GraphMap: Scalable Crowd-Sourced Global Vectorized HD Map Construction via Sparse Visual Graph Fusion

Abstract

High-definition (HD) maps are vital for autonomous driving, offering fine-grained geometric and semantic information beyond onboard perceptions. While recent vision-based methods enable online local HD map detection, constructing accurate global vectorized maps at scale remains a fundamental challenge: (1) individual vehicles have limited sensing range, and (2) the mainstream ego-centric dense fusion approaches for collaborative perception suffer from poor scalability, high computation costs, and sensitivity to spatial-temporal misalignment. In this paper, we present GraphMap, a novel crowd-sourced vehicle-cloud framework for global vectorized HD map construction via sparse graph fusion. Unlike prior collaborative perception methods that rely on dense, fixed-size bird’s-eye-view (BEV) features, our approach is not constrained by ego-centric views or fixed perception ranges. GraphMap encodes local vectorized maps as sparse geometric graphs and incrementally fuses them through a sparse-to-sparse graph fusion algorithm. To reduce network bandwidth and computation overhead, GraphMap incorporates a selective map sensing mechanism that prioritizes data uploads from more informative agents with higher sensing value. Experiments across multiple real-world and simulated driving scenes with up to 60 agents demonstrate that GraphMap constructs accurate global maps under sparse, spatially misaligned observations, outperforming state-of-the-art baselines by 37.2 mAP while reducing communication costs by more than 30%.