Design of a Multistage Radar-Based Human Fall Detection System

Deep neural networks (DNN) have recently been introduced to the radar-based fall detection system to achieve high detection accuracy. However, such systems generally suffer the limitation of increased computational complexity and thus increased power consumption. In this work, a novel multi-stage radar-based fall detection system is proposed to maintain high accuracy while keeping the power consumption at a low level. The proposed system consists of three stages. In the first stage, named event detection, a simple threshold-based method is adopted to determine whether there is motion existing or not. In the second stage, a shallow neural network called preliminary screening network (PSN) with extremely low computational complexity is proposed to determine whether such activity is fall-like or not. Finally, the last step contains a DNN with heavily computational complexity, named reconstruction-based fall detector (CRFD), which is applied to determine whether such a fall-like motion is a fall or not. By adopting the proposed multi-stage architecture, the part with the highest computation cost-the CRFD would be inactivated most time and thus can significantly reduce the complexity of the overall fall detection system. The experimental results show that compared with the conventional one-stage method, the proposed multi-stage system can not only achieve high fall detection accuracy but also has potential for deployment in a much lower power mode.