Raspberry Pi & Phi-3 Mini enable GDPR-compliant visual monitoring with on-device AI

A team from the University of Applied Sciences Upper Austria has published a proof-of-concept pipeline that keeps visual monitoring fully on-device, eliminating GDPR privacy risks. By combining a Raspberry Pi 5 with a Hailo-8L neural-network accelerator running YOLOv5n-seg for real-time object detection, raw pixel buffers are processed and immediately discarded. A stateful trigger engine then forwards minimal JSON event payloads to a locally hosted Phi-3 Mini (3.8B parameters, Q4_0 quantisation), which synthesises one-to-two sentence natural-language alerts for human operators. The system never transmits any image data across a network—only the generated text alert is sent.

Performance measurements show the setup delivers practical inference latency and resource utilisation on single-board hardware. The paper includes representative generated alerts and confirms that combining a dedicated accelerator with an on-device LLM is not only feasible but produces deployable, human-readable output. This architecture inherently complies with the GDPR's data-minimisation principle (Art. 5(1)(c)) by design, offering a template for privacy-preserving visual monitoring in sectors like retail, healthcare, and smart buildings.