2026 · Solo · Personal project

seizureGuard

Vision pipeline that watches a dog through a home camera 24/7 and sends a Telegram alert with a video clip when the motion looks like an epileptic seizure. Five stages, each cheap enough to feed the next: frame-difference motion scoring with a state machine (2 s to arm, 5 s pre-roll so the onset is never lost), event sampling at 2 fps for context plus 10 fps bursts around motion peaks, an optional local YOLO-pose gate that keeps only limb oscillation in the 1.5–4.8 Hz clonic band — strictly fail-open — and two-tier verification by a vision model looking for paddling, rigid posture and rhythmic jerking. The alert itself is a 5–10 second clip cut from the window the verifier was most confident about, taken from a full-rate recording kept alongside the sparse analysis frames, so you see the motion instead of one ambiguous frame. Evaluated on published veterinary footage: 4/4 seizure videos detected, including one documented as missed by a commercial detector, with 6/7 normal clips rejected. Runs on a Raspberry Pi 5. A hobby prototype, not a medical device.