Edge-accelerated machine learning-based face verification using open autonomous UAV platforms

This thesis presents a novel open-source Unmanned Aerial Vehicle (UAV) platform designedto be sensor-agnostic, enabling the integration of various cameras and sensors. The platformincorporates a companion computer, the BeagleBone AI-64, with onboard Artificial Intelligence(AI) capabilities, facilitating real-time data processing and analysis during flight. Akey focus of this research is the evaluation of face verification algorithms for UAV applications.The study investigates the effectiveness of different face verification algorithms andintroduces a novel dynamic distance-based thresholding technique to enhance their accuracywhile varying distances, exhibiting that an improvement of 15% in accuracy can be achieved.Moreover, a new face detection algorithm, UWS-YOLO, is proposed, demonstratingsuperior performance in detecting faces from UAV-captured images compared to existingmethods, enabling effective face detection from UAV footage up to 30 metres with a lowinference time of only 11 ms, 345% faster than RetinaFace and 373% than YOLOv7.Moreover, it achieves 59.29% of accuracy compared with 27.43% of RetinaFace and 46.59%of YOLOv7.The integration of AI algorithms into the open UAV platform is explored, showcasing thepotential for real-time face detection and verification during UAV operations. The researchalso addresses ethical considerations and privacy concerns related to facial verification technologies,emphasising the importance of compliance with regulations such as the EuropeanUnion (EU) AI Act and the General Data Protection Regulation (GDPR).The results demonstrate significant improvements in face verification accuracy andinference speed, validating the proposed methodologies and system architecture. Thisresearch has successfully achieved all its objectives, resulting in a novel, cost-effective, open UAV platform with advanced AI capabilities. This work not only advances the field ofUAV-based biometric verification but also opens new avenues for research and practicalapplications in security, surveillance, and beyond, offering a cost-effective and adaptablesolution for face verification tasks. The open-source nature of the platform encourages furtherresearch and development in this field, promoting collaboration and innovation in UAV-basedfacial verification systems.