Within the StorAIge project, UTIA aims at developing AI-self-trained, ultrasound-based, hand-gesture-recognition sensors implemented and integrated on ST microcontroller circuits with AI accelerator.
The targeted application domains are Graphical User Interface (GUI) sensors. Output from video-based AI algorithms will be used for detection of coordinates of objects like the moving hand of operator performing a gesture by hand. Measured trajectory of hand will be used as an input for training of the final AI algorithm inside an ultrasound-based hand gesture recognition sensor.
Four application notes and evaluation packages, published by UTIA, describe implementation of hardware (HW) accelerated AI algorithms on programmable logic part of Zynq Ultrascale+ devices.
The first application note explains es how to define « reduced size » HW Data Processing Unit (DPU) fitting into small module TE0821-01-2cg-4GB manufactured by company Trenz Electronics (Used Zynq SoC has only limited PL, BRAM and DSP resources in this case). It works with pre-compiled models to accelerate resnet50 in HW for Vitis AI 2.0 facedetect demo with input from USB camera. Note describes how to recompile the AI models for this small version of the DPU.
- Xilinx Vitis AI ‘facedetect’ and ‘resnet50’ Demo on Trenz Electronic TE0821-01-2cg-4GB SoM + TE0706-3 Carrier
The three other application notes describe set of Vitis AI 2.0 applications (recompiled from Xilinx www pages dedicated to Vitis AI 2.0):
- Testing all Samples from Xilinx Vitis AI Library 2.0 on Trenz Electronic board TE0808 SoM + TEBF0808 Carrier
- All VART Examples from Xilinx Vitis AI 2.0 for Trenz Electronic board TE0808 SoM + TEBF0808 Carrier
- Xilinx Vitis AI ‘facedetect’ Demo on Trenz Electronic board TE0808 SoM + TEBF0808 Carrier
More information: https://zs.utia.cas.cz/index.php?ids=projects/storaige