Architecting the Checkout Lane

Technical executives in retail face a persistent challenge: balancing loss prevention while reducing friction in the customer experience. Recent implementations of computer vision systems at Sam’s Club demonstrate the potential for technology to meet both goals. At Groceryshop 2024, the CEO of Sam’s Club described their new exit verification system and reported a 14% improvement in the Net Promoter Score—a metric that validates the investment in advanced technological solutions.

Computer Vision

My local Sam’s Club has two arches that you push the buggy through after checking out (on the app or at the registers). About 25 feet past the arches is an employee who will wave you through or check your receipt. Most of the time these days, I’m waved through. Combine this with the scan-to-go app, and I can be in and out of Sam’s Club in about 15 minutes. Compare this to the experience at Costco, which is more like a social event, because I will have enough time to make friends in line!

The arches use computer vision technology, a branch of artificial intelligence that enables computers to interpret and analyze visual information from the world through cameras and sensors, processing images and video to perform tasks like object recognition and motion tracking. The implementation builds upon existing computer vision capabilities developed at Sam’s Club, a natural outgrowth of their technology already deployed for inventory management by autonomous floor robots.

Computer vision technology is used in other retail applications. Floor robots can scan shelves to ensure the price shown is the same as the pricing database used for checkout; this is especially useful if shelf prices are updated for temporary sales. Heat mapping can determine traffic patterns within a store, to optimize store layout and product placement. Loss prevention is of course a key application. Sam’s Club attempts to manage the entire inventory lifecycle with the technology – as Chris Nicholas shares, “having computer vision tell us what’s coming into the club, having computer vision validate everything, and having computer vision tell us what’s going out of the club” drives efficiency. Theirs is not the only checkout systems using computer vision technology; the Circle K near us has a touchless checkout that uses camera scanning instead of barcodes.

At Sam’s Club, an array of cameras in the arches capture items in the customer’s cart (and hands), scanning multiple angles of the top and sides of the cart. The shapes and colors of the items are compared to the inventory system to identify the item. Then the cart contents are correlated to receipts. When I first experienced not having to show my receipt at the exit, I thought it was because cameras had been recording me while using the scan-to-go app, but it’s not that deep… apparently face and other biometric data are not used. Recent receipts are compared to recent cart scans.

If discrepancies are detected, the exit employee reverts to the previous process of scanning the receipt and sample items. This is a great fallback, especially if it’s rare, since it can also cover lags in image processing or data synchronization.

Data Integration Architecture

The core challenge lies in processing visual data and integrating with multiple systems in real-time, all within that short processing window. While the employee stationed at the exit can reconcile problems, the primary objective is to maintain rapid customer flow through exit points while ensuring transaction accuracy.

This workflow implies that there is a local inventory and receipts database and that checkout data is at least temporarily stored to do the comparison. The training model used for the item detection would benefit from data aggregated across stores, so the data would be synced back to a main database.

Edge computing

To allow for local data processing and synchronization to a centralized location required by a system like the one at Sam’s Club, several solutions are available.

An important requirement to consider is that the data transfer is asymmetrical: the store is sending a lot of data to headquarters, but it only needs to receive a smaller training model back. A good approach, then, is to maintain each store location as an edge location.

Each store location would function as an edge node with local processing capabilities and data storage. This local infrastructure would handle critical functions like tracking real-time inventory records, processing point-of-sale transactions, and running the computer vision analysis. The edge nodes would maintain full operational capability even during network outages, ensuring business continuity.

Data from these functions would be streamed in near real-time to the central location. The central system would aggregate and process data from all store locations to continously train the model used in the exit process. These model improvements would then be pushed back out to the edge locations.

Potential bottlenecks

In streaming architectures, it’s critical to monitor and manage the potential bottlenecks. Network throughput must be carefully monitored, especially during peak business hours when data volume is highest. Using efficient compression algorithms help to minimize network bandwidth usage. Queue sizes at both the edge and central locations need to be managed to prevent memory overflow and ensure timely processing.

The target database must be designed to handle the concurrent updates from multiple locations without becoming a bottleneck itself. A distributed database is a good fit here, providing low latency for in-region writes, horizontal scalability, and resilience to region outages.

This architecture enables a robust, scalable system that can handle the complex requirements of modern retail operations while maintaining consistency and reliability across all locations.

Conclusion

The successful implementation of exit verification systems demonstrates how sophisticated technical architecture can be used to improve the customer experience. By carefully considering edge computing requirements and data synchronization strategies, technical leaders can deliver solutions that scale across retail operations while maintaining system reliability and performance.