Projects    »    

Optimized CNN implemented on TPU camera on autonomous robot and open-source analogue neural network accelerator with a parallel pipeline

Optimized CNN implemented on TPU camera on autonomous robot and open-source analogue neural network accelerator with a parallel pipeline

Our research started as a vision system for a RoboCup competition robot. There we faced many real-life problems such as creating a robust and high-speed system which would also be power efficient. We developed our own architecture optimized for a TPU accelerator and constructed a compact camera containing coral edge TPU together with CM4 capable of running the ubuntu operating system allowing fast image processing with great connectivity, which already brings results in multiple projects. During development, we found many problems such as proprietary documentation, communication bottlenecks and so on. Therefore we created our own fully parallel analogue camera from discreet components with analogue neural network implementation for computer vision which is low cost and fully opensorced.

Jakub Gál

19 years

Stand24
ProjectComputing-05
CountrySlovakia

Share on Social Media