On-device person and vehicle classification is no longer hard to ship; the hard part is shipping it without training the owner to disable notifications in week two. False-alert discipline is a product feature with its own spec.
The knob set that matters
Zones exclude the pavement, the road and the neighbour’s doorway — the single biggest false-alert reduction. Minimum object size filters distant noise. Schedules let the detector be strict at night and lenient in the afternoon. Cooldown stops one event from becoming fifty notifications. Line-crossing and loitering rules turn classification into intent.
Default to verify, not deter
The siren should not fire on every detection. The sane default sends an alert for human verification, and automated deterrence — siren, lights, voice — sits behind separately configured authorization rules. This default is what keeps the alert channel credible; a device that cries wolf gets muted, and a muted device has no security value at all.
Every event should explain itself
A trustworthy event record carries a unique ID, device and camera identifiers, a UTC timestamp, event type, confidence, the zone or rule that fired, and a snapshot reference. When the owner asks “why did it alert?”, the answer should be readable in the event itself — that is what makes threshold tuning a five-minute job instead of a support call.
Validating detection honestly
Daytime accuracy is easy to demo; dusk, headlights, rain, insects on the lens and waving branches are where detectors actually live. A validation plan that includes nuisance-alert testing at night, with agreed thresholds at design freeze, is the difference between a detector the client can ship and one they quietly downgrade.