Smart Surveillance System

Overview

Smart Surveillance System is a desktop application for real-time face detection and recognition.
It frees security personnel from routine monitoring by automatically detecting, identifying, and alerting on “wanted” individuals in live camera feeds.

Key Features

• Real-time face detection (MTCNN, Dlib HOG, Haar-cascade)
• Face embedding via FaceNet (128-dimensional vectors + triplet-loss)
• Face classification with SVM (trained on embeddings)
• Add new suspects: capture name, metadata, profile image
• “Align” & “Train” workflows to preprocess and fit new faces into the model
• User management with login (SQLite backend)
• Simple PyQt5 GUI: Start/Pause stream, Add/Search/Train suspects, Manage accounts
• Instant pop-up of recognized suspect’s full profile

System Requirements

Hardware
• Logitech (or equivalent) webcam, 30 fps, ≥340×240
• PC with ≥2 GHz CPU, ≥8 GB RAM, GPU recommended for TensorFlow

Software
• Python 3.7+
• OpenCV
• TensorFlow
• MTCNN (npy models)
• Dlib
• PyQt5
• SQLite3
• NumPy, SciPy
• scikit-learn

Installation

1. Clone the repo:
2. Create and activate a virtual environment:
3. Install dependencies:
4. Download pretrained MTCNN (det1.npy, det2.npy, det3.npy) into ./models/mtcnn/
5. Download FaceNet model (20180402-114759.pb) into ./models/facenet/
6. Initialize the SQLite database:

Usage

1. Launch the GUI:
2. Login with an existing account or create a new one via the “Add Account” tab.
3. Add Suspect: enter name, gender, race, birth year, criminal record, profile image & training folder.
4. Align: preprocess all raw images into aligned faces.
5. Train: build/update the SVM classifier (classifier.pkl) on all aligned embeddings.
6. Start: begin live video stream; recognized faces will trigger a pop-up with full profile.
7. Pause: temporarily halt detection/recognition.
8. Search: retrieve any suspect’s stored profile by ID, name, or race.

Algorithmic Details

1. Face Detection
   • MTCNN: 3-stage CNN (P-Net, R-Net, O-Net) for fast, accurate bounding-box + landmarks
   • Dlib HOG + linear SVM
   • OpenCV Haar cascades
2. Feature Extraction
   • FaceNet maps each face → 128-dimensional vector
   • Triplet loss minimizes same-identity distances, maximizes different-identity distances
3. Classification
   • SVM (linear kernel, probability outputs) over embeddings
   • Decision boundary defined by support vectors

GUI in Action

Testing & Validation

• Unit & integration tests for database, detection, embedding, classification
• Accuracy evaluations on held-out video sequences
• Real-time performance stress tests at 30 fps