DiffVecMap: A Robust Online Vectorized HD Map Construction Method With a Diffusion Model

Abstract

High-definition (HD) map construction plays a critical role in providing precise and comprehensive static environmental information for autonomous driving systems. However, the performance of existing methods degrades significantly under adverse conditions such as nighttime or rainy weather, where the quality of visual data deteriorates. To address this issue, we propose an online vectorized HD map construction method, called DiffVecMap, which leverages the powerful denoising and generative capabilities of the diffusion model to effectively enhance robustness in such challenging environments. To effectively apply diffusion techniques to the map construction task, a diffusion block based on conditional feature modulation is proposed, in which a time-adaptive gate mechanism is incorporated into the modulation process to improve the controllability of the diffusion trajectory. At the same time, a lightweight hierarchical Bidirectional Feature Pyramid Network (BiFPN) is employed as the architecture of the diffusion model, where multi-scale representations are efficiently fused through adaptively learned weighting coefficients. Experimental results demonstrate that DiffVecMap achieves a competitive balance between accuracy and efficiency while exhibiting strong robustness. On the nuScenes original validation set and the separately partitioned night and rainy splits, DiffVecMap outperforms the baseline method by 3.0 mAP, 4.9 mAP, and 5.2 mAP, respectively.