Computer Vision for personalised drive-thru orders

An AI solution for a quick-service restaurant chain: a camera at the drive-thru reads the licence plate, identifies the guest or the order history of the car and shows a personalised offer on the screen.

The goal was to build a system that personalises service at the drive-thru without manual sign-in or any extra action from the customer.

At drive-thru points guests usually order quickly, without opening the app and without identifying themselves in the loyalty programme. As a result the chain cannot use purchase history, favourite items, bonus mechanics or personal offers at the moment of ordering. Some restaurants already have digital menu boards, while others still use printed menus where personalisation can only be shown on a separate screen next to the order point.

We had to design an AI system that recognises the licence plate in real time, links it to a user account or to the order history of that car, and sends a suitable offer to the screen: a favourite dish, a combo, an upsell, a bonus mechanic or an invitation to sign in to the app.

Project format: R&D prototype of a CV solution
Industry: quick-service restaurant chain
Scenario: drive-thru
Data type: camera video stream, licence plate, order history
Key AI function: real-time licence plate recognition
Service scenarios: 3
Point types: digital menu boards and legacy menu boards with a separate screen
Target recognition time: 1–2 seconds after the car arrives
Target recognition accuracy on a valid frame: 90%+
Project horizon: up to 24 weeks from research to pilot operation

The project had to combine computer vision, real-time event processing, customer identification and a recommendation engine. The camera detects the arriving car, recognises the plate and passes it to a backend service that resolves one of three scenarios:
the car is linked to a signed-in user of the mobile app; the car has ordered at the drive-thru before but is not linked to an account; the car is here for the first time and needs a new order-history profile.
For signed-in guests the system can show the name, bonus balance, favourite dishes and personal offers. For known cars without an account it uses past orders. For new cars it falls back to universal mechanics: bonuses for signing in, welcome offers and non-personalised upsells.

We designed a prototype in which the drive-thru camera captures the car in the ordering zone and passes the frame to the computer vision module. The model locates the plate, recognises the characters, normalises the number and sends it to the backend.

The backend checks whether the plate is linked to a user account in the mobile app. If it is, the system uses loyalty data and order history. If there is no account but the car has been here before, it uses the accumulated order history for that plate. If the car is new, an anonymised profile is created and used for recommendations later.

After identification the system sends a communication scenario to the screen: favourite dish, personal offer, bonuses, combo or upsell. In restaurants with digital menu boards the offer can be shown on the board itself; in restaurants with printed menus a separate screen next to the order point is used.

Integration with push and email communication was also designed for signed-in users: personal coupons, bonus notifications, challenges and win-back offers.

The result is an R&D prototype of an AI system for drive-thru personalisation based on computer vision and order history.

It covers 2 restaurant types: points with digital menu boards and points with legacy boards where personalisation is shown on a separate screen. Three identification scenarios are described: a signed-in user, a known car without an account and a new car. Four types of personalisation are supported: favourite dishes, combo offers, bonus mechanics and upsells.

The project plan covered research, requirements and test programme, training data collection, model training, inference, the DevOps loop, integration and load testing, manual testing, test and pilot operation — up to 24 weeks for the full cycle.

Target metrics of the prototype: plate recognition in 1–2 seconds, 90%+ accuracy on a valid frame, real-time event processing and delivery of the personal offer to the screen before the order is placed.