Cameras & Vision Plan – Zero-UI Home Automation (2025)

Purpose

Local visibility with minimal noise. AI filters events; it does not decide core behavior. Avoid motion-only alert spam.

Desired Outcomes

• Wired, local-first cameras (Reolink/Scrypted/Frigate) with HKSV
• AI filtering for person/package/vehicle; default to silence on uncertainty
• Actionable, rate-limited alerts only when needed
• No coupling from camera failures into lighting/occupancy

Signals

• Primary: camera online/offline state; AI classes (person/package/vehicle) with confidence
• Secondary: time of day, arm/away state
• Never use motion-only as a trigger for notifications

Core Behaviors

• Recording and retention local-first; HKSV for family view
• Alerts only on allowed classes when armed/away (or operator rules), with debounce/cooldown
• Annotate events; no automation chains that hinge on camera state

Lighting/Occupancy Interplay

• No motion-only lighting triggers
• Optional: assist lighting on high-confidence person at entry when dark; must respect manual override and cooldowns

Notifications

• Allowed classes only (person/package/vehicle), armed/away bias
• Rate-limit and suppress repeats; silence on low confidence
• No alerts for “camera offline” unless operator-only and once

Fail-Safes

• Camera offline/AI down → silence; do not affect lighting/occupancy decisions
• Stale classifications → discard
• HA/MQTT restart → no alert storms; resume quietly

Rollout / Testing

1. Observe: classify only; no alerts
2. Assist: operator-only alerts for allowed classes with confidence thresholds
3. Alert: enable rate-limited actionable alerts when armed/away

Naming Examples (one job per automation)

• cameras.entry.alert_person_armed
• cameras.driveway.alert_package_daytime
• cameras.health.offline_operator_notice