Study of a Health Monitoring System for Elderly Living Alone

—Action Recognition using LSTM

Abstract

The purpose of this study is to develop a system to watch over the increasing number of elderly people living alone. We aimed to construct a system with high privacy protection and a high recognition rate by using LSTM based on skeletal coordinates obtained from images. In this study, LSTM models were trained with various optimizers and learning rates, and the model with the best results was verified. Furthermore, we constructed a system that can recognize actions in real time using LSTM models. As a result, we recorded a maximum accuracy of 0.999886 when evaluating the LSTM model on the split training dataset, and a maximum average recall of 0.834148 on the validation dataset that included only the abnormal actions that we captured. In addition, a client-server type system was constructed, and only skeletal information is used for mutual communication to ensure a high degree of privacy protection.