Sistema de detecção de quedas de idosos baseado em deep learning

Abstract

Falls in the elderly are a public health problem, due to the aging of the population combined with the serious consequences they can bring. In this situation the need to avoid these problems and mitigate their consequences arises. A fall detection system fits into the second category, allowing the reduction of the time to attendance for the elderly who suffer falls. The work developed consists of the development of a low cost fall detection system based on deep learning for image processing through the steps of obtaining a database of classified images, training of four conventional architectures of neural networks (AlexNet, VGG- 19, GoogleNet and ResNet) using the transfer learning process to compare performance and implementation in embedded systems. The best performance obtained among the used neural networks is with the AlexNet network. When a fall is detected, an SMS message is sent to a monitor user. The system is technically feasible, as it achieved accuracy and recall rates of over 96 % even in adverse situations. The system must go through improvements before becoming a product.