On-premises AI for industrial sites: turning existing cameras into an operations platform

Most industrial sites already own the hardware for a computer vision program: hundreds of IP cameras and a recorder that stores footage nobody reviews until something has gone wrong. The gap is the software that turns a live stream into attendance records, access policy, gate compliance, and an audit trail that holds up when someone challenges it. This post covers why that software usually has to run on-premises, what it can do today, and how to roll it out without burying the control room in alerts.

Why industrial sites keep video on-site

Cloud video analytics fits a small retail store. It is a poor fit for a cement plant, a steel mill, or a port terminal, for four reasons.

• The data cannot leave. Footage shows employees, contractors, and license plates. Worker agreements, regulators, and customer contracts often prohibit sending it to a third-party cloud.
• Latency is operational, not cosmetic. A weighbridge hold or a zone-breach alert has to fire while the truck is still on the scale or the person is still in the zone. A round trip to a cloud region does not belong in that path.
• Bandwidth does not scale to hundreds of cameras. Streaming every camera off-site continuously costs more than the analytics are worth.
• The plant does not stop when the internet does. Detection, alerts, and gate decisions need to keep working during an outage.

This is why Eagle AI, our on-premises surveillance product, runs entirely on-site in Docker, ingests the IP cameras a site already has over RTSP, and keeps nothing off the premises. It is built as a unified fork of the open-source Frigate NVR, so the recording layer is proven.

What modern computer vision does beyond recording

Recording is table stakes. The value is in the modules that sit on top of real-time detection and tracking of people and vehicles.

• Workforce attendance. With face recognition, the first sighting of an enrolled employee becomes the check-in and the last sighting becomes the check-out. The roster is imported from the existing HRMS rather than re-entered.
• Zone access policy. A department-by-zone policy matrix defines who is allowed where. A maintenance technician in the kiln area is normal; an office visitor in the same place is an alert.
• Gates and vehicles. License plate recognition, inspection compliance checks, weighbridge weight variance with an automatic hold when a load is out of tolerance, loading bay monitoring, and badge-plus-face identity fusion so a borrowed badge does not open a door.
• Semantic footage search. Instead of scrubbing through hours of video, an operator types "person in orange vest" and gets the matching clips.
• Audio detection. Some events are heard before they are seen.
• Audit and reports. Every event lands in a hash-chained, tamper-evident log that nobody can quietly edit.

The pattern is the same throughout: the cameras become sensors for operations, not just evidence for after the fact.

Accuracy: what to expect and how to enrol people

What matters is accuracy measured live on your cameras, at your angles, in your lighting. On the Eagle AI deployment designed for a cement plant, face recognition with a 12-slot enrolment measured 92–94% confidence live, and license plate recognition produced verified reads at 92–97%. Those numbers set the right expectation: high enough to run attendance and gate workflows, not high enough to remove humans from decisions that matter.

A few enrolment practices make the difference between those numbers and disappointing ones:

• Capture enrolment slots under working conditions: hard hat on, hard hat off, daylight, night lighting, and the angle of the camera that will actually see the person.
• Re-enrol when appearance changes materially, and review low-confidence matches weekly.
• Set thresholds per use case. Attendance can tolerate a borderline match that a human resolves later; a gate release should require the higher threshold or fall back to badge-plus-face fusion.
• Keep a human in the loop for any decision that stops a truck, blocks a door, or affects pay.

Hardware acceleration and sizing

Eagle AI supports hardware acceleration on Coral, OpenVINO, TensorRT, ROCm, Hailo, and RKNN, which means a site can choose between an existing GPU server, low-power edge accelerators, or a mix. The industry is moving the same way: August 2026 coverage of physical AI highlighted edge boards such as NVIDIA's Jetson Orin Nano 2 for drones and robots, the same class of hardware that makes on-site inference practical for cameras.

Not every camera needs full-frame-rate detection. Group cameras by criticality, run gates and restricted zones at high rates, and sample the rest. Prove the accelerator on twenty cameras before ordering for three hundred.

Alert fatigue and triage

Sending every detection to a phone kills a surveillance program within weeks. Eagle AI ships 17 alert types split into critical, warning, and info, with storm-guard de-duplication so one event seen by five cameras is not five pages. The tooling helps, but the policy is yours:

• Critical alerts go to a named person and require an acknowledgement.
• Warnings land on the dashboard and are reviewed at shift change.
• Info events feed reports and are never pushed.
• Review alert volume by type every week for the first quarter, and tune zones and thresholds until the critical channel is quiet enough that people still trust it.

A rollout plan that survives contact with the plant

1. Baseline. Record how attendance is captured today, how long gate processing takes, and how many incidents were logged last quarter.
2. One gate, one zone. Twenty cameras, plate recognition at one entrance, one restricted zone with a policy. Run for a month against the baseline.
3. Attendance for one department. Run it alongside the existing method for a full pay cycle before anyone's pay depends on it.
4. Weighbridge and loading bays. Add the auto-hold only after the weight variance thresholds have been validated against manual records.
5. Hand over the audit. Train HR and security on the reports, then expand zone by zone.

What to do this quarter

• Inventory your cameras: count, RTSP availability, resolution, and which ones cover gates and restricted zones.
• Write down the data-residency constraints from HR, legal, and your largest customers.
• Pick one gate and one zone for a pilot and define the metric you will compare against.
• Decide who owns critical alerts on each shift.
• Ask HR for the roster export format so that enrolment does not become a manual project.

How XISLABS approaches this

We built Eagle AI surveillance for exactly this environment: a cement plant with hundreds of cameras and multiple security zones, where nothing can leave the site. The same practice applies to any computer vision development engagement: baseline, instrument, evaluate accuracy on the customer's own cameras before launch, keep humans in the loop by design, and operate the system after go-live. Our AI integration services connect the vision layer to an existing HRMS, access control, or ERP, and AI consulting helps sites still deciding what to pilot. If that sounds like your plant, talk to us.

Sources

• AI news, September 2026 (blog.mean.ceo)
• AI Weekly, 2 September 2026