Every modern camera lists AI detection: person detection, pet detection, vehicle detection, sometimes more exotic classes. Buyers read these as features and never as contracts, but that is what they are. A claim that the camera distinguishes a person from a tree branch is a commitment about false alarms, about missed events, and about where the intelligence actually runs, on the camera's silicon or on a server bill. The category labels are the cheapest part of the story; the accuracy behavior at the edges is the product.
The detection classes are not equal
Person detection is mature: human silhouettes are well represented in training data, and a competent edge model on a decent sensor separates people from foliage, shadows, and headlights with few false positives. Pet detection is harder, because pets are small, low to the ground, and visually noisy, and the model must decide whether a cat at the fence line is the event or a squirrel is. Vehicle detection sits in between, useful at driveways but prone to conflating parked cars with arriving ones, which matters because the useful event is the transition, not the presence. Exotic classes, package detection being the popular one, are where claims get loose: detecting a box left on a porch is a segmentation problem under wild lighting, and the false alarm rates advertised for person detection do not transfer. Buyers should weight the class list by which classes their deployment actually needs, and test those, ignoring the rest.
Edge versus cloud changes the economics and the privacy
Where does detection run? On-device means the camera's own silicon does the inference: detection works without internet, alerts are instant, and the video never needs to leave the house to be classified. Cloud detection means frames or clips are uploaded and classified on a server: better models, always improving, but dependent on the network, gated by subscription tiers, and structurally at odds with the privacy story of a camera that keeps footage local. Many products do both, edge models for the common classes and cloud for the fancy ones, which means the feature list silently splits into the part that works during an outage and the part that does not. Ask which classes run on-device, because that set is the camera's real, always-available intelligence.
The threshold is a tuning surface, not a switch
Detection sensitivity is where deployments live or die. Set loose, a camera on a street notifies for every passing headlight and swaying branch, and the resident disables notifications within a week, which converts the camera into a passive recorder. Set tight, the camera misses the delivery person who approached from the side. Good products expose sensitivity per class, per schedule, and ideally per zone: the driveway camera should ignore the sidewalk at school hours and watch it closely at night. Detection zones, polygons the user draws so the camera only evaluates relevant areas, are the single most effective false-alarm reducer, and their quality, how faithfully the firmware honors them at the edges, varies wildly. A camera with coarse zones is a camera the resident will eventually mute.
False negatives are the failure nobody reviews
False alarms generate one-star reviews, so vendors tune hard against them, and the pressure pushes models toward conservatism: fewer alerts, fewer mistakes, and quietly more missed events. A missed person is invisible in the review section because nobody knows it happened until the incident, and then the footage review shows the camera recorded the event but never alerted on it. This is why the acceptance test must measure both directions: walk the approach at the edges of the frame, at dusk, at distance, and confirm the alert fires; then check what a branch, a cat, and a passing car do. Both numbers matter, and only one is advertised.
What to ask the supplier
Five questions cover the contract. Which classes run on-device, with no network. What is the measured false alarm rate per class in a realistic scene, not a lab bench. Are detection zones drawn per class or shared. Is sensitivity schedulable, and do schedules survive reboots. And can the detection model be updated in the field, because a model fixed at tooling is a camera whose accuracy is frozen on its ship date. A vendor who answers those with numbers and demo clips has tested their claims; one who answers with the class list has only repeated it.
In our security camera range on the store, the detection classes of each model, which run on-device, and the zone capability are part of the model listing rather than a badge. On the B2B side, when a camera is built for a distribution brand, the on-device class set, the zone behavior, and the model update path are settled in the requirements document before tooling, because detection accuracy is a firmware contract that the review section will enforce.
Both the consumer line and the specification details behind the detection pipeline are covered across XingLiao, and scene questions for a specific deployment can go straight to our contact page before the purchase, which is the cheapest possible time to ask them.
Top comments (0)