AI Smart Proctoring System

AI-powered exam proctoring using YOLOv8 and OpenCV for suspicious activity detection, prohibited-object detection, evidence logging, and integrity reporting.

Overview

The system watches a webcam feed or a recorded exam video, runs each frame through a YOLOv8 object-detection model, and applies rule-based suspicion scoring to flag potential cheating — no face/person in frame, multiple people in frame, or a prohibited object (book, cell phone, headphone). Once enough suspicious frames accumulate, it saves a CSV of all suspicious records plus the top-5 highest-scoring evidence frames as images.

For the full step-by-step breakdown of how a frame moves from capture to saved evidence, see WORKFLOW.txt.

Sample Detections

Random sample of annotated detections from the validation set

Model inference on a single frame

Model Performance (best.pt, validated on held-out data)

Class	Precision	Recall	mAP50	mAP50-95
book	0.879	0.956	0.959	0.663
cell phone	0.902	0.946	0.933	0.525
headphone	0.825	0.856	0.850	0.377
person	0.955	0.970	0.980	0.717
All (mean)	0.890	0.932	0.931	0.570

Trained: YOLOv8n, 100 epochs, image size 640, batch size 16, on a 4-class dataset (book, cell phone, headphone, person). Full training log and validation output are in Notebook/object_detection.ipynb.

Project Structure

AI-Smart-Proctoring-System/
├── APP/
│   ├── config.py            # paths, thresholds, class list
│   ├── video_source.py      # webcam/video capture + frame timing
│   ├── detector.py          # YOLO model loading + per-frame detection
│   ├── suspicion.py         # rule-based suspicion scoring
│   ├── evidence_manager.py  # CSV + evidence frame saving
│   └── main.py               # entry point
├── model/
│   └── yolo_v12.pt           # trained YOLOv8n weights
├── Notebook/
│   └── object_detection.ipynb  # dataset prep, training, evaluation
├── Sample images/             # example annotated/detection images
├── reports/
│   ├── cheating_evidence/     # saved evidence frames (generated)
│   └── suspicious_records.csv # suspicious-frame log (generated)
├── testing_vid/                # sample test video (ignored in git)
├── WORKFLOW.txt                # full pipeline walkthrough
└── README.md

How It Works

1. Choose a source — webcam or a video file.
2. Frames are sampled at a fixed processing rate (PROCESS_FPS).
3. Each sampled frame is run through the YOLOv8 model.
4. Suspicion rules check for: no person, multiple people, or a suspicious object (book / cell phone / headphone).
5. Suspicious frames are logged; evidence-eligible frames are buffered.
6. Once the suspicious-frame count crosses a threshold, cheating is flagged, and the CSV + top evidence frames are written to reports/.

See WORKFLOW.txt for the complete diagram and explanation.

Getting Started

git clone https://github.com/ENGABHAY/AI-Smart-Proctoring-System.git
cd AI-Smart-Proctoring-System
pip install -r requirements.txt
python APP/main.py

You'll be prompted to choose a webcam or a video file. Press Q to stop processing at any time.

Tech Stack

• Python
• OpenCV
• Ultralytics YOLOv8
• Pandas

License

See LICENSE.