UAV‐Radar‐Based Through‐Wall Human Activity Recognition via Frequency‐Selective STFT‐ResNet

ABSTRACT This letter proposes a robust human activity recognition approach for through‐wall sensing in high‐rise buildings using an unmanned aerial vehicle (UAV)‐mounted radar. The raw radar echoes are first processed through clutter suppression, coherent accumulation, and region‐of‐interest extraction, after which time–frequency spectrograms are generated via the Short‐Time Fourier Transform. To enhance spectral quality, a ResNet18‐based frequency selection module is employed to eliminate noisy or invalid subcarriers. The refined spectrograms are subsequently classified using another ResNet18 network. Experiments conducted through a 24 cm brick wall yield a recognition accuracy of 98.08% on the test set for three common activities, namely standing, squatting, and lying prone.