University of Glasgow
Leveraging AI and radar for non-invasive fine-grained human activity monitoring
Abstract
dc:description.abstractHuman-centric monitoring demands sensing technologies that are non-invasive, privacy-preserving, and capable of operating reliably across diverse environments and motion scales. Conventional approaches based on wearable sensors suffer from compliance and long-term usability constraints, while vision-based systems raise privacy concerns and degrade under poor lighting or occlusion. Alternative RF-based modalities, such as Wi-Fi channel state information (CSI), offer privacy-preserving device-free sensing but remain sensitive to environmental dynamics and offer limited spatial resolution. In contrast, frequency-modulated continuous-wave (FMCW) radar provides a favourable balance of robustness, privacy preservation, and fine-motion sensitivity, making it well-suited to the human-monitoring use cases targeted in this thesis, namely, ambient assisted living and in-vehicle driver monitoring, across behavioural and physiological motion scales. This thesis positions radar-based sensing as a unified framework for large-scale and fine-grained human activity monitoring and vital signs estimation, and further examines how behavioural and physiological cues can be complemented through multimodal fusion in realistic settings. First, large-scale human activity monitoring is addressed using FMCW radar representation maps, where multiple radar signal-processing domains and deep learning architectures are systematically evaluated under subject-wise testing. Among the evaluated configurations, MobileNetV2 combined with short-time Fourier transform (STFT) representations achieves a recognition accuracy of 96.30% with the shortest inference time of 2.57 ms per sample, while requiring only 0.22 s to preprocess and generate each STFT-based time–frequency radar map, supporting real-time, low-power deployment. Second, the thesis advances fine-grained monitoring by analysing seated upper-body and head movements using mmWave FMCW radar. Generalisation to unseen subjects is initially limited to 60% accuracy; by applying the proposed data augmentation strategy, performance improves substantially to 91.32%, demonstrating robust cross-subject recognition under data-scarce conditions. In parallel, a motion-robust radar-based vital signs estimation framework is developed, which uses variational mode decomposition to suppress motion artefacts. Under non-stationary conditions, vital signs estimation performance improves by 40.59%, highlighting radar’s ability to recover physiological dynamics despite concurrent movement. Finally, the thesis extends toward multimodal human state monitoring by integrating physiological heart rate with behavioural gaze metrics in real-world driving. Analysis of approximately 480 minutes of naturalistic driving data reveals clear scenario-dependent and subject-specific patterns, and a lightweight decision-level fusion algorithm identifies abnormal driving segments with improved robustness and interpretability compared with unimodal cues, demonstrating effective operation under real-world, unconstrained conditions.
Degree
thesis:*- Level dc:type.qualificationlevel
- PhD
- Grantor dc:publisher.institution
- University of Glasgow
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Alhumaily, Basim
Subjects
dc:subject × 2Rights
- Language dc:language
- en
Identifiers
dc:identifier.*- DOI dc:identifier.doi
- https://doi.org/10.5525/gla.thesis.86167