IndustriesUltra-Low Power Face Recognition Using STM32
An OEM required a smart face recognition system, a battery-powered embedded device that could authenticate its employees completely offline, without relying on cloud connectivity..
Most existing solutions relied on remote servers and cloud connectivity for face recognition, failing to meet the client's criteria.
The core challenge was implementing a seamless face recognition system within a compact, battery-powered embedded device while maintaining a long battery life.
We built a face recognition device using the STM32H743 microcontroller: a small, powerful chip running at 480 MHz. A face recognition model, MobileNetV2, was optimized through quantization to run directly on the chip, handling everything from capturing a face to identifying the user, all without any internet or cloud connection.
When a face is recognized, the device instantly triggers a relay to unlock or deny door access. To save battery, the device stays in deep sleep and only wakes up when motion is detected. With this, the device can run for several months on a single charge, making it a reliable solution for locations where a power supply is not readily available.
The device works completely offline, making it suitable for any location including remote sites, basements, and off-grid installations where internet access is unavailable. Using a lightweight MobileNetV2 model, the device can perform accurate face recognition directly on the chip, with no GPU or server required. The wake-on-motion feature keeps the device in sleep mode until needed, allowing it to run for several months on a single charge without any wired power. Since everything runs on the device itself, no external server, extra hardware, or complex setup is needed.
Network Independent
Face recognition runs entirely on-device - fully offline operation with no server, no internet dependency, and no latency.
Extended Life
The wake-on-motion feature keeps the device in sleep mode until needed, allowing it to run for months on a single charge.
Edge Processing
MobileNet model quantized to run on a microcontroller - enrollment, matching, and access control all handled locally on the STM32H743.
Let's build the future, together.
Fullstack engineering partner: from hardware to cloud to Vision AI.
Trusted by
