Localization in indoor applications is usually performed by means of environmental sensors such as lidar, mono and stereo cameras. Nowadays, cost-efficient monocular cameras are widely used in series production vehicles for driver assistance systems. In the context of higher automation, however, they face the problem of missing depth estimation, so that true-scale information cannot be extracted. As a result, localization maps generated with Visual SLAM exhibit arbitrary scaling in addition to the typical drift. Therefore, high-precision localization in a globally referenced coordinate system is not feasible. Addressing this challenge, we present an approach to achieve highly precise localization using only a monocular camera. Based on the opensource algorithm ORB-SLAM3 and ground truth data, we first generate globally referenced, highly accurate feature maps. In the second step, we evaluate the localization in real-world experiments using representative scenarios such as valet driving at the end-of-line at Ford Motor Company's manufacturing plant, automated valet parking in parking garages and outdoor driving in semi-structured areas. Emphasis is placed on investigating suitability under varying influences such as time of day, weather and environmental objects. In typical cases, we achieve an average accuracy of about 3 cm with maximum deviations of up to 8 cm.