Filterbank-Based Micro-Doppler Spectrogram Transformation for Human Activity Recognition

Human activity recognition (HAR) utilizing radar sensors has garnered significant attention due to its ability to robustly recognize activities in various environments without infringing on privacy concerns. In prior radar-based HAR studies, recognition has been predominantly conducted using spectrograms generated from received signals. In this paper, we propose a novel time-frequency domain signal representation method specifically designed for the effective analysis of micro-Doppler signatures arising from human activities. The proposed method applies a filterbank capable of nonlinearly transforming the frequency bands of conventional spectrograms, which possess linear frequency bands. This transformation allows for a more precise analysis of micro-Doppler patterns generated by human activities. We evaluated the recognition performance by applying convolutional neural networks to datasets that utilized the proposed filterbank-based spectrograms in comparison with those employing conventional spectrograms. Our results demonstrate that the modified spectrograms achieved up to $\mathbf{5 . 0 2 \%}$ improvement in recognition performance.