Human Activity Estimation System Based on Continual Learning Using Skeleton Information

Abstract

In recent years, as the society faces a declining birthrate and an aging population, the shortage of caregivers has increased the demand for elderly care. To accurately grasp the health status of the elderly and provide appropriate care, it is crucial to recognize and analyze human activity over time. To address this issue, this study proposes a continual human activity recognition system that uses joint angles calculated from a 3D human skeleton model to recognize daily activities and infer the similarity of each body part's movements. The proposed system extracts human joint angles based on a 3D human skeleton model generated from depth information obtained from multiple RGB-D cameras. Furthermore, using time-series joint angle data of human activity, the system continually recognizes human activities and estimates the similarity of movements for each body part. To validate the effectiveness of the proposed method, detailed verification experiments were conducted using real-world data.