- Home
- Edge AI in retail
Edge AI in retail: why images should stay in the store
Store cameras can be analysed in a cloud or on a device in the store. The choice decides what leaves the building, how fast a screen can react, and what happens when the internet goes down.
Definition
Edge AI in retail means the vision model runs on a device inside the store: images are analysed there, in memory, and only the results leave. A cloud system streams or uploads images to a data centre and analyses them there. Both can count the same things; they differ in what crosses the internet, how quickly an answer comes back, and whether the store keeps working when the line drops.
In the store or in the cloud
| In the store (edge) | In the cloud | |
|---|---|---|
| What leaves the store | Events and counts: which product, when, for how long | Images or video |
| Reaction time | Tenths of a second, on the shop floor | Adds the round trip to a data centre |
| Internet outage | Recognition and screens keep working; events are sent later | Analysis stops until the line is back |
| Bandwidth | Small: text events | Large: continuous video per camera |
| Privacy | Images never leave; easier data minimisation | Images stored or processed off site |
| Hardware | A small computer per store | Cameras only |
| Updates | Need a fleet that is managed remotely | Changed in one place |
Three reasons to analyse in the store
- Privacy. Data-protection law (KVKK in Türkiye, the GDPR in Europe) asks you to collect no more than you need. If images are analysed in memory and only product events leave, there is no footage to store, transfer or leak.
- Speed. A screen that reacts to a product in a shopper's hand has to react while the hand is still there. In GoTrack's test setup the screen showed the product a median 0.35 s after recognition, 0.51 s with two products sharing the screen (method and limits).
- Resilience. Store internet is not always reliable. With the store device's internet unplugged for 7 min 43 s, all 3 showings still reached the screen and all 3 records reached the cloud afterwards, with their real times (lab note).
The trade-offs, honestly
- A device per store has to be installed, powered and occasionally replaced.
- Updates have to reach every store without a visit, and a failed update must not take a store offline.
- Health has to be visible from one place: which stores are recognising products, which camera is down.
- Compute is limited, so the models have to be chosen and tuned for the device, not for a data centre.
A serious edge system answers all four with remote management. GoTrack's store devices report their health and receive updates remotely; trust and security covers how data is kept apart and updates are signed.
Questions to ask any vendor
- Where exactly are images analysed, and are they ever uploaded?
- What leaves the store, field by field?
- What keeps working when the internet goes down, and what is lost?
- How are devices updated, and what happens if an update fails?
- How do we see, from one place, that every store is working?
The privacy checklist has ten more questions for the data-protection side.
Questions
Are images processed in the store or in the cloud?
In the store. Images are analysed on a device in the store and are not sent to the cloud for analysis. Only product events and device health reach the cloud, which is also why recognition and screens keep working without internet.
What happens if the store's internet goes down?
Recognition, screens and event recording carry on in the store. Screens play their stored content, and events wait on the device and are sent with their real times when the connection returns. In our test on 4 September 2026, every showing during a 7-minute outage reached the screen and its record reached the cloud afterwards.
Does GoTrack record camera footage?
Not by default: each analysed frame is discarded. If debug recording is switched on, the files stay on the store device and are deleted after 30 days by default. The only image that reaches the product catalogue is one an authorised user deliberately adds as a product photo.
How are store devices updated?
Remotely. New versions reach store devices without a site visit, are applied at a quiet moment, and roll back on their own if the new version does not work.
How quickly does the screen change when a product is picked up?
In our test measurement on 30 September 2026, the screen acknowledged painting the product within 0.71 s of GoTrack deciding which product was in hand, in all 24 acknowledged fires: a median of 0.35 s with one product and 0.51 s with two. One of the 25 fires sent no acknowledgement. The time from the lift itself to that decision is not included and not yet measured separately. Method: the reaction time note.
See analytics that never send a video.
We will show you what leaves the store, and what never does.
or write to info@gotonom.com