Every day, millions of decisions in Out-Of-Home advertising rely on traffic frequency data — how many cars pass a billboard, how many pedestrians cross a junction, how many eyes a campaign actually reaches. But what if the data itself was quietly wrong? That was exactly the problem our team set out to solve in our diploma thesis, AI4Ads, developed in partnership with R+C Plakatforschung GmbH, an Austrian out-of-home advertising research institute. The errors are subtle: at a road junction, 10,000 vehicles might enter, but only 400 leave across two exits. Physically impossible. Our goal was to build a system that detects and corrects them automatically using machine learning.

The project was divided across four distinct layers of the problem. One team member focused on ingesting and preparing the raw GIS data exported from QGIS; another built and evaluated anomaly detection approaches; a third developed correction strategies for identified violations and the fourth responsible for MLOps designed the infrastructure that would tie it all together. No single part of the project could stand alone — the pipeline only made sense as a whole.

Did everything go according to plan? Not entirely — and we think that is worth being honest about. Integrating machine learning into a domain as specific as GIS-based traffic flow analysis is genuinely hard, and a one-year diploma thesis can only take that so far. What we built is a foundation: a working MLOps infrastructure, a documented data pipeline, and a set of detection and correction approaches that were evaluated against real data. The groundwork is laid. With more time, more training data, and continued collaboration with R+C, the approaches explored here could realistically evolve into a production-ready tool. We hope this project serves not as a finished answer, but as a documented starting point for whoever picks it up next.