Practical Open-Set Face Recognition: Operator-Centric Design with KD-Tree Search

We introduce a real-time, open-set face recognition framework developed with a strong focus on operator usability and compliance-aware AI deployment. The architecture integrates YOLOv8 detection with ArcFace embeddings (512-D, L2-normalized), followed by KD-Tree retrieval with cosine re-ranking for efficient similarity search. A Flask/SQLite backend with a Next.js dashboard enables live monitoring, role-based access, and human-in-the-loop operations such as unknown flagging, approval, denial, and enrollment. The system was piloted under PDPA-aligned conditions with signed consent, signage, and retention limits. In practice, subjects were consistently rejected as unknown prior to enrollment and subsequently identified reliably, including in eyeglass scenarios. Performance measurements show KD-Tree yields 3−12× lower query times than brute-force cosine search as the gallery scales from 25 to 200 entries, enabling submillisecond responses on commodity CPUs. A usability study involving students, faculty, and security staff reported an overall satisfaction score of 3.67/5, highlighting ease of use and responsiveness, with noted challenges in first-sighting sideprofile views. Current constraints include ∼10 FPS video preview due to transport/rendering and the absence of automated incident grouping, both identified as straightforward enhancements for future iterations.