Shelf-fill detection is a signal chain, not a single prediction: a camera view becomes a region of interest, a fill estimate, a threshold decision, and finally a message for staff. Reliability depends on controlling each stage, from sightline quality to temporal confirmation. This walkthrough covers both how the system works and why its alerts are reliable enough to act on.
Step 1: the camera feed
Youtiligent connects to existing ONVIF-compatible IP cameras over the store's local network. No video leaves the premises. During setup, a lightweight agent on an edge device subscribes to the RTSP streams from the cameras covering the shelf zones selected for monitoring.
Camera placement is a material constraint. A ceiling-mounted camera aimed down an aisle at a 45-degree angle can capture shelf fill reasonably well when the aisle is wide enough and resolution is sufficient. The wrong angle, or a view blocked by a display column, can make inference for particular bays unreliable. Onboarding therefore includes a camera audit: which cameras provide usable sightlines for which zones, and which thresholds fit the camera angle and product category. Zones with inadequate visibility are flagged. Frames that cannot support reliable inference are not used for inference.
Step 2: frame sampling and processing
Continuous inference on every frame from every camera is neither necessary nor efficient. Shelf-fill conditions change slowly compared with video framerate. The system samples at a configurable interval, typically every few minutes for shelf-fill detection, with more frequent sampling for high-velocity products or store-designated priority categories.
Every sampled frame is preprocessed to isolate the shelf region of interest (ROI) defined during onboarding. The ROI is a polygon drawn on the camera image around a specific bay or set of bays. Cropping away the rest of the frame speeds processing and filters unrelated aisle activity, such as a customer walking past.
Step 3: shelf-fill detection
The inference model examines the cropped shelf ROI and estimates fill ratio: the fraction of facing area containing product versus empty facing. This pass does not identify specific SKUs. It asks the simpler question: is the shelf adequately stocked or below threshold?
That distinction is important. Identifying a SKU from a typical store camera at a typical mounting angle is more difficult, less reliable, and not required to trigger a restock alert. Staff do not need a message saying "you are low on SKU 4821003." They need to know "bay 3 of aisle 7 is running low, go check it." Product details are visible when they reach the shelf. The camera system addresses timing, not the staff member's visual check at the shelf.
The fill-ratio model was trained and validated on shelf imagery from grocery, pharmacy, and convenience layouts. It accounts for practical variation in product sizes, packaging, and shelf configurations. It does not reliably handle completely dark zones, heavily obstructed views, or shelves hidden by promotional barriers. Those cases are documented and flagged instead of producing a low-confidence inference.
Step 4: threshold check and alert decision
Each zone receives a fill threshold during onboarding, based on the category and the chain's operational preferences. With a threshold of 35 percent, an estimated fill ratio below 35 percent creates a candidate alert for that zone.
One below-threshold frame is not enough to send an alert. The system looks for consistent below-threshold readings across a configurable number of consecutive sample points before confirming a restock event. This temporal confirmation is the main safeguard against false positives. A brief low reading caused by a customer removing and replacing product, or by someone briefly blocking the view, will not consistently cross the threshold and will not produce an alert.
In pilot data, this method keeps false-positive rates below 5 percent for well-placed cameras on standard shelf zones. That result is not universal. Camera angle, lighting conditions, and product type all influence performance. Zones showing high false-positive rates during calibration are flagged for threshold adjustment or camera-angle review rather than allowed to create noise.
Step 5: alert routing and delivery
Once confirmed, an alert event creates a structured message with the store identifier, zone (aisle and bay), trigger timestamp, estimated fill ratio at the trigger point, and a reference to the camera feed. The event then moves to the alert delivery layer.
The chain can configure delivery through WhatsApp, SMS, email, or a combination. Most stores use WhatsApp because it is already available on the devices carried by floor staff, needs no app installation, and is read more often than email. The delivered message is specific: "Aisle 7, Bay 3: shelf fill below threshold (31%). Restock needed." It is not a dashboard notification that requires a login or a log entry to review later. It is a message for a person about a specific place at that moment.
Routing can vary by zone and shift. A dairy aisle alert may go to the dairy stockroom team, while a general grocery alert may reach a floor supervisor. Escalation rules cover alerts that remain unacknowledged beyond a configurable window.
Operational meaning in practice
This signal chain is not magic. It reflects engineering choices with real tradeoffs, especially between sensitivity, which catches stockouts, and specificity, which limits noise. A threshold set too low catches more true stockouts but creates more false alerts, and repeated false alerts make staff stop taking them seriously. A threshold set too high misses genuine events. We use conservative defaults, tune each zone during onboarding, and clearly flag configurations that are not producing reliable results.
Camera-based restock detection is not a replacement for backroom inventory management. It does not establish that a product is out of stock in the store. It indicates that a particular shelf zone is running low. These are separate problems requiring separate solutions. The system sits at the last-ten-meters problem, identifying the gap between "product is somewhere in the store" and "product is on the shelf where a customer can buy it." That is the gap scheduled shelf walks address imperfectly, and where camera-based alerts provide consistent improvement.