RTFDNet: ResNet-Transformer-Based Fall Detection Network Using MmWave Radar

Fall detection plays a crucial role in safeguarding the health and safety of the elderly, representing a vital area within public health and a key application in smart health and intelligent elderly care, which holds significant research value. Traditional fall detection methods relying on cameras, inertial sensors, and other technologies encounter challenges such as privacy concerns and relatively low detection accuracy. In contrast, millimeter wave radar shows great promise in fall detection due to its high penetrability, robust anti-interference capabilities, high resolution, and non - contact monitoring features. However, current research in millimeter - wave radar - based fall detection has several limitations. Most existing methods mainly utilize deep convolutional neural networks to process two - dimensional images, failing to fully exploit the abundant information within 3D point cloud data. Moreover, the fall detection system often struggles to handle dynamically changing interference sources effectively, which severely reduces the accuracy of subsequent models in recognizing fall - related action features. Additionally, the process of data labeling and its associated costs impose substantial constraints on model training and optimization. To address these issues, this study proposes a novel fall detection network, RTFDNet, which combines a deep residual network with the Transformer architecture for the first time. This integration enables the network to comprehensively capture the spatial and temporal characteristics of 3D point cloud data, ensuring a more complete and accurate extraction of essential data information. In terms of data processing, the development of dynamic denoising and physical enhancement strategies significantly improves data quality and the model's generalization ability in complex environments. Furthermore, a semi-supervised learning framework is adopted, requiring only 30 % of the data to be labeled. The framework generates pseudo-labels with the help of model prediction to complete the training. Experimental results demonstrate the effectiveness of the proposed framework. In a four-class classification task, the model achieves a fall detection accuracy of 99.08%, thus providing a new concept for millimeter wave radar fall detection.