Design of a Two-Stage Continuous Fall Detection System Using Multiframe Radar Range–Doppler Maps

This article presents a two-stage radar-based continuous fall detection system designed for high accuracy and low power consumption, suitable for edge device deployment. The system initiates with a barycenter descent detection phase, which screens for potential falls by identifying movements characterized by a downward trajectory of the barycenter. Progressing to the second stage, it leverages a sophisticated neural network, termed the range-Doppler–time network (RDTNet). This network effectively integrates 2-D convolutional neural networks (CNNs) and temporal convolutional networks (TCNs) to meticulously extract essential features related to fall incidents, including range, velocity, and temporal characteristics. Notably, RDTNet enhances efficiency through weight sharing across its convolutional modules, significantly reducing model parameters. Compared with the state-of-the-art method, the proposed model achieves an increase in accuracy by 0.65% and recall by 1.72%, while using only 0.18 times the parameters and 0.0016 times the computational load of the comparison. This model achieves superior precision and recall with minimal parameters and computational overhead. Such a two-stage architecture ensures that the computationally intensive RDTNet activates selectively, thereby markedly reducing power consumption. Experimental results demonstrate that the proposed system not only maintains high accuracy in fall detection with reduced false alarms but also exhibits potential for low-power scenarios.