In complex environments where satellite signals are obstructed or denied, vehicle localization systems that rely solely on GNSS often fail to provide stable and high-precision position estimates. Visual landmarks therefore serve as essential auxiliary information for enhancing the robustness of autonomous navigation. However, traditional visual map construction methods typically suffer from long processing times, dependence on continuous tracking, and strong sensitivity to illumination variations and dynamic objects, resulting in insufficient landmark accuracy and spatial consistency. To address these limitations, this paper proposes a visual landmark construction method incorporating RTK constraints. The method first detects near-straight road segments based on the vehicle’s heading-angle sequence and selects representative image frames via linear interpolation as candidate visual landmarks, while introducing semantic priors to enhance their long-term stability and re-identifiability. Subsequently, a neighborhood image set is constructed around each landmark frame, and the initial camera poses are corrected using globally referenced RTK measurements to improve the robustness and reliability of local map initialization. Furthermore, an adaptive RTK constraint is integrated into both local sparse reconstruction and global optimization, enabling scale-consistent and geographically aligned high-precision landmark construction. A degraded smoothing constraint is activated when RTK signals are disturbed or temporarily unavailable to maintain system stability. Experimental results demonstrate that, compared with conventional approaches, the proposed method achieves notable improvements in map accuracy, reconstruction efficiency, and localization robustness under varying illumination, seasonal changes, and dynamic scenes, yielding an 82.2% increase in mapping accuracy, an 81.3% reduction in processing time, and up to a 70.1% improvement in localization precision.