A Hybrid Spatiotemporal Fusion Model for Real-time Fall Detection

We propose a hybrid spatiotemporal fusion model for real‑time fall detection using monocular camera videos. Per-frame spatial features are extracted by a lightweight CNN with squeeze-and-excitation modules, an AutoEncoder reduces noise to generate core representations, and a two-layer GRU together with a Transformer encoder jointly learns temporal flow and global relations. Multi-task learning that combines classification error (cross-entropy) and reconstruction error (MSE) enables discrimination of subtle pre-fall motions. Evaluated on sequences (30 frames by default) compiled from multiple public datasets with a subject-wise split, the proposed model achieves 97.10% accuracy (67/69), 100% recall for falls (FN=0), and ROC–AUC of 1.00. With 0.91 GFLOPs, 3.22 M parameters, and an average inference delay of 8.16 ms per sample, its usability in edge environments is verified.