Smart Intercom Unlock System

Computer Vision system for automatic door unlocking using face recognition and UI automation.

Overview

This project implements an end-to-end pipeline that:

1. Captures live video from an emulator window
2. Detects a face in the frame
3. Computes face embeddings using a deep neural network (ArcFace ONNX)
4. Compares embeddings with a reference identity
5. Detects the "Open Door" button using template matching
6. Automatically clicks the button when the authorized person is detected

---

Pipeline

Screen Capture (mss)
        ↓
Face Detection (OpenCV Haar Cascade)
        ↓
Face Embedding (ArcFace ONNX)
        ↓
Similarity Check (Cosine Similarity)
        ↓
Button Detection (Template Matching)
        ↓
Automation (pyautogui click)

---

Tech Stack

• Python
• OpenCV
• ONNX Runtime (GPU supported)
• NumPy
• MSS (screen capture)
• PyAutoGUI (automation)
• PyGetWindow

---

Emulator-Based Integration

The system operates through a mobile application running inside an Android emulator (e.g., BlueStacks / LDPlayer).

Since no public API is available for the intercom system, the solution uses:

• Screen capture of the emulator window
• Computer Vision for video stream processing
• UI automation to interact with the app

Why emulator?

• No official API access
• Closed mobile application
• Need to automate real user interaction

This approach allows building a fully functional system without modifying the original application.

---

Setup

1. Clone repository

git clone https://github.com/Mahan1341/Home_FaceRecognition.git
cd Home_FaceRecognition

---

2. Install dependencies

pip install opencv-python numpy mss pyautogui pygetwindow onnxruntime-gpu

---

3. Download models

Download ArcFace model:

https://github.com/deepinsight/insightface/releases

Required file:

w600k_r50.onnx

Place it into:

models/w600k_r50.onnx

---

4. Prepare your face data

Add your image:

me_1.jpg

---

5. Run

python main.py

---

Configuration

You can tune thresholds in code:

FACE_THRESHOLD = 0.5
BUTTON_THRESHOLD = 0.8
GLOBAL_COOLDOWN = 20

---

Notes

• Face recognition is based on cosine similarity between embeddings
• Thresholds are chosen empirically
• System works in real-time

---

Limitations

• Haar Cascade face detector is not robust
• No anti-spoofing (photo/video attacks possible)
• UI detection depends on button appearance

---

Demo

TODO

---

Project Motivation

This project demonstrates:

• Computer Vision
• Deep Learning inference
• Real-world system integration
• Automation

---

Author

Mahan1341