
Computer Vision Exit Verification for a Regional Warehouse Club
Let's Connect
Overview
What we built
A regional warehouse club's exit doors were the slowest, most frustrating part of a member's visit. We replaced manual receipt checks with cameras that verify carts automatically, cutting the wait almost to nothing.
In plain terms: every member leaving one of the club's 14 warehouses had their loaded flatbed cart checked by hand against a receipt. On a quiet afternoon that was manageable, but at weekend peaks the queue backed up past three minutes per cart, and staff under that kind of pressure could only give each cart a quick look. Shrinkage at the exit kept climbing, and complaints about the door experience were among the most common thing members had to say about the whole chain.
We built a computer vision system that watches carts as they pass through the exit archway, classifies what is loaded onto them, and checks that against the member's digital receipt in real time. A verified cart gets a green light and walks straight through, while anything that does not match is sent to a staffed lane for a quick, targeted check instead of a blanket one. 92% of carts now clear automatically with no staff involved, peak weekend waits fell from over 3 minutes to under 40 seconds, exit shrinkage is down 18% across all 14 warehouses, and door-related complaints dropped by 55% in the first quarter.
The Problem
Manual receipt checks at exits
Every one of the club's 14 warehouses ran the same exit process: a staff member manually matched each loaded flatbed cart against the member's receipt before letting them leave. It was a task built for careful comparison, done under constant time pressure, at a door that never stopped moving during opening hours.
Weekend traffic exposed the model's limits fastest. Queues at the exit stretched past three minutes as the same manual check ran again and again, and with a line building behind every cart, the check itself became necessarily cursory. Staff were being asked to catch mismatches at a glance while a queue grew behind them.
The two symptoms fed each other. Rushed checks let more discrepancies through, and exit shrinkage kept climbing as a result. At the same time, the slow, imperfect door experience was becoming one of the most common feedback themes members raised about the chain, regardless of which of the 14 warehouses they shopped at.
Manual receipt matching
Every exit relied on a staff member manually comparing a loaded flatbed cart against the member's receipt, a slow check repeated at every one of the 14 warehouses.
Weekend queue backups
Queues at the exit stretched past three minutes during peak weekend hours, with the same manual matching process running under constant time pressure.
Necessarily cursory checks
With queues building behind every cart, staff could only give each receipt and cart a quick glance rather than a careful comparison.
Rising exit shrinkage
Exit shrinkage kept climbing as rushed manual checks let more mismatches through, while door complaints became one of the most common member feedback themes.
What it was costing them
Every rushed check was a chance for a mismatch to walk out unnoticed, and exit shrinkage kept climbing as a result. Weekend queues stretching past three minutes cost the club goodwill at the last moment of every visit, and door complaints becoming one of the most common feedback themes meant the exit experience was shaping how members felt about the whole chain, not just the checkout.
The Solution
Camera-verified exit lanes
We replaced the manual check with a computer vision system built around the exit itself. Ceiling-mounted cameras at each exit archway watch every cart as it passes, detecting and classifying the items loaded onto it without asking members to stop or present anything beyond their normal receipt.
A matching service cross-references what the cameras see against the member's digital receipt in real time, before the cart reaches the door. Verified carts get a green light and walk straight through, so most members never break stride on their way out.
Carts that do not match are not turned away outright. They are flagged to a staffed lane where a person runs a quick, targeted check instead of the blanket manual comparison every cart used to get, so staff attention goes where a mismatch is actually likely.
Key decisions
Verify automatically, check exceptions
We built the system so verified carts pass automatically and only mismatches route to a staffed lane, concentrating staff attention on the carts that actually need a look.
Cameras at the archway
Ceiling-mounted cameras sit at each exit archway itself, so detection happens exactly where carts already pass, with nothing new for members to stop for.
Match against the digital receipt
The matching service checks detected items against the member's existing digital receipt in real time, reusing data the club already captured at checkout.
Green light, not a gate
Verified carts get a simple green light rather than a hard stop, keeping the exit moving for the vast majority of members.
Roll out across all warehouses
The system was built to run the same way at every one of the 14 warehouses, so the exit experience became consistent across the whole chain.
Measurable Impact
What changed after launch
The exit stopped being the slowest part of the visit. 92% of carts now verify automatically with no staff intervention, and the average wait at peak weekend hours fell from over 3 minutes to under 40 seconds. Members leave the same way they always have, just without the delay a manual check used to add.
The numbers moved on the things that mattered most. Exit-related shrinkage is down 18% across all 14 warehouses within 6 months of rollout, and member complaints about door checks fell by 55% in the first quarter, showing up in feedback as quickly as it showed up in the wait times.
Exit verification
Staff manually matched every receipt to cart
92% of carts verified automatically
Peak wait time
Queues stretched past three minutes
Under 40 seconds at peak weekend hours
Exit shrinkage
Climbing steadily across all 14 warehouses
Down 18% within 6 months
Door complaints
Among the most common member feedback themes
Fell 55% in the first quarter
Headline results
92% of carts verified automatically with no staff intervention required
Average exit wait at peak weekend hours cut from over 3 minutes to under 40 seconds
Exit-related shrinkage down 18% across all 14 warehouses within 6 months
Member complaints about door checks fell by 55% in the first quarter after rollout
Tech & Tools Used
What powered the build
Every tool below earned its place in this engagement. Here is the part each one played.
Python
Glues the exit verification pipeline together, connecting camera capture, the detection model and the matching service into one running system at each archway.
PyTorch
Trains and runs the model that detects and classifies items on each passing cart, the core of what decides whether a cart matches its receipt.
YOLOv8
The detection architecture behind the vision model, spotting and classifying individual items on a cart from the ceiling-mounted camera feed as it passes.
OpenCV
Handles the image capture and preprocessing from the ceiling-mounted cameras before frames reach the detection model at each exit archway.
NVIDIA Jetson Orin
Runs inference on site at each exit archway, so cart detection and classification happen locally rather than over a network round trip.
TensorRT
Optimises the detection model to run fast enough on the edge hardware that a cart is verified before it reaches the door.
Apache Kafka
Streams detection events from each archway to the matching service in real time, so a cart's verification result is ready the moment it arrives at the door.
PostgreSQL
Stores receipt data, detection results and verification outcomes, giving staff a record to check when a cart is flagged to the staffed lane.
React
Powers the staffed-lane interface where staff see why a cart was flagged and run their targeted check against the receipt.
AWS S3
Holds captured cart imagery and verification records, giving the club a durable archive to review flagged cases or audit shrinkage trends.
Ready to Build your Retail & Wholesale Business with Image Recognition & Classification
Ask Byte
Ask Byte
Typically replies instantly
just Now
Hi! I'm OrganByte's assistant. How can I help you today?
AI-generated content may be incorrect

