SmartFallNet: A Vision Transformer and GRU‐Based Dynamic Model With Adaptive Kernel Attention for Precision Fall Detection

Abstract

ABSTRACT Falls among the elderly remain a critical public health concern, often leading to severe injuries or fatalities. In response, we propose SmartFallNet, an advanced deep learning framework designed for accurate and real‐time fall detection in elderly care settings. The architecture leverages a dual‐stream design that independently processes multimodal data from both RGB and depth frames using Vision Transformers (ViTs) for spatial feature encoding and Gated Recurrent Units (GRUs) for temporal sequence modeling. To enhance the model's sensitivity to fall‐relevant patterns, we introduce a novel Dynamic Kernel Attention Mechanism (DKAM) that selectively emphasizes critical temporal frames. Features from both modalities are fused via a Power‐Weighted Aggregation strategy, resulting in a rich and context‐aware representation of human motion. Evaluated on two benchmark datasets—UR Fall Detection (URFD) and the Falling Detection Dataset (FDD)—SmartFallNet achieved state‐of‐the‐art performance with 99.1% accuracy on URFD and 98.5% on FDD, while maintaining a low inference latency of 0.2 s per sample, supporting near real‐time application. Extensive experiments, including ablation studies, demonstrate the model's robustness, efficiency, and superiority over existing methods. These results underscore the potential of SmartFallNet for deployment in real‐world healthcare and ambient‐assisted living environments.