Abstract
In Japan, the rapid aging of the population has led to a critical shortage of caregiving staff, necessitating the urgent improvement of efficiency in care environments. Traditional video surveillance systems pose privacy concerns and have limitations in effectively monitoring elderly patients. In this study, we propose a non-contact, privacy-preserving system for monitoring and recognizing the Actions of elderly patients using depth cameras and human
pose estimation techniques. Specifically, we employ depth cameras to collect data and utilize bounding box detection (BB) and human pose estimation (Keypoint) methods to analyze elderly Actions. To achieve detailed action recognition, we compare the effectiveness of the BB method with the Keypoint method and apply a Graph Convolutional Network (GCN) to model the relationships between body joints. Furthermore, we use a Transformer model for action recognition and compare two preprocessing methods—standardization and Min-Max scaling—to develop a more accurate action estimation model.
