Advancing UAV Safety and Efficiency through Machine Learning Based Open-Set Detection

Abstract

This thesis presents the development of an open-set object detection framework for air-to-air scenarios with unmanned aerial vehicles (UAVs), aimed at enhancing perception reliability in real-world flight conditions. The proposed method is model-agnostic and operates on feature embeddings extracted from the detector, applying Gaussian Mixture Models (GMMs) to model semantic uncertainty through entropy estimation. To improve stability and discrimination, spectral normalization and temperature scaling techniques are integrated, while targeted data augmentation with simulated corruptions is employed to reflect aerial imaging conditions. The implementation process includes the adaptation of the framework for integration into modern detectors and its optimization for embedded UAV systems. The methodology is documented alongside the theoretical background, providing a practical reference for future work in UAV perception and open-set detection. This thesis was written in English to be accessible to a wider audience. A comprehensive summary in Greek follows.