Efficient Learning Methods for Human Sensing at the Wireless Edge: from Sparse Input to Spike-based Processing

Contactless perception of human activity through Radio Frequency (RF) sensing has the potential to enable a new generation of unobtrusive monitoring and interaction systems. By exploiting the reflections of transmitted signals, RF-based approaches provide a privacy-preserving alternative to camera systems, while being robust to lighting conditions and able to sense through common occlusions such as walls or furniture. These properties make RF human sensing attractive for a wide range of applications, from healthcare and assisted living to security, human-computer interaction, and crowd monitoring. At the same time, the growing proliferation of connected devices and the shift toward edge computing highlight the need for solutions that are not only accurate, but also energy-efficient and deployable under strict resource constraints. Despite its promising properties, RF-based human sensing remains a challenging problem. The reflections produced by complex human movements result in high-dimensional, noisy, and incomplete signal representations. Traditional feature extraction methods struggle to capture the fine-grained temporal and spatial dynamics of human motion, while modern DL (Deep Learning) approaches, although powerful, often require large models and significant computational resources that hinder their applicability in resource-constrained environments. Addressing these challenges requires novel approaches that combine accurate signal representations with efficient learning models tailored to real-world deployment. This thesis makes contributions along two converging research directions. The first focuses on learning methods for RF-based human sensing, addressing how to derive and reconstruct informative signal representations from noisy, sparse data. We introduce novel algorithms for gait recognition from sparse millimeter-wave (mmWave) radar point clouds, addressing the open-set classification problem, where systems must handle unseen identities at inference time. We also introduce a lightweight and interpretable approach for reconstructing micro-Doppler signatures in the context of Joint Communication and Sensing (JCS), enabling accurate spectral recovery from incomplete channel measurements with minimal communication overhead. These contributions advance the state of the art in RF human sensing by proposing methods that are robust to sparse, noisy, or irregular data, while remaining suitable for edge deployment. The second research direction focuses on the use of Spiking Neural Networks (SNNs) for RF-based human sensing, motivated by their promise of order-of-magnitude reductions in energy consumption when deployed on neuromorphic hardware. We investigate the problem of spike encoding, introducing a data-driven approach to encode RF channel responses into sparse spike trains while explicitly controlling the trade-off between accuracy and sparsity. Furthermore, we explore the application of SNNs to complex human activity recognition tasks, using aircraft marshalling gestures as a representative case study. Through hybrid architectures combining convolutional modules with spiking neurons, we analyze the advantages and drawbacks of SNNs in comparison to conventional Artificial Neural Networks (ANNs), with particular attention to their relevance for energy-efficient edge computing. The methodology adopted in this thesis combines principled signal modeling, data-driven learning, and neuromorphic-inspired computation. By advancing signal representation, spike encoding, and spiking architectures for RF human sensing, this work contributes to the broader goal of enabling accurate, robust, and energy-efficient human sensing systems suitable for deployment in real-world, resource-constrained environments.