FocusSense 2.0

Real-time, privacy-first study attention analytics using MediaPipe/OpenCV, an optional trained CNN/LSTM classifier, an interpretable focus engine, and a Flask + SQLite dashboard.

MCA minor project, built with Sameer Patel.

Screenshots

Focus monitor	Focus analysis
Student records	Student history
Live focus detection

What changed

• Clean Flask API with session lifecycle and event logging
• Focus score from gaze, head pose, eye state and face presence
• Temporal smoothing to reduce frame-to-frame flicker
• Focused / Distracted / Drowsy / Away / Unknown states
• Optional use of the existing focussense_model.h5; heuristic scoring remains available when TensorFlow/model loading is unavailable
• Live student dashboard, analytics, student history, CSV/PDF exports
• Responsive UI and privacy-oriented local processing
• Evaluation script for measured model metrics

Run

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python init_db.py
python dashboard_server.py

In another terminal:

source venv/bin/activate
python run_focus_live.py --student-id S1 --name "Student 1"

Then open http://127.0.0.1:5000.

Architecture

Webcam -> MediaPipe FaceMesh -> signal extraction -> temporal FocusEngine
                                      |                    |
                                      +-> optional CNN/LSTM +
                                                           v
                                                   Flask REST API
                                                           v
                                                     SQLite database
                                                           v
                                               Dashboard / Analytics / Reports

Model evaluation

Run:

python evaluate_model.py

The script reports metrics computed from the repository's labeled .npy sequences. Do not copy metrics into a report until you have run the evaluation and verified the test protocol.

Privacy

The camera client processes frames locally and sends derived scores/signals to the local dashboard. Raw webcam frames are not stored by the application.

Original project materials

The existing training data, model, PPTs, screenshots and original helper scripts are retained in this release where useful. The live path uses the cleaned 2.0 implementation.

About me

Mohit Raj, MCA graduate from RV College of Engineering. GitHub · LinkedIn · LeetCode