Robotic nurses help patients thanks to ground-breaking AI training
Text by:
ITCL
Robotic nurses can help patients when they perform exercises with their arms or legs. Or see when a patient waves to call for help, or even when a patient falls to the floor. In HosmartAI Pilot 5, these functionalities are being developed and will be tested during the coming year. Its goal is to help hospital staff focus on less routinary tasks, as part of the strategies to deal with the effects of ever-increasing staff shortages.
With its camera and computer vision tech, the nurse detects the posture and movements of the patients. Remarkable here is the way that the underlying Artificial Intelligence is trained. Where normally photos and videos are used to train the neural networks, here synthetic images are used, created with virtual reality techniques.
Over the last years, ITCL Technology Centre (itcl.es) has been developing this technology called SIGEDA – Synthetic Images Generator for Data Augmentation. This is a tool for generating synthetic datasets that allow the rendering of thousands of labelled images and/or videos to train Deep Learning Computer Vision algorithms.
More often than not, too few adequate images/videos are available for training. This service enables AI developers to train their neural networks with labelled images/videos created with virtual reality. For example, for people detection, an enormous variety of people can be generated thanks to the ability of configuration of the VR avatars, avoiding bias and ethics issues like the infamous unbalanced training in terms of gender, age and race. There are no privacy issues as no pictures from real people are used.
Moreover, since these images are created, the precise 3D positions are known (ground truth), instead of unprecise 3D position estimations from 2D photographs.
The developed tool set can be personalized to tackle a wide variety of computer vision jobs for any kind of subject, not only for people. Past implementations include the generation of labelled data for ergonomic pose evaluation for company employees and industrial object localisation & classification.
In HosmartAI, the synthetic videos functionality was included, enabling the detection of movements. SIGEDA is now published as a project result in the Horizon Results Platform:
