SecureProctor AI

AI-based interview/exam proctoring system with:

• Live webcam monitoring
• Face-count checks (NO FACE, MULTIPLE FACES)
• Head-pose + gaze checks (LOOKING AWAY)
• Real-time web dashboard and violation log
• Audible alert (buzzer) on violations

This project runs locally and serves a browser UI from FastAPI.

Features

• External camera support with robust backend probing (DSHOW, MSMF, ANY)
• Camera selection from UI (/cameras, /set_camera/<index>)
• Real-time status polling and MJPEG video stream
• Alert engine with timeout smoothing
• Trained CNN model support from local weights (eye_cnn_final.pth)
• MediaPipe + automatic OpenCV fallback when MediaPipe fails

Project Structure

• server.py: Main FastAPI server (recommended entrypoint)
• proctor_live.py: Core proctoring engine
• index.html, style.css, app.js: Frontend dashboard
• eye_cnn_model.py: CNN model definition
• eye_cnn_final.pth: Final trained gaze model used during proctoring
• pretrained_eye_cnn.pth: Stage-1 pretrained weights
• dataset_loader.py, train_pretrain.py, train_finetune.py: Training pipeline
• people_detector.py: YOLO-based multi-person utility
• capture_dataset.py, collect_images.py: Dataset collection utilities
• PROCTORING_SYSTEM_EXPLAINED.md: Extended architecture explanation

Requirements

Install all dependencies:

pip install -r requirements.txt

Main dependencies:

• FastAPI
• uvicorn
• pydantic
• Flask
• flask-cors
• opencv-python
• torch
• numpy
• mediapipe
• ultralytics

Quick Start (Windows CMD)

cd /d E:\ritesh_exp\proctoring_fastapi
python -m pip install -r requirements.txt
python server.py

Open:

• http://127.0.0.1:5000

How to Use

1. Open UI in browser.
2. Wait for camera list to load.
3. Select your external camera index.
4. Click Start Proctoring.
5. Monitor SAFE/ALERT state and violation log.

API Endpoints

Core Control

• GET /start → start proctoring thread
• GET /stop → stop proctoring thread
• GET /status → current status JSON (SAFE or ALERT + reason)

Camera

• GET /cameras → list camera indices that can actually provide frames
• GET /set_camera/<index> → switch active camera

Video

• GET /video_feed → MJPEG stream
• GET /frame → single JPEG frame

Detection Logic (Current)

Face / Presence

• NO FACE when no face is detected
• MULTIPLE FACES when more than one face is detected

Head Pose + Gaze

• Uses MediaPipe FaceMesh when available
• Computes head orientation (HEAD LEFT/RIGHT/UP/DOWN)
• Uses CNN (eye_cnn_final.pth) for gaze-like classification from face crop
• Triggers LOOKING AWAY when smoothed predictions cross threshold

MediaPipe Fallback

If MediaPipe import/init fails:

• System falls back to OpenCV Haar face detection
• Proctoring continues (no crash)
• NO FACE, MULTIPLE FACES, and CNN-based LOOKING AWAY still work
• Head-pose checks are unavailable in fallback mode

Model Usage

proctor_live.py loads:

• Model class: EyeCNN from eye_cnn_model.py
• Weights: eye_cnn_final.pth from the same folder

So yes, your local trained model is used at runtime.

Training Pipeline

Stage 1: Pretraining

python train_pretrain.py

Outputs pretrained_eye_cnn.pth.

Stage 2: Fine-tuning

python train_finetune.py

Loads pretrained weights and writes eye_cnn_final.pth.

Dataset Utilities

Capture face crops to dataset/good and dataset/bad

python capture_dataset.py

Collect full-frame images into my_dataset

python collect_images.py

Troubleshooting

External camera not opening

• Close Zoom/Teams/Meet/OBS or any app locking webcam.
• Check available cameras:
  • Open http://127.0.0.1:5000/cameras
• Set camera manually:
  • http://127.0.0.1:5000/set_camera/1 (replace index)
• Keep webcam connected before starting server.

MediaPipe error (module 'mediapipe' has no attribute 'solutions')

• Install dependencies from requirements.txt.
• Runtime now automatically falls back to OpenCV if MediaPipe fails.

Import "flask" could not be resolved in editor

• Usually VS Code interpreter mismatch.
• Select the same interpreter where you installed requirements.

Notes

• server.py is the preferred backend entrypoint.
• app.py is an older Flask server variant kept for compatibility.
• Buzzer uses winsound, so alert sound is Windows-oriented.

License

No license file is currently included in this repository.