Through-Wall Human Activity Recognition Based on Adaptive Doppler Feature Enhancement

Through-Wall human activity recognition (HAR) holds significant potential in fields such as intelligent security and human–computer interaction. Due to the presence of walls, clutter suppression is an inevitable step in concealed HAR. However, existing studies overlook the impact of clutter suppression methods on the information of through-wall human activities, leading to incomplete Doppler representations of target behaviors. In this study, we systematically analyze the suppression characteristics of common clutter suppression methods, including two-pulse cancellation and average cancellation. The analysis reveals that two-pulse cancellation highlights limb information, while average cancellation preserves torso information. These two methods exhibit complementary characteristics in the representation of human activity information. To fully exploit the complementary characteristics of two-pulse cancellation and average cancellation spectrograms for constructing a more comprehensive Doppler representation, we propose the adaptive Doppler feature enhancement network (ADFE-Net). Specifically, the two-pulse cancellation spectrogram and average cancellation spectrogram are processed through an adaptive Doppler feature enhancement module (ADFEM), where frequency weighting factors are adaptively learned to fuse and enhance complementary frequency components to generate Doppler-enhanced spectrogram. The enhanced spectrogram is subsequently input into the 3D Coordinate Attention (3DCA)-ConvMixer network to obtain recognition results. Experimental results demonstrate that compared to existing methods, the recognition accuracy improved by 4.24% with proposed ADFE-Net.