A Lightweight Multiscale Neural Network for Indoor Human Activity Recognition Based on Macro and Micro-Doppler Features

Through-the-wall radar (TWR) achieves indoor human activity recognition (HAR) by extracting Doppler and micro-Doppler features. However, the conventional deep learning-based HAR methods have the shortcomings of low accuracy and long inference time. To solve these problems, a lightweight multiscale neural network for indoor HAR based on macro and micro-Doppler features (TWR-FMSN) is proposed in this article. In the proposed method, the trajectories of macroscopic Doppler and microscopic Doppler features are defined first and the integrated models are applied to label the trajectories at both scales for recognition. An efficient attention-mechanism-based lightweight target detection neural network with the Lagrangian trajectory estimation is proposed to obtain macro-Doppler features of human motion. In addition, a kernel-distance-based micro-Doppler labeling method is utilized to obtain the micro-Doppler features of human motion. Finally, all the extracted macro-Doppler and micro-Doppler features are concatenated together for the decision of indoor HAR. The effectiveness of the proposed method is verified by experiments, and the results show that the proposed method can significantly reduce the inference time while retaining high recognition accuracy, which shows great potential in real-time deployment for the practical application.