Mike — Face Recognition Project

Face Recognition

A Python application for real-time face recognition from a webcam. The system uses OpenCV for video capture, drawing bounding boxes, and rendering the interface, while the face_recognition library powered by dlib handles face detection and computes face encodings.

The project is designed for autonomous operation: new faces automatically receive a UUID, are saved as Unknown, get their first photo, and store an initial embedding. The system keeps a global event history, separate logs for each person, and can accumulate new encodings over time to improve recognition stability.

Category :Computer Vision / Python
Stack :Python, OpenCV, face_recognition, dlib, pickle
Mode :Offline / Real-time Webcam
Data Store :known_faces.pkl + faces/<user_id>/
Identity Model :UUID + Auto-learning Encodings
Controls :q / w / d
Status :Portfolio Concept / Implementation Ready

Core Features

The system is designed as a single file with clear functions, making the code easy to move, run, and extend without a complex project structure.

Frame capture from the webcam and on-screen rendering via OpenCV.
Face detection and encoding comparison through face_recognition.
Automatic creation of a new ID and saving the first face as main.jpg.
Attempting to restore the same ID by position when the encoding temporarily fails.
Updating a profile with new encodings when the face matches but the angle is new.

"The goal of the project is not just to recognize a face in a frame, but to maintain a stable identity for the same person across frames, even when the model receives noisy or incomplete embeddings."

Design Note - Identity Persistence

Architecture

The known face database is stored in known_faces.pkl: for each UUID, it keeps the name, a list of encodings, the first and last seen timestamps, and service metadata used for matching. On each frame, the system gets face coordinates, computes a face encoding, compares it with existing profiles, and selects the best match by distance. If no direct match is found, a proximity check against the most recently seen position is used to restore the previous ID during short recognition failures. Visibility time is calculated as the interval between the first detection of a face in the current presence session and the last frame where that face is still visible.