Hand Recognition — Landmark-based Verifier (Prototype)

What this is

• A complete, ready-to-run prototype for a hand/finger landmark-based biometric verifier using a webcam.
• Uses MediaPipe for 21 hand landmarks, stores landmark vectors per-user, trains a lightweight SVM verifier, and performs live verification.
• Includes simple liveness (motion) check and threshold calibration script.

Contents

• capture.py — capture landmarks from webcam into data/hand_landmarks.csv
• train.py — train an SVM verifier and save models/hand_verifier.joblib
• verify.py — live verification GUI (uses trained model)
• calibrate.py — compute thresholds, EER-style curve and suggested threshold
• utils.py — shared helper functions (normalization, IO)
• requirements.txt — Python packages
• example_data/sample_hand_landmarks.csv — tiny example CSV with two users
• LICENSE — MIT

Quick start (Linux / Windows WSL / macOS)

1. Create virtualenv and install:

   python -m venv venv
   source venv/bin/activate   # or venv\Scripts\activate on Windows
   pip install -r requirements.txt

2. Capture samples for user alice:

   python capture.py capture alice --n 30

   Repeat for other users (e.g., bob).
3. Train:

   python train.py

4. Verify live:

   python verify.py verify alice

Notes & recommendations

• Good lighting and consistent hand pose improve results.
• Normalize landmarks (translation, scale, optional rotation) is performed in utils.py.
• For production: use secure storage for templates, stronger embeddings (siamese nets), and robust liveness (NIR/depth).