Human activity classification via trainable fractional fourier transform

Abstract

The human activity classification regarding micro-Doppler characteristics has been a captivating research area since the introduction of the micro-Doppler concept in various fields such as security surveillance, healthcare monitoring/diagnostics, and gait analysis. The extraction of micro-Doppler signatures through radar sensing and subsequent time-frequency analysis of humans or other targets enables precise and reliable characterization of their motion dynamics, allowing for robust classification even in cluttered or low-resolution environments. Compared to conventional vision-based or wearable sensing methods, radar systems are favored due to their inherent ability to preserve target privacy without capturing identifiable visual features. To extract the meaningful micro-Doppler signatures embedded in the radar data, assorted time-frequency analysis approaches have been employed in the literature, especially the Short-Time Fourier Transform (STFT), due to simplicity and computational efficiency. The Fourier transform-based approach might be broadened by utilizing the Fractional Fourier Transform (FrFT), which generalizes the classical Fourier transform by considering the continuum of infinitely many representations between the time and frequency domains. Thanks to the opportunity to work in intermediate domains provided by the Fractional Fourier transform, more distinguishable representations of the micro-Doppler signatures might be obtained. Motivated by the limitations of conventional approaches that depend on fixed time-frequency representations, this thesis proposes a deep learning-based classification methodology for the micro-Doppler signature classification of human activities based on the Fractional Fourier transform. Instead of using the singular usage of the fractional domain representations by searching for optimal transform order empirically, trainable Fractional Fourier transform blocks are engaged in different deep learning architectures. Spectrograms representing the time-frequency characteristics are constructed within the deep-learning architectures, aiming for optimum representations via trainable fractional Fourier transform blocks dynamically. The proposed approach is evaluated with multiple radar-based datasets, including both simulated and real-world measurements with different radars and configurations. A comprehensive set of experiments is conducted with various deep learning architectures, including single-branch CNN, LSTM, and GRU based architectures, with the multi-branch configurations including the simultaneous use of different time-frequency representations obtained via fractional Fourier transform and multi-input Siamase-based models. The proposed approach is compared in detail with the conventional Fourier transform approach in distinct experiments, and the behavior of the fractional representations across the models and datasets is examined. As far as we are aware, this study is the first to combine a trainable Fractional Fourier Transform with micro-Doppler signature-based human activity classification. The experimental results demonstrate that the proposed approach consistently outperforms traditional Fourier Transform-based methods. Moreover, the simultaneous utilization of multiple time-frequency representations with distinct FrFT orders is, to the best of our knowledge, a novel contribution to the literature, yielding considerable performance improvements. It is believed that the findings of this study pave the way for future research on adaptive time-frequency analysis in radar-based sensing applications.