Cross-Attention Augmented End-to-End Architecture for SISO FMCW Radar Activity Recognition With Swin-Transformer

Frequency-Modulated Continuous-Wave (FMCW) radar is extensively used in human activity recognition (HAR) due to its non-intrusive operation, all-weather performance, and anti-interference ability. However, FMCW signals predominantly relies on the Discrete Fourier Transform (DFT) method, which is limited by its fixed basis functions and restricted time-frequency resolution (TFR). Therefore, this paper proposes a cross-attention augmented end-to-end architecture for FMCW radar HAR called CASA-SWIN. The preprocessing module of CASA-SWIN learns the potential Range-Time (R-T), Range-Doppler (R-D), Dopper-Time(D-T) features of rawdata through multiple complex fully connected layers with preset parameters. Furthermore, it utilizes a three-branch cross-attention mechanism (TBCAM) module to extract intrinsic related information for R-T, D-T, R-D branches. For the D-T branch, proposed multi-window filtering module (MWFM) comprises multiple 1D complex convolutions with varying kernel sizes, emulating different window lengths in the Short-Time Fourier Transform (STFT). Finally, a three-branch swin-transformer is used as the HAR classifier. Moreover, proposed CASA-SWIN is evaluated on self-collected HAR, Gait recognition datasets, and open-source gesture recognition datasets. Compared to 12 advanced HAR algorithms, a series of experiments including comparative experiments, cross-environment experiments, open-source datasets experiments, and ablation experiments verify the effectiveness of the proposed module and CASA-SWIN.