Transparent Assessment of Automated Human Detection in Aerial Images via Explainable AI
Sara Narteni, Maurizio Mongelli, Joachim Rüter, Christoph Torens, Umut Durak · 2025
The widespread adoption of Artificial Intelligence (AI)-based software technologies supporting Unmanned Aircraft Systems (UASs) demands new validation methods to ensure that these systems operate safely and reliably. This paper investigates the role of eXplainable AI (XAI) and, in particular, of rule-based models in monitoring the performance of a deep learning-based human detector from aerial images. Starting from several image attributes extracted from the images and information about the performance of the detection model, decision rules are extracted to map the image attributes onto the performance quality. Besides shedding light on the logic of the humans detection successes and failures, these rules can serve as a performance monitor at runtime, by triggering alerts in case input images do not satisfy them. The obtained rules have been adopted to filter out inputs associated with bad performance, showing improved precision and recall with respect to the original model, thus opening the road to promising future developments.