Hybrid graph-temporal modeling with uncertainty-aware Transformers for elderly fall detection

The accurate detection of falls, particularly among elderly populations, remains a critical challenge in smart health monitoring due to the severe physical and psychological consequences of delayed intervention. In this work, a novel hybrid fall detection framework is proposed, which simultaneously captures spatial and temporal dynamics of human motion through the integration of Graph Convolutional Networks (GCNs), Graph Attention Networks (GATs), Inverse Graph Fourier Transform (IGFT), and uncertainty-aware Transformer (UATransformer) architectures. Spatial dependencies are modeled via GCNs to extract local joint interactions, while GATs adaptively prioritize critical joints, enhancing robustness to occlusions, misaligned camera angles, and visually similar motions. Motion features are further refined in the spectral domain using IGFT and multi-scale spectral filtering, allowing the separation of global posture drifts, coordinated joint movements, and high-frequency fall accelerations. Temporal modeling is performed using an uncertainty-aware Transformer, which assigns higher attention to ambiguous or fall-onset frames, enabling early detection of imminent falls. To address class imbalance and improve sensitivity to rare events, a multi-component loss function combining weighted cross-entropy, focal loss, and fall-onset prediction loss is employed. The framework is evaluated on the NTU RGB + D and University of Rzeszow Fall Detection datasets, comprising 114,480 samples across 120 classes of daily activities and 40 daily living activities, as well as 30 fall sequences, achieving accuracies of 99.00% and 99.80%, respectively. Collectively, these contributions yield a model that robustly distinguishes falls from similar actions, anticipates fall onset, and demonstrates strong potential for deployment in clinical and elderly care environments.