Design of a Tiny Machine Learning system for UWB radar based multi-target detection

Resumen

Tiny Machine Learning (TinyML) is a novel field of research which consists on designing machine and deep learning models at a reduced size, enabling them to be executed on tiny devices such as Internet-of-Things units, edge devices or embedded systems. The increasing use an research on TinyML follows the scientific trend of moving data processing closer to where data is generated. This offers an increase in the autonomy of tiny devices and achieving the most efficient use of energy possible.

One of the most promising applications for TinyML is presence detection. However, current automated presence detection systems rely on the processing of images taken from video cameras or microphones. While this is a very appropriate task for Machine Learning-based applications, the acquisition and processing of this data can lead to a risk in privacy vulneration, since images, videos or audio of people can be considered sensitive information. The use of Ultra-Wideband (UWB) radar might be a solution for this privacy issue, and aims to be a promising technology for tiny devices, since it is characterized by high precise recordings, low energy consumption and fast acquisition of data.

Combining these two technologies, we propose the development of automated presence detection solutions with the use of Ultra-Wideband (UWB) radar and TinyML. The goal of this research is to develop functional deep learning solutions for person counting and object detection with the use of UWB radar and applying TinyML algorithms to be able to deploy them on tiny devices. Different approaches have been considered, and several optimizations are offered in order to further shrink down the networks or improve their results.