Aplicação de Algoritmos de Aprendizagem Profunda com Foco na Detecção de Quedas Através da Classificação de Imagens Térmicas

Abstract

Falls tend to cause physical and psychological damage, especially for people over 65 years of age. Such damage can be aggravated if they live alone and do not get help in the right time. This is one of the motivations for investigating smart devices in homes, acting as a way of monitoring some action or incident. Machine learning techniques have been applied to increase the effectiveness of these devices by looking for patterns in audio signals and in images. When using DNN (Deep Neural Networks) the success rates increase, but these solutions usually use cloud processing, which can bring a feeling of insecurity to the user. This work investigates an environment monitoring system, capable of detecting fall incidents, without invading the user's privacy, as it is based on low resolution thermal images. Some DNN models were tested to choose a model that best fits the problem, looking for architectures with few processing layers so that they can be embedded for local processing. Images from two different sensors were used, resized to the standard size of 64x64 pixels, resulting in an average accuracy of 91.66% for the bidirectional LSTM (Long short-term memory) network, and of 97.88 % using the VGG16 network.