Joint vehicle trajectory prediction via multi-scale and future motion interaction modelling

Accurate trajectory prediction is crucial for autonomous driving, but existing models often fail to capture complex, multi-scale interactions and future motion dependencies. To address this, we propose FMMSNet, a novel framework for autonomous vehicle (AV) trajectory prediction. FMMSNet integrates two key components: the Multi-Scale Interaction Network (MSINet) and the Future Motion Interaction Network (FMINet). MSINet captures spatiotemporal interactions at varying scales using a three-stage attention mechanism, generating diverse future trajectories via a Laplace mixture. FMINet explicitly models dynamic interactions by encoding historical trajectories and predicted futures, ensuring consistency across predicted paths. Experimental results on the Argoverse dataset show that FMMSNet outperforms baselines, improving minADE by 13.5% and MR by 30.8%. Ablation studies emphasize the value of multi-scale interaction layers and future-aware reasoning, demonstrating the effectiveness of FMMSNet in improving trajectory prediction accuracy in complex traffic environments.