Application of Tiny-ML methods for face recognition in social robotics using OhBot robots

Abstract

The aim of this paper is to show the possible application of Tiny-ML family neural networks to social robots for face recognition. Social robotics is a constantly developing field that allows the production and development of robots whose task is to accompany humans, participate in social situations and perform specific educational, entertainment and therapeutic tasks. One of the fundamental problems of social robotics is the proper recognition of humans by robots. This poses a critical problem because it is the moment when human-robot contact is initiated. Widespread solutions, in addition to high efficiency, also require adequate computing power, which in social robots cannot always be provided. For this purpose, solutions from the Tiny-ML stream are used, i.e. such a construction of neural networks and machine learning that would be adapted to limited technological resources and, at the same time, equally effective. The paper uses a YOLOv4-tiny network, which was compared to a YOLOv5s solution, both in terms of efficiency and processing time. The proposed networks were tested on social robots of the OhBot type and with extended capabilities, by using Neural Sticks. The results obtained show the highest efficiency of the implemented YOLOv5s network using a Raspberry Pi along with an accelerator. The presented research is an opportunity to draw attention to the problem of computational complexity in robotic applications, and also has the potential to popularize social robots and their use in everyday life.