Using Autoencoders and Support Vector Machines (SVMs) for Face Classification in Telepresence Robots

Artash Nath

COVID19 is a global pandemic. Rising number of cases, limited health care facilities and shortage of personal protection equipment (PPE) mean health care workers are under stress. Robots can play a big role in monitoring patients in nursing homes and long-term care facilities for symptoms of COVID19.

Presentation — Getting ready to present my project remotely at the University of Toronto, Deep Learning Class

I have presented below a very simple description of my project that I undertook as a part of the School of Continuing Studies course on Deep Learning at the University of Toronto.

Project Goal

Provide accurate, quick and real-time face detection in standardized, off-the-shelf telepresence robots to identify and recognize patients to deliver personalized health monitoring.

Off-the-shelf robots have limited computational power and having a machine learning algorithm that can identify and recognize faces in this constrained environment would be a big advantage.

How do we recognize human faces?

Neurons in our brain respond to particular features of a face and allow us to identify a person. We store these features and use them to recreate the face and recognize it when we see it again.

Features humans use to identify and recognize other human faces

Use of Autoencoders Neural Network for Face Detection

Autoencoder makes it possible to efficiently compress input images and encode data. The decoder then reconstructs the original image using the compressed encoded data. And the goal is that the recreated image has to be as close as possible to the original image.

I transferred this model to one of my home-made robots: ARTEMIS to do a real-time prediction of faces and it worked very well and was able to clearly distinguish the features between two faces.