GATE: Gated Attention Technology for Ensemble-Based Fall Detection

Fall Detection Systems (FDS) often rely on accelerometer and gyroscope data to identify fall events, with cross-attention mechanisms commonly employed to explore the complementary relationship between these modalities. However, the strength of this relationship may vary in real-world scenarios, leading to suboptimal feature representations when the modalities exhibit weak complementarity. To address this, we propose GATE (Gated Attention Technology for Ensemble-Based Fall Detection), which combines Convolutional Neural Networks (CNNs) with self-attention and cross-attention mechanisms for robust feature extraction and fusion. The CNNs with self-attention capture intra-modality features, while cross-attention mechanisms extract inter-modality interactions. A novel gating mechanism evaluates the strength of the complementary relationship between modalities and adaptively selects either cross-attended or original features, mitigating the impact of noisy or irrelevant information. Finally, the fused features are passed through a classification layer to distinguish fall events from activities of daily living. Extensive experiments on two benchmark datasets validate the effectiveness of GATE. On the MobiFall dataset, GATE achieves an accuracy of 98.89%, specificity of 96.45%, sensitivity of 98.90% and an F1-score of 97.66%. On the MobiAct dataset, it achieves an even higher accuracy of 99.88%, specificity of 99.78%, sensitivity of 99.40% and an F1-score of 99.59%.