Non-invasive AI-driven human activity recognition

Ayaz, Fahad (2025) Non-invasive AI-driven human activity recognition. PhD thesis, University of Glasgow.

Abstract

The rapid proliferation of Internet of Things technologies, coupled with artificial intelligence driven applications, has revolutionised human activity recognition, enabling pervasive real-time monitoring across smart homes, healthcare, security, and ambient-assisted living environments. This transformation holds particular significance for healthcare systems, as radar-based recognition of physical and physiological activities facilitates continuous remote monitoring through invasive and non-invasive technologies, supporting personalised care and early intervention at scale. Traditionally, activity recognition systems have relied primarily on invasive or contact based devices, such as wearables and biosensors, which often lead to user discomfort, require frequent maintenance or charging, and risk non-compliance, especially among elderly individuals. Conversely, cameras, Wi-Fi, and radar are all treated as non-invasive sensing modalities; however, cameras raise serious privacy concerns and are constrained by lighting conditions, whereas Wi-Fi-based sensing suffers from multipath interference and spectrum-sharing challenges. Radar sensing emerges as a promising tool and privacy-preserving alternative with robustness to environmental variations. Despite these advantages, systems built on radar for activity recognition face significant challenges in real-world applications. This thesis addresses three critical challenges in radar-based human activity recognition: enabling non-intrusive recognition of both macro-level physical activities (e.g., falls, gait) and micro-level physiological signals (e.g., heart rate, respiration rate); data diversity and radar domain adaptation; and ensuring energy-efficient, privacy-aware edge deployment. The first contribution addresses the challenges of non-intrusive recognition of macro-level human activities and radar domain adaptation by developing a radar signal processing framework that transforms complex signals into four two-dimensional domain representations for robust activity recognition. By integrating domain-specific preprocessing with transfer learning, the framework improves adaptability across environments and reduces the complexity of raw signal data. Experimental results show up to 29.36% improvement in recognition accuracy compared to a baseline convolutional neural network, with transfer learning models achieving 96.03% on the primary dataset and demonstrating strong generalisation across two additional radar datasets. Building upon this, the second contribution focuses on finer-grained sensing, addressing the challenge of non-intrusive monitoring of micro-level physiological signals by extending the system to support radar-based extraction of vital signs such as heart rate and respiration rate. This is achieved using two radar modalities: ultra-wideband and millimetre-wave frequency-modulated continuous wave radar. A comprehensive analysis was conducted to evaluate the impact of varying distances and radar positioning configurations on the accuracy of vital sign extraction. The third contribution addresses the challenges of domain adaptation and energy efficiency by optimising transfer learning models for lightweight, energy-efficient, and privacy-aware deployment on edge devices. Using post-training quantisation and selective domain-model pairing, the system significantly reduces computational costs while maintaining high recognition performance across radar domains. Results indicate energy consumption as low as 0.42 mWh and response times of 1.32 seconds for 5-second activities, confirming its suitability for real-time, on-device monitoring. Additionally, the framework incorporates differential privacy techniques to strengthen local inference privacy with minimal loss in accuracy. Collectively, these contributions enhance the scalability, robustness, and efficiency of activity recognition systems, paving the way for non-invasive, AI-driven applications in healthcare and real-world environments.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Keywords: Human activity recognition, macro- and micro-level activity monitoring, radar sensing, transfer learning, energy efficiency, edge computing, radar domain adaptation.
• Subjects: T Technology > T Technology (General)
• Colleges/Schools: College of Science and Engineering > School of Engineering
• Supervisor's Name: Zoha, Dr. Ahmed
• Date of Award: 2025
• Depositing User: Theses Team
• Unique ID: glathesis:2025-85362
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 07 Aug 2025 13:24
• Last Modified: 07 Aug 2025 13:30
• Thesis DOI: 10.5525/gla.thesis.85362
• URI: https://theses.gla.ac.uk/id/eprint/85362

Related URLs

• Article DOI
• Article DOI
• Conference proceeding
• Conference proceeding
• Conference proceeding
• Conference proceeding
• Conference proceeding
• Book section