Through-wall human activity recognition via radar

Abstract

Through-wall human activity recognition (TW-HAR) has important application value in intelligent security, medical monitoring and human-computer interaction. However, signal attenuation, multipath effect and noise interference in through-wall environment seriously affect the recognition accuracy. To solve these problems, this thesis proposes a hybrid deep learning model based on GoogLeNet and Swin Transformer to optimize the feature extraction and classification performance of radar signals. This method combines GoogLeNet's local feature extraction capability and Swin Transformer's global spatiotemporal modeling capability to identify multi-person activities through micro-Doppler spectrogram.

Based on frequency-modulated continuous wave~(FMCW) radar data, the experiments were evaluated in barrier-free, through-wall and cross-environment generalization test scenarios. The results show that the recognition accuracy of the model reaches 97.50\% in barrier-free environment and 95.88\% in through-wall environment, and shows strong generalization ability in cross-environment test. Compared with the traditional CNN method, the model is more robust in complex scenes. This dissertation proves the effectiveness of deep learning in the TW-HAR task, and provides a new idea for the application of radar sensing technology in HAR.