Facetrack: Personalized information retrieval through face recognition

Ooi, Yunn Suen (2025) Facetrack: Personalized information retrieval through face recognition. Final Year Project, UTAR.

Preview

PDF Download (3245Kb) | Preview

Abstract

This project addressed the challenge of managing and recalling personal information in scenarios involving high-volume interactions, infrequent encounters, and age-related memory decline using face recognition. Existing face recognition solutions were primarily limited to government applications, such as national identification systems, suspect tracking, and border control [1], [2], highlighting the need for accessible, personalized tools for everyday users. The system employed OpenCV and YuNet for face detection and movement analysis with a sharpness threshold above 1.8, while DeepFace (FaceNet512) enabled face matching with a similarity threshold of 0.7. WebSocket facilitated low-latency communication between the frontend and backend, supported by user interaction modules with form validation and notifications, and PostgreSQL handled database operations. The system was tested across multiple stages, including initialization, detection, recognition, user interaction, and database management. It achieved an end-to-end latency of 2–5 seconds and a recognition accuracy of 98.54% for known faces under controlled lighting. The novelty of this system lay in its movement-adaptive processing, which cleared the recognition queue during rapid motion. This allowed the system to quickly process new faces entering the camera view, as high movement often indicated a change in person. Additionally, its sharpness-based image capture mechanism ensured that only high-quality images were processed. Combined with WebSocket-driven pause/resume functionality, these features enabled a seamless and uninterrupted user experience. Overall, the system demonstrated the feasibility and scalability of a personalized face recognition solution for everyday environments such as kindergartens, hotels, and events.

Actions (login required)