Self-localization is becoming increasingly important in areas such as augmented reality, robotics and autonomous driving. Localization using positioning systems is often not possible or too imprecise. In this thesis, camera-based localization using a classification model, specifically YOLOv8, is investigated in the field of miniature autonomy. The HAW Hamburg’s Mikrowunderland was used as a test environment, as it depicts a large number of road environment features. It was investigated how accurately the model can predict its position on the route and how many images are required per class. It was also analyzed what the model sees and what the reasons for an incorrect prediction are. It turned out that even a small number of images in training can provide a prediction accuracy of 2 centimeters. In most cases, incorrect classifications were due to incorrect image acquisition and not to the model’s ability to correctly recognize differences.