Where to Go and How to Move: Goal–Motion Latent Flow for Multimodal Trajectory Prediction

Trajectory prediction plays a critical role in autonomous driving and intelligent surveillance. Recent generative models have shown promising results in capturing the uncertainty of future trajectories. However, many of these models do not explicitly distinguish long-term motion goals from local motion details in their trajectory representations, potentially affecting the accuracy of both local paths and endpoints. To address this challenge, we propose Goal–Motion Latent Flow (GMFlow), a novel dual-latent flow model comprising a goal–motion representation module and a conditional flow generation module. The representation module decomposes deviations from a constant-velocity reference trajectory into two complementary latents: a goal latent encoding long-term goals and a motion latent capturing fine-grained local variations. Conditioned on observed history and learnable mode queries, the conditional flow generation module jointly generates both latents through flow matching, producing multiple future trajectories in a single flow step. A goal-preserving decoder allows the motion latent to adjust local path shapes without changing the positions specified by the goal latent. Together, these designs enable accurate and diverse trajectory prediction with efficient flow-based generation. Experiments on NBA, SDD, and ETH-UCY demonstrate state-of-the-art performance, with particularly strong results in long-term endpoint prediction.