Challenge
A motion alert alone does not explain what actually happened. The goal is to connect independent door and motion observations with available camera clips and a readable description, without letting missing footage or an unverified model output create a false sense of certainty.
Solution
SONOFF sensors communicate through a Zigbee bridge and Mosquitto/MQTT with a Mac Studio. A separate motion-triggered camera records clips to a Synology NAS; a read-only watcher imports them for local processing. YOLO26s with ByteTrack on Apple MPS detects and tracks objects and stores evidence frames. In a separate, manually triggered experiment, Qwen3-VL drafts Czech descriptions of selected frames. The dashboard distinguishes sensor events, detections and VLM drafts; a completed human review determines which sentences become the main description. AI processing and storage remain local, while Cloudflare provides protected web access.
Outcome
- One view of the door state, sensor event history, available videos and detection results.
- Two separate data paths: sensors over MQTT and camera clips through Synology into local processing on a Mac.
- Automatic YOLO analysis of new MP4 clips with evidence frames; experimental VLM descriptions remain manually triggered and reviewed.
- Traceable outputs: human corrections are separate from the original model, and a candidate video time overlap is not presented as a confirmed relationship.
