Falls are a major global cause of accidental death

posing a particular threat to the elderly.A video-based human pose fall detection method is proposed to address the limitations in current technologies in across scenario robustness and real-time performance of.Key innovations include developing a behavior feature extraction(BFE)module which effectively capture dynamic changes and relationships between adjacent skeletal nodes

and improving the spatiotemporal graph convolutional network(ST-GCN)framework.This improvement introduces a dynamic graph convolution network(D-GCN) module and a multi-scale temporal convolution network(MS-TCN)module

which enables effective utilization of skeletal features to capture motion information and long-range temporal dependencies.Experiments show that this method maintains high accuracy(up to 99.45%) across multiple public datasets and demonstrates excellent cross-scenario generalization ability.Additionally

the model achieves real-time performance of 56fps on non-GPU platforms and 92fps on GPU platforms.