Elderly fall detection based on skeleton data using OpenPose

Falls are a major concern for the elderly, often leading to severe injuries and reduced quality of life. This paper presents a fall detection approach utilizing skeletal data extracted by OpenPose and classified using a Graph Neural Network (GNN). The proposed method effectively captures the spatial relationships between human joints, allowing for precise differentiation between fall and non-fall actions. A publicly available dataset consisting of 8,655 images was used to train and test the model, with the GNN achieving a 90% accuracy. Comparative experiments demonstrated that the GNN outperforms a traditional Convolutional Neural Network (CNN), which achieved only 75% accuracy. The results indicate that the combination of skeletal data and GNN provides a reliable method for fall detection, with potential applications in elderly care systems.