5 things to consider before deploying visual AI at the edge in retail

5 things to consider before deploying visual AI at the edge in retail

Thinking about leveraging visual AI at the edge in your business? Here are five things to consider before deploying your solution.

1. Data privacy

Whether it’s shelves, social distancing, or entry and exits, your visual AI system will be constantly “watching” parts of your stores and accumulating a large amount of visual data. Once your system is installed you should assume it’s only a matter of time before the general public gets wind of you using some form of AI or AI enabled video monitoring, and your customers are going to have questions around what kind of data you’re collecting and how you’re using it. Your company is going to need to have a specific policy and/or point of view on how to handle privacy issues, both to govern your solution and to answer any questions your customers have. Having a solid privacy policy will not only assuage any potential concerns from questions, but it will help your company avoid potential negative PR.

General governance of your solution is another reason for having a security policy as it will help inform business process design related to your solution. For example: a stockout solution could be configured to “throw away” all images of a stockout that have customers in them, and just keep the underlying information, i.e. “Cheerios went out of stock at 14:02 on August 12”. Solutions that detect license plates can be configured to automatically delete the image of the license plate once the customer has received their order.

Whatever your solution is, it’s important to make sure your company’s privacy stance is well known so that customers don’t feel as if they must sacrifice their privacy to shop at your grocery store.

2. Cameras

One of the most important features we’ve discovered in our experiences around building visual AI solutions for retailers is having at least a 12x optical zoom, which provides two main benefits:

Higher image quality

You want the highest image quality possible as it will help ensure that your visual AI solution is as accurate as possible, as lower quality images are harder for the AI models to interpret. In the past we’ve tried to go with lower zoom cameras or ones that used a “digital zoom” and found that the pictures were of much lower quality, which led to significantly higher error rates.  To quote our Director of Data Science: “No algorithm is going to overcome bad images”.

Observe a larger area in store

In both our lab testing and experiences in actual stores, we’ve found that the higher zoom cameras can observe a much larger area in the store. This means that while you will pay more for the cameras on an individual basis, you will need fewer of them overall.

Neal Analytics has spent hundreds of hours testing cameras in terms of camera angles, shelf heights and product size and are well equipped to help you select and evaluate cameras for your visual AI solution.

3. Infrastructure

There are two main infrastructure pieces that need to be considered for an edge-based AI solution:

  1. Power: for certain high demand or industrial applications you need to use an especially powerful edge device that requires one or more dedicated 15 AMP power line, make sure you talk with a qualified electrician to ensure your current electrical infrastructure can support the device. In a recent customer engagement, our client’s infrastructure team elected to install a dedicated power line, as their facilities people felt that was the best option given the edge device’s power demands. While not all edge devices are “power hungry” per se, it is important that you keep this mind when installing the device.
  2. Internet Connectivity: having the edge device operate on a separate internet connection than the one the store uses for point of sale and other purposes is a good idea, so that the edge solution doesn’t take up needed bandwidth for processing transactions and other purposes. However, this solution does need a public facing IP not just so it can connect to the cloud but so that the team supporting the solution can push code to the edge device, configure cameras and make other adjustments as necessary

4. Data collection

Visual AI models need to be “trained” with data relevant to the scenario it’s being used for:

  • A stockout solution will need hundreds of photos of the aisles the solution is monitoring so the solution can “learn” to detect the various combinations of possible stockouts. While a human mind can extrapolate the concept of a stockout to every shelf in the store after being shown 1-2 of them, a computer will need to see hundreds of variations of stockouts on a given shelf in order to “learn” when a stockout has occurred and will need to see these variations for every aisle/shelf being monitored.
  • Solutions that track people’s movements in the store in order to monitor social distancing will need videos of people shopping within your stores to “learn” good levels of distancing from bad.

When creating the implementation plan for your solution, you will need to factor in time to gather data, data collection that will often involve simulating the event(s) you want your solution to monitor. This could include having store associates take items on and off shelves, capturing videos of people moving around the store, etc.

You will also need “labelers” who take data or images collected by the cameras and label things as “stockout” or “persons too close” in way that the AI models can interpret properly. However, the need for aggressive data collection will decline as you onboard more and more stores, potentially to the point where you can use data generated from day to day business for updating the models and/or just needing to expend a small amount of effort to gather new data a couple of times a year.

5. Business transformation

Visual AI solutions will completely change how you do business, whether it’s changing the way you approach safety, facilities management, transactions, supply chain management or even managing the number of people in your physical retail space. While the technical people are installing the technology and fine-tuning things, another team needs to be working in parallel to think through the required changes to business process, space management, ordering and the like.

For example: in a retail stockout scenario someone should be tasked with reviewing patterns in stockouts in order to refine product ordering. Every process that these solutions touch need to be reviewed so that the proper adjustments can be made so that your company can

  • Adopt the solution successfully
  • Get the most out of its investment.

While there are numerous other factors and considerations that go into building a visual AI solution for you retail business, keeping the aforementioned things in mind will not only help your edge adoption go smoother, but will also ensure that your company receives more value. Neal Analytics’ stockout solutions are backed by hundreds of hours of R&D and in real life experience, please contact us if this is something you’re interested in pursuing, as we have the right combination of managerial consulting experience and technical know how to make your retail AI solution successful.

Learn more about visual AI at the edge in retail: