Vectorized HD Map Learning for Autonomous Driving: A Survey of Representations, Architectures, and Future Directions

Abstract

High-definition (HD) maps are essential for perception, localization, and planning in autonomous driving, providing geometric and semantic descriptions of road environments. This survey reviews recent advances in learning-based HD map construction with a focus on vectorized and structured representations that support high-level autonomy tasks. We analyze the evolution from rasterized to vectorized and graph-based approaches, discussing their respective advantages, limitations, and implications for downstream planning and forecasting. The paper categorizes dominant architectural paradigms, including transformer-based set prediction, diffusion models, and graph neural networks, and compares their performance on major benchmarks such as nuScenes, Argoverse2, and OpenLane-V2. We further examine sensor fusion strategies, topological reasoning mechanisms, and evaluation metrics tailored to structured map learning. Finally, we identify key open challenges, including generalization, real-time scalability, and topological consistency, and outline future research directions toward deployable learning-based HD mapping systems.