Yellow Car Bot 🚕

An automated bot that monitors Norwegian traffic cameras for yellow cars and posts them to Bluesky with the classic Norwegian phrase "GUL BIL!" (Yellow Car!).

How It Works

1. Downloads images from 814 Norwegian traffic cameras
2. Detects vehicles using YOLO26 real-time object detection
3. Validates yellow color using HSV filtering for accurate yellow car detection
4. Posts to Bluesky with "GUL BIL!" and annotated bounding boxes when a yellow car is found

Note: This bot has been migrated from GPT-4o/OWLv2 (cloud-based AI) to YOLO26 (local inference) for improved speed, cost-efficiency, and reliability. YOLO26 provides ~50x faster detection (100-200ms vs 5-10s per image) while maintaining high accuracy.

Features

• Fair Processing: Shuffles camera order to ensure all cameras get processed at different times of day
• Budget Optimized: Designed to stay within GitHub Actions' 2000 minutes/month limit
• Rate Limit Aware: Gracefully handles API rate limiting
• Persistent State: Remembers progress between runs
• Statistics Tracking: Monitors detection rates and posting success

Setup

1. Repository Structure

your-repo/
├── gulbilbot.py
├── pyproject.toml
├── valid_webcam_ids.txt
├── .github/
│   └── workflows/
│       ├── post.yml
│       └── update-dummy-md.yml
└── README.md

2. Required Files

pyproject.toml and requirements.txt:

requests
python-dotenv
torch
torchvision
Pillow
atproto
ultralytics
opencv-python-headless
numpy

requirements.txt:

torch==2.13.0
torchvision==0.28.0
requests
python-dotenv
atproto
Pillow
ultralytics
opencv-python-headless
numpy

valid_webcam_ids.txt: One traffic camera URL per line (814 URLs total). Example:

https://webkamera.atlas.vegvesen.no/public/kamera?id=3000063_1
https://webkamera.atlas.vegvesen.no/public/kamera?id=3001004_1
...

3. GitHub Secrets

Set up these secrets in your repository (Settings → Secrets and variables → Actions):

• BSKY_HANDLE: Your Bluesky handle (e.g., username.bsky.social)
• BSKY_PASSWORD: Your Bluesky app password (not your main password!)

Note: No API tokens required! YOLO26 runs locally on GitHub Actions runners.

4. Bluesky App Password

1. Go to Settings → Privacy and Security → App Passwords
2. Create a new app password for this bot
3. Use this password (not your main password) in the BSKY_PASSWORD secret

Configuration

Edit these variables in gulbilbot.py:

CONF_THRESHOLD = 0.3          # YOLO26 vehicle detection confidence (lower = more sensitive)
YELLOW_RATIO_THRESHOLD = 0.15 # Minimum ratio of yellow pixels to confirm yellow car
SAMPLE_SIZE = 100             # Number of images to test per run

Supported Vehicle Types: car, truck, bus, van, threewheel

Schedule

The bot runs 3 times per day:

• 6:00 AM UTC
• 2:00 PM UTC
• 10:00 PM UTC

Budget: ~60 minutes/day, ~1800 minutes/month (within 2000 limit)

Coverage: 300 random cameras per run, all 814 cameras covered every ~3 runs

Statistics

Each run tracks:

• Images downloaded successfully
• Yellow cars detected
• Successful Bluesky posts
• Processing time
• Results saved to test_results.txt

Yellow Detection Algorithm

Two-stage process:

1. YOLO26 Object Detection:

   • Detects vehicles in real-time using YOLO26n (Nano variant)
   • Identifies: cars, trucks, buses, vans, and threewheels
   • Confidence threshold: 0.3 (30%)
   • Processing speed: 100-200ms per image
   • ~50x faster than previous cloud-based approach

2. HSV Color Validation:

   • Validates detected vehicle contains sufficient yellow pixels
   • HSV Range: Hue=[15-35], Saturation=[80-255], Value=[80-255]
   • Minimum yellow pixel ratio: 15%
   • Eliminates false positives from non-yellow vehicles

3. Bounding Box Annotation:

   • Draws bright yellow rectangles around detected yellow cars
   • Labels include vehicle type and confidence score
   • Annotated images posted directly to Bluesky

Troubleshooting

Common Issues

"Rate limit reached":

• Bot automatically stops to preserve GitHub Actions minutes
• Will resume in next scheduled run

"Shuffle state not found":

• Normal on first run
• Bot will create new shuffle and start processing

"No yellow cluster detected":

• Most images won't have yellow cars - this is normal
• Adjust YELLOW_THRESHOLD if too sensitive/insensitive

Logs

Check workflow logs for:

• Processing progress
• Detection statistics
• Error messages
• Session summaries

Manual Testing

Trigger a manual run:

1. Go to Actions tab
2. Select "Yellow Car Bot - Budget Optimized"
3. Click "Run workflow"

File Descriptions

• gulbilbot.py: Main bot script (downloads 300 random images, detects yellow cars, posts to Bluesky)
• pyproject.toml: Python dependencies
• requirements.txt: Pinned runtime dependencies for pip-based installs
• valid_webcam_ids.txt: List of 814 traffic camera URLs
• .github/workflows/post.yml: GitHub Actions workflow for the main bot run
• .github/workflows/update-dummy-md.yml: GitHub Actions workflow (updates dummy.md weekly)
• test_results.txt: Results from latest run (auto-generated)

Privacy & Ethics

• Only processes public traffic camera feeds
• No personal data collected or stored
• Images deleted immediately after processing
• Respects API rate limits and terms of service

Contributing

Feel free to:

• Adjust detection parameters for better accuracy
• Add more traffic cameras to the list
• Improve the yellow detection algorithm
• Enhance error handling

License

This project is for educational and entertainment purposes. Please respect:

• Traffic camera usage terms
• Bluesky community guidelines
• API rate limits and quotas

---

Lykke til med GUL BIL-jakten! 🚕