A Dual-Stream CNN-BiLSTM for Human Motion Recognition with Raw Radar Data

With the development of science and technology, human motion recognition as an important auxiliary technology for smart home has important research significance and broad application prospects in the fields of security monitoring, family aging and human-computer interaction. At present, most of the traditional radar-based human motion recognition methods perform multi-dimensional Fast Fourier Transform (FFT) on human motion echo signals to obtain distance, Doppler, angle information and to construct various data spectra, which are fed into the neural networks for classification and recognition. However, the data pre-processing process is complicated. To address this problem, our paper proposes a dual-stream network structure model, which consists of convolutional neural network (CNN) and bi-directional long and short-term memory network (BiLSTM) for human motion recognition with radar raw data. First, digital compensation technology and background frame difference method are used to preprocess the original in-phase/orthogonal (I/Q) raw data to eliminate static interference. Then the processed I/Q signal are converted into amplitude/phase (A/P) data format. Finally, the processed I/Q signal and A/P signal are simultaneously fed into the network structure to extract the temporal and spatial characteristics of human motions, and then the dual-flow network is integrated to enhance the feature interaction and improve the recognition accuracy. The experimental results show that the method is simple in data pre-processing, makes full use of the inter-frame correlation of action data, has fast model convergence, and achieves 98.57% recognition accuracy.