PHR-Net: Proposal-Level Historical Retrieval for Non-Stationary Temporal Consistency in Trajectory Prediction

Multi-agent trajectory prediction serves as a critical component in autonomous driving systems, bridging environment perception, behavior understanding, and motion planning. Its outputs not only affect candidate trajectory evaluation and interactive decision-making but also directly influence downstream processes such as risk anticipation, braking and yielding, and safety margin allocation. Therefore, obtaining accurate and stable prediction results is of great importance. Although existing methods have achieved remarkable progress in single-timestep prediction accuracy, most of them still adopt an independent decoding paradigm under a sliding-window setting. As a result, during continuous online prediction, these models are prone to frequent mode switching, temporal discontinuities in overlapping segments, and local trajectory jitter, which become particularly pronounced in complex interactive scenarios such as yielding, merging, and unprotected turning. To address these issues, this paper proposes PHR-Net, a two-stage proposal-level historical retrieval framework that introduces cross-timestep historical context to perform consistency-aware refinement of current predictions on top of multimodal coarse proposals. Experiments on the Argoverse 1 benchmark show that PHR-Net achieves competitive performance under both Top-1 and Top-6 settings. PHR-Net obtains a Top-1 minFDE of 1.0834 and MR of 0.1046 and achieves an MR of 0.1027 under the Top-6 setting. In the overlapping-interval consistency evaluation, PHR-Net reduces the summed ADE to 2.08. These results show that proposal-level historical retrieval improves endpoint reliability and cross-timestep temporal consistency.