🎨 Graffiti Detection AI Model

AI-Powered Graffiti Detection Using YOLOv8

Note

Real-Time Anti-Vandalism Detection System

Instantly detect and alert on any sign of graffiti vandalism across walls, buildings, bridges, and vehicles. Built with YOLOv8 for immediate response and prevention.

A powerful AI-driven system that instantly identifies and alerts on graffiti vandalism, enabling rapid response to prevent urban decay. Deploy on surveillance cameras, mobile apps, or edge devices for continuous 24/7 monitoring and immediate threat detection.

Key Features:

• 🚨 Instant Vandalism Alerts - Immediate detection and notification of graffiti incidents
• ⚡ Real-Time Processing - Sub-second inference for rapid response (<50ms per frame)
• 📹 24/7 Surveillance Integration - Deploy on CCTV cameras for continuous monitoring
• 🎯 Multi-Context Detection - Walls, buildings, bridges, vehicles, trains, and public spaces
• 🔔 Alert System Ready - Integration with SMS, email, push notifications, and security systems
• 📊 Incident Logging - Automatic timestamping and geo-tagging of vandalism events
• 🛡️ Edge Deployment - Run on edge devices (NVIDIA Jetson, Raspberry Pi) for offline operation
• 🌐 API-Ready - RESTful API for integration with existing security infrastructure

Table of Contents

• Requirements
• Installation
• Quick Start
• Real-Time Surveillance
• Alert System Integration
• Training
• Inference
• Evaluation
• Export & Deployment
• Contributing
• Author

Requirements

• Python 3.8 or higher
• CUDA-capable GPU (recommended for training)
• 8GB+ RAM and 10GB+ disk space
• Annotated graffiti dataset in YOLO format

Installation

1. Clone the repository

   git clone https://github.com/yourusername/graffiti-detection-ai-model.git
   cd graffiti-detection-ai-model

2. Create virtual environment

   python -m venv venv
   source venv/bin/activate  # On macOS/Linux
   # or on Windows: venv\Scripts\activate

3. Install dependencies

   pip install -r requirements.txt

Quick Start

1. Prepare Your Dataset

Collect and annotate graffiti images (aim for 1500+ diverse images):

# Organize your raw images and YOLO format labels
python scripts/prepare_dataset.py --data-dir data/raw --output-dir data

Supported Annotation Tools:

• LabelImg - Desktop tool for bounding box annotation
• CVAT - Web-based collaborative annotation
• Roboflow - Cloud-based annotation with AI-assist

2. Train the Model

Basic Training:

python scripts/train.py --data configs/dataset.yaml --model yolov8n --epochs 100

Advanced Training:

python scripts/train.py \
    --data configs/dataset.yaml \
    --model yolov8m \
    --epochs 100 \
    --batch-size 16 \
    --device 0

3. Run Inference

Single Image:

python scripts/inference.py --model models/best.pt --source image.jpg

Batch Images:

python scripts/inference.py --model models/best.pt --source images_folder/

Real-time Webcam:

python scripts/inference.py --model models/best.pt --source 0 --show

Real-Time Surveillance

Deploy on CCTV/IP Cameras

Monitor live camera feeds for immediate graffiti detection:

RTSP Stream Monitoring:

python scripts/inference.py \
    --model models/best.pt \
    --source rtsp://username:password@camera-ip:554/stream \
    --conf-threshold 0.3 \
    --save-detections

Multiple Camera Monitoring:

# Monitor multiple cameras simultaneously
python scripts/multi_camera_surveillance.py \
    --model models/best.pt \
    --cameras cameras_config.json \
    --alert-webhook https://your-alert-system.com/webhook

Edge Device Deployment (NVIDIA Jetson):

# Export to TensorRT for edge deployment
yolo export model=models/best.pt format=engine device=0

# Run optimized inference on Jetson
python scripts/inference.py \
    --model models/best.engine \
    --source /dev/video0 \
    --conf-threshold 0.25

Alert System Integration

Immediate Notification Setup

Configure instant alerts when graffiti is detected:

Email Alerts:

python scripts/inference.py \
    --model models/best.pt \
    --source rtsp://camera-ip/stream \
    --alert-email [email protected] \
    --smtp-config smtp_settings.json

SMS/Push Notifications:

python scripts/inference.py \
    --model models/best.pt \
    --source 0 \
    --alert-sms +1234567890 \
    --alert-service twilio \
    --min-confidence 0.4

Webhook Integration:

# Send detection events to your security system
python scripts/inference.py \
    --model models/best.pt \
    --source rtsp://camera-ip/stream \
    --webhook-url https://security-api.com/graffiti-alert \
    --include-image \
    --geo-tag

Discord/Slack Alerts:

# Example: Discord webhook integration
from detection_alerts import DiscordAlert

alert = DiscordAlert(webhook_url="YOUR_DISCORD_WEBHOOK")
alert.send_graffiti_detection(
    image_path="detections/graffiti_001.jpg",
    confidence=0.87,
    location="Building A, Camera 3",
    timestamp="2026-01-02 14:23:15"
)

Automated Response System

Integration with Security Systems:

• Trigger alarms when graffiti detected
• Activate additional cameras to capture perpetrators
• Log incidents with timestamp and location
• Generate daily/weekly vandalism reports
• Alert nearest security personnel via mobile app

Training

The training pipeline supports multiple YOLOv8 variants with configurable parameters:

Model	Speed	Accuracy	Use Case
YOLOv8n	⚡⚡⚡	⭐⭐	Real-time, edge devices
YOLOv8s	⚡⚡	⭐⭐⭐	Balanced performance
YOLOv8m	⚡	⭐⭐⭐⭐	High accuracy needed
YOLOv8l	🐌	⭐⭐⭐⭐⭐	Best accuracy

Key Training Features:

• 🎨 Advanced augmentation with weather, lighting, perspective transforms
• 📊 Progress tracking via TensorBoard
• 💾 Automatic checkpointing
• 🔄 Resume training from checkpoints
• ⚡ Mixed precision (FP16) for faster training

Inference

Usage Examples

Single image detection:

python scripts/inference.py --model models/best.pt --source image.jpg --conf-threshold 0.25

Save detection crops:

python scripts/inference.py --model models/best.pt --source image.jpg --save-crops

Custom confidence threshold:

python scripts/inference.py --model models/best.pt --source image.jpg --conf-threshold 0.5

Evaluation

Evaluate your trained model on the test set:

python scripts/evaluate.py --model models/best.pt --data configs/dataset.yaml

Metrics Reported:

• [email protected] - Standard object detection metric
• [email protected]:0.95 - COCO-style average precision
• Precision/Recall - Classification accuracy
• F1-Score - Harmonic mean of precision and recall

Export & Deployment

Export your model for different deployment scenarios:

ONNX (Cross-platform):

yolo export model=models/best.pt format=onnx

TensorRT (NVIDIA GPUs):

yolo export model=models/best.pt format=engine device=0

CoreML (iOS/macOS):

yolo export model=models/best.pt format=coreml

Deployment Scenarios:

Deployment	Use Case	Response Time	Best For
🌐 REST API	Centralized monitoring hub	~100ms	Multiple camera integration
📱 Mobile App	Field inspection & reporting	~50ms	Property managers, inspectors
🖥️ Edge (Jetson)	Standalone surveillance	<30ms	24/7 real-time monitoring
☁️ Cloud (AWS/GCP)	Large-scale city deployment	~150ms	City-wide surveillance networks
🏢 On-Premise Server	Private security systems	~80ms	Corporate/institutional security

Quick Deploy Examples:

# Deploy as REST API service
uvicorn api.graffiti_detector:app --host 0.0.0.0 --port 8000

# Docker deployment with auto-restart
docker run -d --restart always \
  --gpus all \
  -p 8000:8000 \
  graffiti-detector:latest

# Kubernetes deployment for high availability
kubectl apply -f kubernetes/graffiti-detection-deployment.yaml

Contributing

Contributions are welcome! Please fork the repository and submit a pull request with your changes.

Areas for Contribution:

• 🎯 Dataset expansion with diverse vandalism examples (tags, murals, scratchiti)
• 🔍 Multi-class detection for graffiti types and severity levels
• 🎨 Segmentation model for pixel-level vandalism mapping
• 📱 Mobile deployment optimization for field inspectors
• 🤖 Automated perpetrator tracking across multiple cameras
• 📊 Vandalism hotspot analytics and predictive modeling
• 🔔 Advanced alert filtering to reduce false positives
• 🌐 Multi-language support for international deployment

Author

Pierre-Henry Soria

Passionate software AI engineer building intelligent systems to solve real-world problems.

☕️ Enjoying this project? Buy me a coffee to support more AI innovations!

License

This project is distributed under the MIT License.

Disclaimer

This model is designed to assist in identifying graffiti for maintenance and urban management purposes. Always respect local laws and privacy regulations when deploying computer vision systems in public spaces.