Prior-conditioned edit prediction for real-time incremental updating of vectorized high-definition map

Abstract

High-Definition (HD) maps are essential for intelligent transportation systems and autonomous driving, but their maintenance remains costly and time-consuming. Existing end-to-end perception models reconstruct vectorized HD maps from onboard sensors, whereas practical deployment typically starts from an existing but possibly outdated prior. We propose MapUP, a prior-guided edit prediction framework built on MapTR for real-time incremental HD map updating. MapUP encodes stale prior polylines as instance-level embeddings and injects them into Transformer queries, enabling decoding from structured map hypotheses rather than from scratch. For database-compatible maintenance, MapUP predicts KEEP, MODIFY, and DELETE for prior-associated instances; ADD operations are assembled from high-confidence predictions unmatched to valid priors. To overcome the lack of real before–after map annotations, we synthesize stale priors through controllable deletion, geometric perturbation, and spurious polyline insertion, and train MapUP with vector map detection, edit classification, and geometry regression losses. Experiments on nuScenes using a before–after protocol show that MapUP consistently improves maintained maps across three prior-quality regimes, increasing mAP from 35.6 to 62.9 under severe staleness, from 63.1 to 82.9 under moderate staleness, and from 91.6 to 95.9 with strong priors, while achieving 9.32 FPS in a controlled batch-1 inference benchmark on a single NVIDIA RTX 4090D GPU. Under moderate staleness (S2), MapUP outperforms reconstruction-centered alternatives: MapTR + Geometric Differencing and MapUP prediction + Geometric Differencing reach 46.06 and 70.20 mean Average Precision (mAP), respectively, whereas MapUP achieves 82.90 mAP with only 0.67% unnecessary updates. Prior-shift experiments show positive update gain under localization drift, and Argoverse 2 results further demonstrate cross-dataset generalization. These results indicate that prior-conditioned initialization and learned edit prediction offer an effective, interpretable, and maintenance-oriented solution for incremental HD map updating.