Why Vörðr needs so little hardware.
Vörðr never decodes the video it records, and it runs AI only when motion needs confirming. Every camera records continuously, and an event only changes how long the footage around it is kept.
Recording copies the stream
Vörðr writes each camera's main and sub streams to disk exactly as the camera encoded them. Recording uses very little CPU because nothing is re-encoded, and you can play back the file that is still being written.
Motion detection uses the small stream
Only the sub stream is decoded, at 320×180 in grayscale and five frames a second, and the motion detector runs on that. When it finds motion, the object detector checks full-resolution frames to confirm what moved. If the detector is unavailable, recording and motion detection keep running.
Events decide what is kept
By default, sub stream footage is kept for 90 days and HD footage for 15 minutes. Confirmed events keep their HD footage for 30 days, and events the AI rejects release it within hours. Because recording never stops, the clip always includes the moments before the motion started.