Opt-FITP: optimization approach for multiagent trajectory prediction with fused importance

To address the deficiencies in interaction modeling and trajectory optimization, we propose a trajectory prediction framework (Opt-FITP) that integrates importance perception and adaptive calibration. We first construct a neighborhood aggregation module based on the attention mechanism to extract importance scores. Then, we dynamically prune low-importance interactions via adaptive thresholding and globally propagate high-importance scores to enhance cross-regional interactions. This enables the model to capture inter-regional agent interactions more effectively. In addition, for the trajectory optimization problem, a two-stage strategy is proposed, which employs calibration regularization constraints to align the variance and prediction error, and uses post-processing networks to generate scenario-adaptive scaling coefficients to adjust the variance estimation, and the joint optimization stage integrates the trajectory regression, modality selection, and calibration objectives to enable the model to better balance the trajectory accuracy and the uncertainty estimation. Our experiments on the Argoverse1 dataset show that Opt-FITP improves on minADE, minFDE and MR compared to existing methods.