Residual Neural Network Driven Human Activity Recognition by Exploiting FMCW Radar

In recent years, radar-based human activity recognition has attracted the interest of a large number of researchers. Many researchers have proposed various effective processing algorithms. However, a good data processing algorithm not only has high recognition accuracy but also should be closer to the real application environment, such as having better detection robustness and detection sensitivity. This paper proposes a residual-bi-LSTM-attention hybrid multi-network, which has high recognition accuracy and the advantages of strong robustness and detection sensitivity. First, we collected data on five different activities of 13 volunteers in a laboratory setting. Then, after processing the collected data through the proposed algorithm, the micro-Doppler characteristics of each action are obtained. Finally, these desired features are fed into the proposed network for classification and recognition. Experimental results confirm the efficiency and accuracy of the proposed algorithm.