Challenge

Our customer, a major retail chain, wanted to determine how much shrinkage was attributed to Point-of-Sale (POS) losses.

To do this, they would need to detect actions like missed scans and “sweethearting” (scanning a lower-priced item instead of a higher-priced item) while sufficiently protecting employee and customer privacy. However, the existing store cameras were incapable of the IoT and AI features needed for a cognitive vision solution.

Solution

The Neal Analytics team worked with the customer to select cameras and strategize camera placement in a way that would best support the use case. We installed eight cameras across 4 locations (2 per store) to run a POC leveraging Nvidia Deepstream SDK.

Our data science experts then gathered and labeled the data to train a vision AI model. The solution would cross reference time stamps to compare against transactional data and help identify missed scans.

To protect Personally Identifiable Information (PII) data, Neal Analytics also included the ability to redact the faces of the cashier and customers if they appear in the camera’s zone of interest.

Example of face redaction for PII purposes

Results

Neal Analytics provided the customer with a trained vision AI model that could be used to

• Automatically compare POS scans with what the AI algorithm sees
• Leverage the power of edge computing with accelerated Edge AI to detect and track the products throughout a transaction
• Insights are integrated with other enterprise systems for alerting and investigation
• Flexible and extendible edge to cloud architecture for further expansion to additional stores, registers, and in-store video analytics scenarios

With these new capabilities, the customer has been able to better prevent shrinkage within their stores and rapidly scale out the solution across the United States.