Abstract: Objectives: Road maps play an important role in localization and motion planning for autonomous driving. To address the challenges of high cost and low efficiency in the construction of high-definition maps, this paper proposes a lightweight collaborative visual-semantic mapping method. Methods: This method constructs a sparse local vector map using semantic features of lane lines and employs a deep learning-based raster image matching technique for robust association of single-user and multi-user submaps, ensuring effective local map matching. Furthermore, vehicle poses are optimized through a factor graph framework, and global calibration is achieved by leveraging the local associations between map elements and poses, enabling efficient two-step map optimization. Results: Experiments conducted in urban road scenarios demonstrate that the proposed method successfully constructs a globally consistent and lightweight semantic map. Compared to VINS-Fusion odometry and ORB-SLAM3 mapping algorithms, the overall trajectory accuracy is improved by 89.9% and 80.5%, respectively, under cross-time and cross-season conditions. Conclusions: The proposed method enables effective multi-user collaborative mapping across large-scale environments, generating road feature maps that maintain global consistent while remaining highly lightweight.