A Simulation-Based Framework for the Design of Direction-Independent Human Activity Recognition Systems Using Radar Sensors

Human activity recognition (HAR) systems play an important role in understanding and interpreting human movements across various domains, with applications ranging from automobiles to smart homes and health. This dissertation focuses on HAR within the realm of radio frequency (RF) sensing, with a primary focus on modeling the intricate influence of human motion on wireless channel characteristics, particularly in the context of frequency-modulated continuous wave (FMCW) radar systems. It presents a paradigm shift from experimental- to simulation-based approaches tailored for RF sensor-based HAR systems. The core innovation lies in a sophisticated channel model capable of transforming three-dimensional (3D) trajectories into high-fidelity simulated RF signals, offering substantial control over signal parameters for simulating diverse environmental conditions. This research addresses two main challenges in HAR: accommodating multiple directions of human motion and tackling the scarcity of radar data for diverse scenarios. To overcome motion direction challenges, a distributed multiple-input multipleoutput (MIMO) radar configuration is introduced, capturing multi-perspective radar signatures of multi-directional human activities. The configuration, complemented by the dynamic time warping (DTW) distance metric, facilitates the development of a direction-independent step counting system for multi-directional walking activities. To mitigate the problem of cross-channel interference, a novel range gating method is implemented, leveraging distinct RF delay lines within the distributed MIMO radar setup. This distributed MIMO radar configuration, providing complementary RF sensing, is well-suited for realizing direction-independent human activity recognition (DIHAR) systems. An experimental-based DIHAR system is developed, utilizing the multi-perspective MIMO radar configuration, to classify various multidirectional human activities. The system involves training a machine learning model with a large dataset of radar signatures, necessitating a comprehensive measurement campaign. The dissertation highlights the limitations of experimental data-driven approaches, emphasizing the challenges of acquiring diverse and representative datasets for radarbased classifiers. It advocates simulation-based solutions, offering control over radar parameters, reducing training efforts, addressing user privacy concerns, and enabling the generation of varied training datasets tailore