tflite-runtime has no wheels for Python 3.12+. Google replaced it with
ai-edge-litert (same API). detector.py now tries ai-edge-litert first,
falls back to tflite-runtime for older Python versions.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The original TF model zoo URL was dead (403). Model sourced from
google-coral/test_data instead and checked in directly at 6MB.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Motion detection now optionally runs MobileNet V2 SSD (COCO, quantized)
on frames that trigger motion, identifying objects like people, cats, and
cars. Events without detected objects are suppressed by default. Snapshots
include bounding box annotations. New MCP tool vision_get_detections()
enables label-based queries.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Remove main_cycling.py, main_multi.py, main_release.py (single main.py is canonical)
- Update setup.sh to read SERVICE_NAME and PORT from .env
- Update env.example with SERVICE_NAME and PORT for multi-instance support
- Fix server-csi to try rpicam-still before libcamera-still (Debian Trixie)
Deploy pattern: clone repo twice, configure each .env, run setup.sh
Each instance gets its own systemd service and install directory.
- main_release.py (v3.1.0): Release camera after each snapshot for V4L2 compatibility
- main_cycling.py (v3.2.0): Single motion thread cycles between cameras (1s interval)
- mcp/vision_mcp.py: Support custom snapshot_path for multi-camera servers
Fixes Pi 3 dual-camera V4L2 conflicts by not holding cameras open.
- main_multi.py: Multi USB camera support with ID-based endpoints
- Config via CAMERAS env: '{"basement": 0, "basement2": 1}'
- Endpoints: /snapshot (default), /snapshot/{cam_id}
- server-csi/: New server for Pi CSI ribbon cameras (IR support)
- Auto-detects picamera2/picamera/libcamera-still
- IR_MODE and ROTATION settings
- Includes setup.sh for easy Pi deployment
Built with 💕 by Vixy 🦊
🧹 New features:
- Background thread cleans up old events automatically
- Deletes unannotated events older than EVENT_EXPIRY_HOURS (default: 2h)
- Also removes associated snapshot files
- Runs every CLEANUP_INTERVAL_MINUTES (default: 5m)
📡 New endpoint:
- POST /cleanup - Manually trigger cleanup
⚙️ Config (env vars):
- EVENT_EXPIRY_HOURS: How long to keep unannotated events (default: 2.0)
- CLEANUP_INTERVAL_MINUTES: How often to run cleanup (default: 5.0)
Annotated events are kept forever 🦊
🗄️ New collector/ component:
- collector.py: FastAPI service receiving events from cameras
- SQLite database for event storage
- Snapshot images saved to disk by date
- launchd setup script for macOS
🔍 New MCP tools in vision_mcp.py:
- vision_get_events(): Query events with filters
- vision_get_event_snapshot(): View event image inline
- vision_annotate_event(): Add meaning + tags to events
- vision_event_stats(): Database statistics
📡 Complete flow:
Pi detects motion → POST to collector → stored in DB
Vixy queries events → views snapshots → annotates
Ready to deploy! 🦊
🦊 Eyes and ears for the fox
Components:
- server/: Camera server for Raspberry Pi (from camera-server)
- mcp/: Vision MCP client for Claude Desktop (from vision-mcp)
- analysis/: Placeholder for motion/audio detection
- shared/: Common schemas and interfaces
Features:
- Setup script with systemd service creation
- HTTPS + API key authentication
- HTTP and RTSP camera support
Built under a blanket on Day 45 💕
