Self-Checkout Shrink Detection: Camera Accuracy Without False Flags

How much does self-checkout actually lose?

The NRF's 2023 National Retail Security Survey put US shrink at $112.1 billion for 2022, about 1.6 percent of sales. Self-checkout is a known contributor: ECR Retail Loss research led by Adrian Beck found that loss rates climb as retailers push a larger share of transactions through unattended lanes. The failure modes are specific:

• Missed scans. The item passes the scanner without registering, deliberately or not.
• Mis-scans. A produce code entered for a steak, the classic banana trick.
• Bagging without scanning. Items go straight from cart to bag.
• Walkaways. The cart is full and payment never completes.

Kroger and Walmart both run Everseen's vision system to catch missed scans in real time, which tells you the majors have already decided cameras beat spot audits. Each unattended lane removes a cashier's passive deterrence, and nothing replaces it unless software does. The lane saves labor, then leaks margin.

What accuracy stops losses without false accusations?

Run the math on your own volume before picking a threshold. A supercenter pushing 2,000 self-checkout transactions a day with a 1 percent false positive rate flags 20 honest customers daily, over 7,000 a year, per store. At that rate staff stop trusting alerts and shoppers start posting videos. The workable design is two tiers:

• Soft prompts, 80 to 90 percent precision. The screen pauses and asks the shopper to rescan one item. A wrong call costs five seconds, so you can tune for recall here and catch most of the loss.
• Staff interventions, 95 percent precision or better. A person approaches a shopper. Fire this only on high-confidence, repeated, or high-value events, with the triggering video clip attached so the associate sees exactly what happened.

Measure precision at the intervention, not the model. A detector at 90 percent frame-level accuracy can still deliver 99 percent alert precision if you require corroboration across cameras plus a missing POS scan event. Accuracy is policy, not a model score.

How does product-in-hand tracking work at the lane?

The pipeline behind missed-scan detection has four stages. Overhead cameras cover the scan zone and bagging area. A hand-object interaction model detects when a shopper picks up an item, which is harder than shelf detection because a hand can occlude most of a small package. The item is classified, or at least visually fingerprinted, and its trajectory is tracked from cart to scanner to bag. Finally the system reconciles that trajectory against the POS event stream: an item that crossed the scan zone and entered the bag with no scan event inside the time window is a missed scan.

Two engineering constraints dominate:

• Latency. The prompt must fire before payment completes, so inference budgets are sub-second, which in practice means edge compute at the lane rather than cloud round trips.
• Occlusion. A single overhead view misses items palmed under produce; multi-camera fusion resolves what one angle cannot.

POS correlation is the quiet workhorse, because video alone rarely proves anything. The scanner log plus the trajectory does.

Why false accusations cost more than the shrink

In December 2023 the FTC banned Rite Aid from using facial recognition for five years after false matches led staff to confront innocent customers, some of them children. That order is the template for what regulators do when loss prevention AI misfires on people. The way out is architectural: track the product, not the person.

• Object trajectories carry no biometric identifiers, which keeps the system outside Illinois BIPA and outside the EU AI Act's remote biometric identification categories.
• Attach the triggering clip to every alert so an associate reviews evidence, never a hunch.
• Log every intervention and its outcome. Your false positive rate should be a number you can state in a deposition.

One wrongful accusation lawsuit or one viral confrontation video erases years of recovered shrink, and the reputational damage lands on the retailer, not the vendor. Design the system so the worst failure mode is a wasted rescan prompt. Bad alerts cost more than stolen steaks.

Which patents cover product-in-hand tracking, and should you build or license?

Vision checkout is crowded IP territory. Amazon alone has filed hundreds of applications around Just Walk Out, and granted claims in this space read on the exact pipeline above: detecting an item in a shopper's hand, tracking it through a zone, and reconciling it against a transaction record. Position Imaging holds granted US patents in this area, including US 11,774,249, US 12,079,006, US 12,066,561, and US 12,000,947, part of a computer vision and machine learning portfolio that Apple and Bosch have cited in their own filings.

For a team building shrink detection, that changes the build-or-license math in two ways. First, freedom to operate: a claim chart against the granted art costs a few weeks now, versus discovery costs later. Second, speed: licensing proven tracking IP lets a team stand up a lane pilot in months instead of spending a year reinventing hand-object tracking and POS correlation that the art already covers. Run FTO before your Series A diligence does.