基于未来潜状态交叉注意力的多模态轨迹预测方法

A Multimodal Trajectory Prediction Method Based on Future Latent State Cross-Attention

摘要: 准确预测多个交通参与者的未来运动轨迹对于实现安全可靠的自动驾驶至关重要。尽管近年来的轨迹预测方法通过建模智能体交互已展现出优异性能，但现有方法主要关注历史交互而忽视了智能体未来运动之间可能出现的复杂依赖关系，这导致预测轨迹在密集交通场景中可能缺乏全局交互一致性。为了应对这一挑战，本文设计了一种基于未来轨迹交互建模的轨迹预测算法框架，通过考虑未来潜在空间内的多模态交互关系，在解码阶段对其进行显式建模。具体而言，本方法将轨迹预测分解为基于历史的粗粒度预测和基于全局交互一致性的未来轨迹优化两个阶段。第一阶段通过矢量化场景表征，采用分层式编码机制融合局部上下文特征与全局场景交互特征，从而实现对各种交通场景中丰富的历史时空交互进行建模。为捕捉历史观测之外的交互关系，第二阶段进一步在解码器中引入未来潜在状态的空间交叉注意力模块(Future Latent Cross-Attention, FLCA)，设计了跨智能体的交互掩码机制，使每个预测模态都能关注其他智能体的未来运动，同时自身不同模态之间不出现干涉。最后，在大规模自动驾驶基准数据集Argoverse1上的实验表明，本方法能够实现更具全局交互一致性和准确性的轨迹预测。

Abstract: Accurate prediction of the future movement trajectories of multiple traffic participants is crucial for achieving safe and reliable autonomous driving. Although recent trajectory prediction methods have demonstrated excellent performance by modeling agent interactions, existing approaches mainly focus on historical interactions while neglecting the complex dependencies that may arise between agents’ future movements. This leads to the possibility that the predicted trajectories may lack global interaction consistency in dense traffic scenarios. To address this challenge, this paper designs a trajectory prediction algorithm framework based on future trajectory interaction modeling, explicitly modeling the multimodal interaction relationships in the future potential space during the decoding stage. Specifically, this method decomposes trajectory prediction into two stages: coarse-grained prediction based on history and future trajectory optimization based on global interaction consistency. In the first stage, a vectorized scene representation is used, and a hierarchical encoding mechanism is adopted to fuse local context features and global scene interaction features, thereby enabling the modeling of rich historical spatiotemporal interactions in various traffic scenarios. To capture interaction relationships beyond historical observations, the second stage further introduces a Future Latent Cross-Attention (FLCA) module in the decoder and designs an interaction masking mechanism across agents, allowing each prediction modality to focus on the future movements of other agents while avoiding interference among different modalities of the same agent. Finally, experiments on the large-scale autonomous driving benchmark dataset Argoverse1 show that this method can generate trajectory prediction results with better global interaction consistency and accuracy.