Kingham Xu's FryPi Aims to Deliver a Compact Open Source STM32F4 Platform for TinyML Projects
Demos including handwriting, gesture, and facial recognition showcase the STMicro chip's capabilities for low-power ML work.
Maker Kingham Xu, creator of the open source OV-Watch, has unveiled a follow-up design: an STMicroelectronics STM32F4-based "mini devkit" that targets projects ranging from real-time control to tiny machine learning (tinyML): the FryPi.
"The original purpose of making this board is because of the previous smartwatch project OV-Watch," Xu explains of the new board's origins. "Many […] people said that the device is very difficult to weld, secondary development is not convenient. In addition, I also plan to deploy some AI [Artificial Intelligence]-related things on STM32 and do a tutorial, so the FryPi 'fried chicken pie' was born. This development board is not only suitable for beginners, but also for advanced developers."
The compact development kit is built around the STMicro STM32F411REU6, chosen for its compatibility with the -CEU6 used in Xu's earlier OV-Watch project. There's a single 100MHz Arm Cortex-M4F microcontroller core with digital signal processor (DSP) and an ART accelerator, 128kB of static RAM (SRAM), and 512kB of on-board flash with support for external expansion.
While the board can be used as a simple development kit, bringing its general-purpose input/output (GPIO) connectivity to two 0.1" double-row pin headers at either side, there's support for optional expansion modules including a touchscreen display, which Xu uses to showcase compatibility with the OV-Watch firmware, and both thermal and visible-light camera modules.
To showcase the microcontroller's capabilities for on-device machine learning, Xu has written a set of tutorials that include handwritten digit recognition, gesture recognition from a thermal camera input, and facial recognition from an Omnivision OV2640 sensor.
More information on the project is available on Xu's Hackaday.io page, while source code and hardware design files are available on GitHub under the reciprocal GNU General Public License 3.