How edge AI cameras are transforming the future of smart retail

Application domains:

• Smart appliances (refrigerators)
• Smart retail (vending machines)
• Warehouse management

Traditional inventory monitoring in refrigerators and retail shelves relies on manual inspection, weight-sensor systems, or cloud-based image recognition. Each approach carries significant limitations: manual checks are labor-intensive and prone to delayed restocking, weight sensors require costly hardware modifications and are sensitive to product placement, and cloud-based solutions demand high bandwidth, introduce latency, and raise privacy concerns. Conventional MCUs also lack the computing power to run accurate detection models locally, leading to high false alarm rates.

CamThink addressed these challenges by deploying the NE301, built on ST's STM32N6, to perform AI-based product recognition entirely on-device. When inventory changes occur, only minimal structured data (e.g., "Coke: 3 cans") is transmitted instead of images. This fundamentally changes the economics and privacy profile of smart inventory systems.

Key benefits of this edge AI approach:

• Dramatic traffic reduction: Over 99% less data transmitted compared to image-upload solutions, reducing cloud costs by up to 90%
• Flexible power options: Supports battery, PoE, and Type-C power, enabling deployment in diverse environments
• Privacy by design: Visual data never leaves the device; only structured results are shared
• Lower operational costs: Eliminates manual inspections and reduces restocking delays
• Real-time awareness: Sub-50 ms inference delivers instant inventory visibility without cloud round-trips

By harnessing edge AI on the STM32N6, CamThink demonstrates how intelligent, scalable, and cost-effective inventory monitoring can be achieved in modern retail and appliance environments.

The NE301 runs a complete detection pipeline locally, from image capture to structured result output, with no cloud processing required.

The NE301 is periodically awakened on schedule and captures images of the refrigerator interior through its wide-angle CMOS sensor. The integrated ISP performs noise reduction and automatic exposure adjustment to ensure clear input even under complex lighting conditions such as reflections and low-light environments.

The built-in Neural-ART NPU runs a YOLOv8 Nano model to detect and count beverages directly on the device. Inference completes in under 50 ms, enabling real-time inventory awareness without uploading any images.

Only the recognition results, in JSON format, are transmitted to the backend system via MQTT (e.g., Home Assistant or a custom server). This keeps bandwidth usage minimal and ensures immediate actionability for restocking or alerts.

CamThink's solution fully leverages the heterogeneous architecture of the STM32N6:

• Neural-ART NPU — Accelerates YOLO model inference to under 50 ms, enabling real-time inventory detection without uploading images to any server.
• CamThink AI Toolstack — Provides a complete end-to-end toolchain from data collection and labeling to training and quantized deployment (.bin). Developers can build customized product recognition models within hours, even without deep AI expertise.
• Ultra-low power architecture — Combines the STM32N6 edge AI capability with an STM32U0 co-processor for power management and sleep control. The device wakes only when needed, dramatically reducing power consumption and deployment costs.

CMOS Image Sensor: OS04C10, HFOV 137°

The NE301 performs image preprocessing through its ISP (including noise reduction and automatic exposure), ensuring that the AI model receives clear and reliable input even under complex lighting conditions inside refrigerators, such as reflections and low-light environments.