Video analytics is software, running inside the camera or on a central server, that turns raw footage into events and answers: a person crossing your fence line at 2 a.m., a count of Saturday's foot traffic, the red pickup that left the yard at 4:17. As of 2026 the marketing has outrun the technology, and every camera brochure says AI on it somewhere. This guide sorts the five features that earn their cost on commercial systems from the badges that do not.
Worth paying for: person and vehicle classification, line-crossing and intrusion zones, purpose-built license plate reading, people counting, and search by attribute. Mostly fluff: consumer-grade facial recognition claims and an AI sticker on cheap hardware with no processing behind it.
Which AI camera features actually matter?
- Person and vehicle classification. The workhorse. The camera distinguishes a person from a deer, headlights, rain, and a flag in the wind, which cuts false alerts from a nightly flood to a trickle. It is also the feature that makes remote video monitoring affordable, since operators respond to classified events instead of every pixel change.
- Line-crossing and intrusion zones. Draw the fence line or the after-hours zone on screen and alert only when it is crossed, in the direction and during the hours you choose. Simple, reliable, and the backbone of perimeter protection.
- License plate reading. Real LPR works, and it requires a purpose-built LPR camera with the shutter speed, infrared, and mounting angle for the job, budgeted per lane. A normal camera pointed at a gate will not read plates at night; it will record a glowing white rectangle where the plate should be.
- People counting and occupancy. An operations feature more than a security one: traffic by hour, conversion against sales, staffing to the pattern instead of the guess.
- Search by attribute. The quiet time-saver. Ask the system for every red truck since Friday and get answers in seconds, instead of scrubbing hours of timeline for one frame.
Which features are mostly fluff?
Two categories, as of 2026. First, consumer-grade facial recognition claims. Reliable face matching needs controlled angles, controlled lighting, and serious processing; a cheap camera watching a parking lot has none of the three, so the feature demos well and then misidentifies your own staff all winter. Real deployments are controlled-entry projects with legal questions attached, not a checkbox. Second, the AI badge on cheap hardware. A $60 camera advertising a dozen analytics has no processor behind the sticker, and the features fail exactly when you need them testifying for you. Treat analytics claims as hardware claims: they are only as credible as the chip and the OEM behind them, the same provenance problem covered in our NDAA compliance guide.
Edge or server: where should analytics run?
Edge analytics run inside the camera: no extra hardware, no single point of failure, and the system scales one camera at a time. The cost is per camera, and weak edge chips do weak analytics, which is where the fluff problem above comes from. Server or cloud analytics run centrally: more horsepower, deeper search across every camera at once, one place to upgrade, and they can add intelligence to simpler cameras. The cost is the central box or subscription, and every stream has to reach it. Small sites usually land on good edge cameras; larger and search-heavy sites justify central processing; plenty of systems sensibly mix the two.
What about privacy and Missouri law?
Analytics do not change your legal obligations, but they raise the stakes on them, because a system that classifies, counts, and searches people is doing more than passively recording. Camera placement rules, employee notice, and especially audio recording (where Missouri law is stricter than most owners assume) all apply before the first alert fires. Read our Missouri recording laws guide before you spec analytics, not after.
How should you buy analytics?
Buy analytics to answer a defined operational question, not to fill a feature checklist. Good questions: who is in the yard after close (classification plus intrusion zones, and possibly monitoring), which vehicles came through the gate (LPR), how many customers walked in Saturday (counting), where did that ladder go (attribute search). If a feature does not answer a question you actually have, it is a line item, not a capability, and we would rather sell you three analytics you will use than a spec sheet of forty you will not. Our manufacturer-trained technicians tune zones, schedules, and sensitivity at commissioning, because analytics that were never tuned are the ones that get turned off by Christmas. The site assessment that starts that conversation is free.
Frequently Asked Questions
No, but person and vehicle classification eliminates most of the classics: deer, rain, headlights, moving shadows. Expect a large reduction plus some tuning, not perfection. What remains is small enough for a human to act on, which is what makes monitored systems work.
Almost certainly not at night. Plate capture requires a purpose-built LPR camera with fast shutter, infrared, and the correct angle over each lane. A general-purpose camera at the gate records a glowing rectangle where the plate should be.
No, not in its consumer-grade form as of 2026. The claims rarely survive real-world angles and lighting, and the privacy exposure is real. Classification, intrusion zones, and attribute search deliver the practical value at a fraction of the trouble.
Often, yes. Some features are built into better cameras; others carry per-camera or annual software licensing. A quote should state which, plainly. Buying the two or three features that answer your question usually costs far less than buying the checklist.