Time-Aware Graph Neural Networks for Human Activity Recognition from mmWave Radar Point Clouds

The rapid advancements in machine learning and computer vision have transformed the way signals and images are processed, opening new opportunities in areas such as autonomous driving, human–machine interaction, healthcare, and surveillance. Among the sensing technologies enabling these applications, millimeter-wave radar has recently gained increasing attention. Compared to cameras and LiDARs, radar sensors are low-cost, robust to adverse weather conditions, and inherently privacy-preserving. Their relevance is further amplified by the forthcoming 6G integrated sensing and communications (ISAC) paradigm, where sensing and wireless connectivity will converge into a unified infrastructure. Despite these advantages, radar data remain challenging to process: they are low-resolution, sparse, and noisy. Traditional approaches often project the raw data into two-dimensional grids (e.g., range–Doppler maps) in order to leverage convolutional neural networks, but this inevitably discards structural information. Graph neural networks (GNNs) have emerged as a promising alternative, as they can directly operate on irregular point clouds while preserving spatial relationships. This thesis investigates the use of GNNs for Human Activity Recognition (HAR) from millimeter-wave radar, proposing an innovative approach for the construction and exploitation of spatio-temporal graphs. By jointly modeling the spatial structure of radar point clouds and their temporal evolution, spatio-temporal graphs provide a more faithful representation of human motion. In parallel, special emphasis is given to computational efficiency, which is crucial for embedded and real-time applications. The proposed spatio-temporal GNN models achieve state-of-the-art performance on benchmark datasets while significantly reducing both parameter count and computational cost. Overall, this work demonstrates the potential of spatio-temporal GNNs to combine accuracy and efficiency in radar-based HAR.