A Two-Stage Low-Complexity Human Sleep Motion Classification Method Using IR-UWB

With increasing concern about sleep quality, human sleep motion recognition has come into sight. Recent years have witnessed that almost all researches use deep neural networks for human motion recognition based on radar sensors. However, it is difficult to deploy these methods in edge computing in the case of limited hardware resources. This paper proposes a two-stage low-complexity method for real-time human sleep motion classification using an ultra-wideband (UWB) radar, which can be applied in edge computing. Firstly, we implement power burst curve (PBC) detection to decide whether sleep motions occur. Secondly, a method named zero-forcing processing for radar spectrum is proposed in order to reduce the number of parameters of classifiers. Later, in stage one, we propose a three-feature extraction approach based on micro-Doppler signatures and utilize support vector machine (SVM) to roughly classify micro-motion and body-movement motions. In stage two, a long short-term memory (LSTM) neural network with fewer parameters is applied to further classify body movements. Experimental results show that the proposed method has a low computational complexity, and can reach the average accuracy of the five typical sleep motions over 97% using train and test split samples, and 91.50% using the leave-two-subject-out (LTSO) samples.