Cameras may be installed throughout a store, yet deliver little operational value. A 15-store regional chain has, conservatively, 8 to 12 cameras per store. At standard recording quality, that creates 150 to 200 terabytes of footage per year across the chain. Most is reviewed only after something goes wrong, and then only for a specific incident. The footage is there. Its information is not data because nothing extracts it.
That is the central problem camera-based operations intelligence addresses. It is not a storage issue, a bandwidth issue, or primarily a computer vision issue. The real gap is inference: converting a stream of frames into structured signals that staff can act on without asking someone to watch the video.
Why raw footage lacks operational value
Video contains substantial information, but it is unstructured. A grocery aisle frame can show whether shelves are stocked, whether people are lingering unusually near a display, or whether promotional placement matches the planogram. Those details are encoded in pixels. A person watching could identify them, but people cannot monitor every camera continuously or review everything at once. As a result, the information remains trapped in the footage instead of reaching someone who can respond.
Moving from "footage" to "data" requires inference: a model processes the pixels to answer a defined operational question and create a structured result. A fill ratio estimate, dwell-time value, or compliance flag can be stored, compared with thresholds, sent to alert channels, and combined in reports. Raw footage cannot do those things. The inference output is data; the footage itself is not.
Three inference types from one camera
A ceiling-mounted IP camera covering a shelf aisle performs three roles in our system, each tied to an operational question for that zone.
Shelf fill inference asks whether facing coverage for this bay is above threshold right now. It is a spatial, per-frame question. The model examines the shelf ROI crop and estimates the share of facing area containing visible product. If the result stays below threshold, it creates a restock alert. Inference runs on sampled frames at regular intervals.
Behavioral pattern inference asks whether activity at a location matches known loss patterns. It requires tracking across frames rather than reviewing one frame. Extended dwell near high-value fixtures, repeated product handling without cart movement, and unusual movement toward exits from specific zones are temporal patterns. The model maintains state across a window of frames and checks whether the accumulated behavior reaches the alert threshold. The inference window is typically a few minutes, and no alert is issued until the pattern is confirmed.
Display compliance inference asks whether the current shelf arrangement matches the reference planogram for that zone. It compares the current frame crop with a reference image showing the correct display state. The model detects structural product-placement differences, missing items in important facing positions, and category-level deviations. This runs less often, typically hourly or on-demand during compliance review cycles.
All three inferences use the same camera feed. Each result goes to its relevant channel: restock alerts to floor staff, behavioral flags to loss prevention contacts, and compliance deviations to the operations manager's weekly report.
What gets kept and what does not
The video stream remains on the store's DVR or NVR, just as it did before. The inference layer creates a separate stream of structured events: timestamped records showing what was inferred at which zone, with what confidence, and what decision followed. These events are small enough to send to the cloud, store economically, aggregate across stores, and query for reporting without handling the video.
This separation matters for two reasons. First is data sovereignty. Video that may include customer faces and other privacy-sensitive content stays on the premises. Inference outputs are operational signals about shelf conditions and zone events, contain no personal data, and can move through standard cloud pipelines. Second is scale. Structured events from 15 stores fit within any standard database. Raw video from 15 stores creates a petabyte problem.
The calibration step most systems skip
The distance between a camera feed and dependable operational data is not only a matter of computer vision quality. It is also calibration: defining zones, setting thresholds, and checking that outputs reflect operational reality for a particular store, including its lighting, camera placement, and product mix.
We devote time to calibration during onboarding because that is where reliability is established. A shelf zone that is too broad can undercount fill when an empty facing at one end reduces the measured result for a full facing in the middle. A behavioral threshold set too sensitively can alert on ordinary browsing. A planogram reference image captured under lighting different from the live camera view can create false compliance deviations.
These are calibration failures, not fundamental technical failures. The inference models are functioning as intended. The issue is whether the model inputs, including zone definitions, thresholds, and reference images, are configured to produce results aligned with operational intent.
Calibration is part of onboarding because a system producing unreliable outputs during its first two weeks will likely be ignored in week three. During the first week of a new installation at any location, alerts are generated but not immediately sent to staff. An operations team member reviews them with the related footage clips to verify that they reflect the intended operational meaning. Adjustments are made before alerts reach floor staff. It is not a flashy feature, but it determines whether a system is used or ignored.
What inference cannot do, and should not pretend to
Camera-based inference works when the question is visible within the camera's field of view and the signal is consistent enough to separate from noise. It performs poorly when coverage is obstructed, lighting varies, or the answer depends on information hidden from the available camera angle.
We flag those limits directly. If a display unit blocks the camera sightline, that zone cannot support reliable fill-ratio inference, so we flag it rather than generating an inference we cannot trust. If a behavioral use case involves a product category at a fixture outside every camera's coverage, the existing hardware cannot support it. Extending coverage may be appropriate, but that is a separate discussion. Our role is to maximize the value of cameras already installed while stating clearly where coverage gaps limit what inference can deliver.
The aim is not a system that claims to know everything. It is a system reliable enough that staff act immediately when an alert appears, and an operations manager can trust that no alert means the covered zone is in a normal state. That trust depends on accurate inference in covered zones and honest silence where coverage is insufficient.