Evaluating Depth-Based Human Pose Estimation in Real-World Nursing Home Environments

Human pose estimation in real-world indoor environments remains challenging due to privacy concerns, occlusions, varying numbers of people, and noisy sensor data. In this work, we investigate pose estimation from depth images captured over an extended period in a nursing home setting. We explore multiple input representations, including depth-only and silhouette-based representations, as well as depth-based Sobel gradient features. To systematically analyze the impact of model complexity and input representation, we design an experimental framework comparing lightweight Convolutional Neural Networks (CNNs), graph-based models, and deeper architectures, including ResNet and HRNet. Among the evaluated configurations, CNN+GCN with Depth+Sobel input achieved the highest test PCK on the seven-joint baseline, reaching 77.18% in the room-dependent setting and 76.03% in the room-independent setting. Lightweight models generally outperformed higher-capacity architectures on this dataset, but the experiments do not identify the mechanism behind this difference. CNN+LSTM reduced frame-level performance compared to the corresponding frame-wise CNN. Under realistic conditions, performance decreased for partially visible and multi-person scenes. Adding a visibility prediction head produced modest improvements in the evaluated scenarios. Unoccupied-room training reduced spurious pose predictions. The results indicate that lightweight architectures with structural depth representations are promising for privacy-sensitive pose estimation in real-world nursing home environments. However, substantial challenges remain for partial visibility, multi-person scenes, annotation uncertainty for occluded joints, and end-to-end deployment. Computational measurements show low model inference time on the tested hardware but do not establish end-to-end real-time system performance.