InstGPMap: Real-Time Instance-Level Global Prior Mapping via Historical Predictions Fusion

Abstract

Recent online methods for HD map construction directly infer local maps from sensor observations, yet suffer from limited perception range, particularly under challenging scenarios such as occlusions by large vehicles or poor visibility in rainy conditions. Inspired by human perception, which incrementally integrates previous observations to form stable prior knowledge about the environment, several methods have proposed utilizing global priors constructed from historical observations to enhance local map inference. However, existing global prior approaches often require significant storage overhead or complex post-processing, limiting their practical real-time usability. To address these challenges, we propose InstGPMap, a novel online framework that explicitly represents and maintains global prior map (GPMap) elements at the instance level. Specifically, our method explicitly leverages historical predictions instead of relying on implicit intermediate representations such as Bird's-Eye-View (BEV) features. We assign consistent instance identifiers (IDs) to map elements detected across frames, enabling direct instance-level association and updating. InstGPMap comprises two core modules: (1) the GlobalMapUpdate Module, which dynamically associates and manages GPMap elements across frames to form instance-level GPMap elements; (2) the PriorMapEncode Module, which encodes these instance-level GPMap elements into track queries, significantly enhancing real-time prediction accuracy. Extensive experiments demonstrate that InstGPMap achieves state-of-the-art performance on the nuScenes datasets, demonstrating superior accuracy and storage efficiency.