POS Camera Module

The Project

For a locally based industrial company, we developed a specialized camera system that integrates into modern point-of-sale systems. The module uses AI inference to detect whether a shopping cart is empty or still loaded — directly at the checkout. With a latency of under 4 milliseconds, the system operates entirely in real time.

Challenge

At self-checkout stations, it must be reliably detected whether a shopping cart is actually empty after the scanning process. The solution had to run on constrained POS hardware without a GPU, support two different industrial camera types, and trigger automatically via a magnetic field sensor as soon as a shopping cart is in position. Additionally, Qt could not be used due to licensing restrictions.

Core Features

• Automatic Camera Detection: The system automatically detects via USB whether a TIS (The Imaging Source) or IDS camera is connected and loads the appropriate drivers.
• Magnetic Field Sensor Integration: An industrial magnetic field sensor automatically detects shopping carts via field changes and triggers image capture.
• AI Classification: A lightweight neural network classifies captured images in real time with interchangeable inference backends.
• Production GUI: State machine with four states (INIT → MINI → WORKING → ALARM) — live preview in thumbnail format, full-screen notification for non-empty carts, all operable via touch.
• ML Training Pipeline: Complete workflow from data collection through 5-fold augmentation to Docker export — without dependencies on the target system.
• Debug & Diagnostics: CLI tools for camera status, sensor training, live viewer, and inference tests — all accessible through a developer menu.

Technical Highlights

• Under 4ms inference latency on embedded POS hardware without a dedicated GPU
• Modular architecture with cleanly separated layers — interchangeable camera drivers, sensor connectivity, and inference backends
• Direct hardware communication with the sensor via the camera data channel — no additional wiring required
• Flexible configuration with extensive parameters and validation for various deployment scenarios
• Service integration with automatic restart and process management in production operation

Result

The checkout camera module runs in production on POS terminals in retail. The real-time detection operates reliably during checkout operations and can be easily extended with new camera types and ML models thanks to its modular architecture. The comprehensive developer documentation enables smooth maintenance and further development.