shield_lock
JinDaGe
← Back to skills
extension
Category: Content & MediaNo API key required

vision-monitor

Vision-based monitoring system using camera + AI to detect specific objects or activities. Features: real-time photo capture, visual analysis via GLM-4V-Flash, customizable detection targets (spoons, cups, dangerous activities), Feishu notification alerts. Actions: monitor, detect, watch, alert, capture, analyze.

personAuthor: jakexiaohubgithub
open_in_newRepositorydownloadDownload package

Vision Monitor Skill

AI-powered visual monitoring system using your Mac's camera and GLM-4V-Flash for real-time object/activity detection.

Features

  • 📸 Real-time Capture - Automatic photo capture at configurable intervals (default: 5 seconds)
  • 🔍 Visual Analysis - Uses GLM-4V-Flash to detect specific objects or activities
  • 🎯 Customizable Targets - Monitor for spoons, cups, people, dangerous activities, etc.
  • 📱 Feishu Alerts - Instant notifications when targets are detected
  • 🧹 Auto Cleanup - Keeps only latest N photos to prevent disk overflow

Requirements

  • macOS with built-in camera (FaceTime HD Camera)
  • FFmpeg installed (brew install ffmpeg)
  • GLM-4V-Flash API key (智谱AI)
  • Feishu app (optional, for notifications)

Installation

# Install ffmpeg
brew install ffmpeg

Configuration

API Key Setup

Set environment variables:

export GLM_API_KEY="your-glm-api-key"
export FEISHU_TOKEN="your-feishu-token"  # Optional
export FEISHU_USER_ID="your-user-id"     # Optional

Or configure in ~/.clawdbot/skills/vision-monitor/config.yaml:

glm_api_key: "your-glm-api-key"
feishu:
  token: "your-feishu-token"
  user_id: "your-user-id"

Usage

Start Monitoring

# Monitor for spoons
./monitor.sh --target spoon

# Monitor for cups
./monitor.sh --target cup

# Monitor for dangerous activities
./monitor.sh --target danger

# Custom target
./monitor.sh --target "red phone"

Check Status

# View latest photo
cat /tmp/camera_monitor_latest.jpg | open -f

# View monitoring logs
tail -f /tmp/vision-monitor.log

Stop Monitoring

pkill -f vision-monitor
# or
kill $(cat /tmp/vision-monitor.pid)

Configuration Options

| Option | Description | Default | |--------|-------------|---------| | --interval | Capture interval in seconds | 5 | | --target | What to detect | required | | --max-photos | Max photos to keep | 10 | | --output | Photo output directory | /tmp/camera_monitor |

Example:

./monitor.sh --interval 10 --target "suspicious person" --max-photos 20

Project Structure

vision-monitor/
├── SKILL.md          # This file
├── README.md         # User guide
├── monitor.sh        # Main monitoring script
├── detect.sh         # Detection logic
├── config/           # Configuration files
│   └── prompts.yaml  # Detection prompts
└── scripts/
    ├── capture.sh   # Photo capture
    ├── analyze.sh   # Image analysis
    └── notify.sh    # Alert system

Supported Detection Targets

  • Objects: spoon, cup, phone, laptop, bag, etc.
  • People: count people, specific actions
  • Activities: climbing, danger, theft
  • Custom: Any object or behavior you define

Example Prompts

Spoon Detection

prompt: "请仔细看这张图片。如果有人手里拿着勺子，请说'发现勺子'。如果没有人拿勺子，请说'安全'。"
alert: "🚨 有人拿勺子了！"

Danger Detection

prompt: "请检测是否有人做危险动作，例如爬高、攀爬、站在高处等。如果发现危险动作，请说'危险警报'。"
alert: "🚨 检测到危险动作！"

Troubleshooting

Camera not detected

ffmpeg -f avfoundation -list_devices true -i ""

FFmpeg pixel format error

# Use correct pixel format
ffmpeg -f avfoundation -framerate 30 -video_size 1280x720 -i "0" -pix_fmt uyvy422 ...

API errors

  • Check API key is valid
  • Ensure GLM-4V-Flash has sufficient quota
  • Verify network connectivity

API Credits

Uses GLM-4V-Flash from 智谱AI (BigModel).

  • Free tier available
  • Fast inference
  • Good for real-time monitoring

Get API key: https://open.bigmodel.cn/

License

MIT License

Author

Created for OpenClaw personal AI assistant

Changelog

v1.0.0 (2026-02-03)

  • Initial release
  • Basic monitoring
  • GLM-4V-Flash integration
  • Feishu notifications