🛡️ VisionGuard: AI-Powered Smoking & Restricted Area Compliance System

Real-time cigarette detection and human compliance tracking system built using YOLO26, ByteTrack multi-object tracking, spatial containment, temporal multi-frame confirmation, evidence snapshot generation, and a live web monitoring dashboard.

---

📌 System Architecture

flowchart TD
    A[Webcam / RTSP / Video Feed] --> B[Frame Capture & Preprocessing]
    B --> C[Person Detector YOLO26s + Native ByteTrack]
    B --> D[Cigarette Detector Custom YOLO26 best.pt]
    
    C -->|Person Bounding Box & Persistent Track ID| E[Spatial Association Engine]
    D -->|Cigarette Bounding Box & Confidence| E
    
    E -->|Spatial Containment Ratio >= 0.20 & Ambiguity Margin| F[Candidate Smoking Event]
    
    F --> G[Sliding Window Temporal Confirmation]
    G -->|Threshold Met| H[🔴 CONFIRMED SMOKING VIOLATION]
    
    H --> I[📸 Capture Evidence Snapshot JPEG]
    H --> J[💾 Log Incident in SQLite Database]
    H --> K[📊 Live Web Dashboard Alert & Stream]

Loading

---

📂 Streamlined Repository Structure

Smoking-detection/
├── visionguard/               # Main VisionGuard Core Package & Web Application
│   ├── app.py                 # FastAPI Web Application & API endpoints
│   ├── camera.py              # Decoupled 50+ FPS Camera capture & MJPEG streaming loop
│   ├── detector.py            # YOLO26s Person Tracking + YOLO26 Cigarette Inference Engine
│   ├── draw.py                # Live Video Overlays (HUD, Bounding Boxes, Alert Banner)
│   ├── violation_engine.py    # Spatial Association & Temporal Multi-frame Confirmation
│   ├── db.py                  # SQLite Event Database Connection & Logging
│   ├── models.py              # Database Models & Event Schemas
│   ├── config.yaml            # Tunable System Thresholds & Parameters
│   └── static/                # Live Monitoring Dashboard UI (HTML / JS / CSS)
│
├── person_detector.py         # Standalone Person Detection & ByteTrack Tracking Script
├── train_cigarette.py         # Roboflow Dataset Downloader & YOLO26 Training Pipeline
├── requirements.txt           # Project Dependencies
└── README.md                  # System Documentation

---

⚡ Quick Start Guide

1️⃣ Installation

Ensure Python 3.10+ and PyTorch are installed:

pip install -r requirements.txt

2️⃣ Run VisionGuard Live Web Dashboard

Launch the web application on port 8000:

cd visionguard
python -m uvicorn app:app --host 0.0.0.0 --port 8000

Open http://localhost:8000 in your browser.

3️⃣ Standalone Person Tracker (YOLO26s + ByteTrack)

To run standalone webcam human tracking with persistent IDs:

python person_detector.py

4️⃣ Train Custom Cigarette Detection Model (YOLO26s)

To train or fine-tune a cigarette detection model using the Roboflow dataset:

python train_cigarette.py

---

🌐 Features

• Strict Human-Only Person Detection: Uses pretrained COCO YOLO26s to track human beings with persistent ByteTrack IDs (#1 person, #2 person).
• Precision Cigarette Detection: Trained on 4,000+ labelled smoking dataset images with Small-Target-Aware Label Assignment (STAL).
• Decoupled 50+ FPS Video Streaming: Camera streaming is decoupled from deep inference passes for smooth, lag-free playback.
• Automatic Evidence Snapshots: Automatically captures JPEG evidence snapshots and logs violations to an SQLite database (visionguard.db).

---

Dashboard Screenshots

Live Monitoring	Violation Photos	Analytics