Markerless Human Movement Analysis with Artificial Intelligence: From Pose Estimation to Clinical Biomarkers

Human movement is a rich, non-invasive source of information about neural, musculoskeletal, and cognitive function, and its quantitative analysis can provide objective markers for assessing and monitoring neurological and neurodegenerative disorders. Conventional analysis relies on marker-based motion capture: accurate, but costly, complex, and confined to specialised laboratories, which limits its use for large-scale screening and continuous monitoring. Advances in computer vision and artificial intelligence have enabled a video-first paradigm, in which human pose estimation extracts body keypoints directly from ordinary images and video. This thesis argues, however, that pose estimation on its own only localises visual keypoints and does not yield a clinically meaningful description of movement: visual keypoints are optimised for image-level localisation rather than anatomical correspondence, so high visual accuracy does not guarantee biomechanical validity, and uncertainty introduced at the keypoint stage propagates to downstream joint angles, spatio-temporal parameters, and derived biomarkers. Starting from this representational gap, the thesis investigates which computational representations - and which transformations between them - are required to convert markerless visual observations into reliable biomechanical and clinically interpretable descriptions of movement, across both gross and fine motor scales. It develops four contributions. Poseidon, a temporal context-aware architecture for multi-frame human pose estimation, extends ViTPose with adaptive frame weighting, multi-scale feature fusion, and cross-attention, improving robustness to occlusion and motion ambiguity over strong video benchmarks. A study of learning-based mappings from generic 3D keypoints to anatomical landmarks evaluates them by positional accuracy, biomechanical fidelity, robustness, and uncertainty, showing that predictive uncertainty can act as a quality-control layer that ranks frames by expected error. A markerless gait-analysis pipeline for Parkinson's disease derives gross motor biomarkers from accessible RGB video through pose trajectories, feature extraction and selection, machine-learning classification, and explainability; and a handwriting-analysis study for Alzheimer's disease extends the biomarker perspective to fine motor behaviour by integrating multiple tasks and fusing offline and dynamically derived representations. The central conclusion is that markerless movement analysis becomes clinically meaningful only when the representation supports the interpretation: accurate visual keypoints are necessary but not sufficient, because they do not by themselves establish anatomical correspondence, enforce kinematic continuity, or guarantee clinically valid biomarkers. The systems developed here should therefore be understood as research contributions toward screening, monitoring, and decision support rather than as standalone diagnostic tools.