Using an End-to-End Convolutional Network on Radar Signal for Human Activity Classification

Almost all existing methods for human activity classification based on micro-Doppler radar first manually convert the raw radar signal into a spectrogram using a short time Fourier transform. Then, the spectrogram features are either manually extracted using hand-crafted feature engineering or automatically extracted using deep convolutional networks and fed into a classifier such as a support vector machine, k-nearest neighbor or multi-layer perceptron. However, the optimality of this two-step process is limited by the use of spectrograms, which are a hand-crafted representation. In this paper, the first time truly end-to-end deep network that incorporates the signal representation process into the network is proposed. In the proposed network, which is called RadarNet, two one-dimensional convolutional layers are used to replace short time Fourier transform to obtain a learned radar signal representation. The experimental results show that the proposed RadarNet can achieve 96.35% accuracy in human sleep activity classification and 96.31% accuracy in human daily activity classification, which is 1.96% and 3.26% higher than those of the best existing method, respectively.