations of noise sampling, which often lacks semantic information. Experiments conducted on the real-world datasets nuScenes show that the proposed model outperforms existing models in prediction performance, and the generated trajectories demonstrate enhanced diversity and physical realism. Our model offers a unique perspective in guiding multimodal trajectory generation and deducing plausible potential driver intentions, thereby enhancing prediction fault tolerance and facilitating the analysis of potential vehicle conflicts and risks in complex scenarios.