IR-ST: A Lightweight Transformer Network for Human Fall Detection Based on FMCW Radar

Without timely treatment, falls often cause serious injuries in elderly individuals. Due to the aging of the global population, an efficient and privacy preserving fall detection system is exceptionally indispensable. According to the above requirements, this article proposes a lightweight neural network inverted residual module and Swin-Transformer module (IR-ST) based on the combination of inverted residual module and Swin-Transformer block. First, the original data are collected by a millimeter-wave radar, and the range-Doppler-average matrix (RDAM) method is used to extract the micro-Doppler map. Second, the spatial feature information in the radar micro-Doppler map is extracted using the inverted residual module. The self-attention mechanism in the Swin-Transformer module is used to learn timing feature information. Finally, the classification results are generated and displayed after processing the data through the fully connected layer. The experiments indicate that the IR-ST neural network exhibits superior accuracy in fall detection, with optimized model parameters that result in reduced prediction time.