Privacy monitoring and abnormal posture detection for home-based elderly care using virtual avatars

In the context of home-based elderly care, abnormal posture monitoring systems must detect human postures in real-time, which raises legitimate privacy concerns. This paper proposes NARF-YOLOv8, an enhanced detection model based on virtual avatar representations of the elderly. To address the challenges of low recognition accuracy, computational redundancy, and insufficient sensitivity to small targets, the model incorporates three key optimizations: the NAM attention mechanism for adaptive feature selection, the RFAConv module to enhance feature extraction under complex backgrounds, and the Focal-SIOU loss function to improve small-target detection. On the self-built APVH (Abnormal Postures of Virtual Human) dataset, NARF-YOLOv8 achieves 97.7% mAP and 95.2 FPS on a server-grade GPU, outperforming the baseline YOLOv8. Cross-domain experiments on the RHPD (Real-World Human Pose Dataset) reveal a substantial virtual-real domain gap; however, fine-tuning with only 20% real data improves performance to 79.3%, demonstrating that the model possesses domain adaptation capability. The virtual avatar mechanism is adopted as a privacy-preserving design strategy to alleviate elderly users’ potential reluctance toward continuous visual monitoring. While the reported real-time performance is achieved with GPU acceleration and further optimization is required for edge-device deployment, this study explores a technical pathway for privacy-friendly smart elderly care systems.