The increasing demand for continuous, unobtrusive, and remote monitoring of human movement is shifting biomechanics toward wearable sensors, particularly inertial measurement units (IMUs/MIMUs). While laboratory-based motion capture remains the gold standard, its cost, complexity, and limited acquisition volume prevent long-term and real-world use. Wearable inertial sensors offer a valid alternative, enabling prolonged monitoring with minimal intrusion, but they introduce methodological challenges such as sensor–anatomy alignment, signal drift, and context interpretation, especially in populations with motor impairments. In this context, this thesis investigates the methodological foundations, computational strategies, and practical applications of MIMU-based systems for human motion analysis, focusing on three areas: human activity recognition (HAR), lower-limb joint kinematics, and dynamic stability assessment through the margin of stability (MoS). The aim is to develop accurate, robust, and clinically meaningful wearable solutions for everyday monitoring. Inertial-based HAR has been addressed through two frameworks: one using shallow machine-learning classifiers and another relying on deep-learning models. Both approaches show that a single wrist-mounted MIMU can reliably recognize clinically relevant upper-limb gestures, such as drinking or pill intake. High classification accuracy was achieved even with minimal sensor configurations. Carefully engineered features enabled machine-learning models to match more complex networks, while hybrid deep-learning architectures further improved accuracy and robustness using raw inertial data. The contribution of individual signals and their combinations was also examined. The thesis also explores lower-limb kinematic estimation for gait analysis and rehabilitation. A new anatomical axis identification method was developed, using functional movements and least-squares minimization of instantaneous helical axes to align sensor and anatomical frames. This enabled meaningful and clinically compliant joint angle estimation. The study also assessed whether calibration movements could be reduced without compromising sagittal-plane accuracy, while also evaluating performance in the frontal and transverse planes. Validation extended beyond straight-line gait to include clinically relevant tasks such as turning and the Timed Up and Go, demonstrating the method’s applicability to complex, real-world movements. Dynamic stability was further investigated through MoS estimation using only three inertial sensors. Fully MIMU-based MoS computation is still largely unexplored, and this work contributes to establishing it as a feasible, unobtrusive technique for quantifying dynamic stability—a key predictor of fall risk, especially in older or frail individuals. Although preliminary, the results show good agreement between inertial- and marker-based measures in both healthy and Parkinson’s disease participants, highlighting the potential of this approach and the diagnostic importance of turning tasks. Overall, these contributions support the development of next-generation wearable technologies for early diagnosis, personalized rehabilitation, and long-term monitoring of motor function in real-world environments.
La crescente richiesta di monitoraggi continui, poco invasivi e a distanza del movimento umano sta orientando la biomeccanica verso l’uso di sensori indossabili, in particolare le unità di misura inerziali (IMU/MIMU). Sebbene i sistemi di motion capture da laboratorio restino il riferimento più accurato, costi elevati, complessità e volume di misura limitato ne impediscono l’uso prolungato e in contesti reali. I sensori inerziali rappresentano quindi un’alternativa promettente, pur introducendo sfide metodologiche come l’allineamento sensore–anatomia, la deriva del segnale e l’interpretazione dei dati in vita quotidiana, soprattutto in popolazioni con deficit motori. In questo contesto, questa tesi approfondisce le basi metodologiche, le strategie computazionali e le applicazioni pratiche dei sistemi basati su MIMU per l’analisi del movimento umano, concentrandosi su tre aree: riconoscimento delle attività (HAR), stima della cinematica degli arti inferiori e valutazione della stabilità dinamica tramite il margine di stabilità (MoS). L’obiettivo è sviluppare soluzioni indossabili accurate, robuste e clinicamente utili per il monitoraggio quotidiano. Il tema dell’HAR è stato investigato tramite due approcci: uno basato su algoritmi di machine learning “shallow” e uno su modelli di deep learning. Entrambi mostrano che una singola MIMU al polso può riconoscere gesti clinicamente rilevanti come bere o assumere una pillola. Configurazioni sensoriali essenziali hanno comunque prodotto elevata accuratezza. Caratteristiche ben ingegnerizzate hanno permesso ai modelli più semplici di raggiungere prestazioni comparabili a reti più complesse, mentre architetture ibride di deep learning hanno ulteriormente migliorato accuratezza e robustezza usando direttamente i segnali grezzi. È stato inoltre valutato il contributo informativo delle diverse tipologie di segnale. La tesi analizza anche la stima della cinematica degli arti inferiori per la valutazione del cammino e la riabilitazione. È stato sviluppato un nuovo metodo di identificazione degli assi anatomici, basato su movimenti funzionali e minimizzazione ai minimi quadrati degli assi elicoidali istantanei, per allineare il frame del sensore a quello anatomico e ottenere stime articolari significative e clinicamente coerenti. È stata inoltre valutata la possibilità di ridurre i movimenti richiesti nella fase di calibrazione senza compromettere l’accuratezza nel piano sagittale, analizzando anche le stime nei piani frontale e trasverso. La validazione è stata estesa oltre il cammino lineare, includendo compiti funzionali come la svolta e il Timed Up and Go, confermando l’applicabilità della metodologia a movimenti complessi e reali. Infine, la stabilità dinamica è stata esaminata tramite la stima del MoS utilizzando solo tre sensori inerziali. Un metodo completamente basato su MIMU per questa misura è ancora poco esplorato, e questo lavoro contribuisce a definirne la fattibilità come tecnica non invasiva per quantificare la stabilità dinamica, un indicatore cruciale del rischio di caduta, soprattutto negli anziani e nei soggetti fragili. Sebbene preliminari, i risultati mostrano un buon accordo tra misure inerziali e marker-based sia in soggetti sani sia in persone con Parkinson, evidenziando il potenziale dell’approccio e il ruolo diagnostico dei compiti di svolta. Nel complesso, questi contributi gettano le basi per tecnologie indossabili di nuova generazione a supporto della diagnosi precoce, della riabilitazione personalizzata e del monitoraggio a lungo termine della funzione motoria in contesti reali.
Exploiting inertial sensing for human motion analysis and motor dynamics recognition / Scattolini, Mara. - (2026 Mar).
Exploiting inertial sensing for human motion analysis and motor dynamics recognition
SCATTOLINI, MARA
2026-03-01
Abstract
The increasing demand for continuous, unobtrusive, and remote monitoring of human movement is shifting biomechanics toward wearable sensors, particularly inertial measurement units (IMUs/MIMUs). While laboratory-based motion capture remains the gold standard, its cost, complexity, and limited acquisition volume prevent long-term and real-world use. Wearable inertial sensors offer a valid alternative, enabling prolonged monitoring with minimal intrusion, but they introduce methodological challenges such as sensor–anatomy alignment, signal drift, and context interpretation, especially in populations with motor impairments. In this context, this thesis investigates the methodological foundations, computational strategies, and practical applications of MIMU-based systems for human motion analysis, focusing on three areas: human activity recognition (HAR), lower-limb joint kinematics, and dynamic stability assessment through the margin of stability (MoS). The aim is to develop accurate, robust, and clinically meaningful wearable solutions for everyday monitoring. Inertial-based HAR has been addressed through two frameworks: one using shallow machine-learning classifiers and another relying on deep-learning models. Both approaches show that a single wrist-mounted MIMU can reliably recognize clinically relevant upper-limb gestures, such as drinking or pill intake. High classification accuracy was achieved even with minimal sensor configurations. Carefully engineered features enabled machine-learning models to match more complex networks, while hybrid deep-learning architectures further improved accuracy and robustness using raw inertial data. The contribution of individual signals and their combinations was also examined. The thesis also explores lower-limb kinematic estimation for gait analysis and rehabilitation. A new anatomical axis identification method was developed, using functional movements and least-squares minimization of instantaneous helical axes to align sensor and anatomical frames. This enabled meaningful and clinically compliant joint angle estimation. The study also assessed whether calibration movements could be reduced without compromising sagittal-plane accuracy, while also evaluating performance in the frontal and transverse planes. Validation extended beyond straight-line gait to include clinically relevant tasks such as turning and the Timed Up and Go, demonstrating the method’s applicability to complex, real-world movements. Dynamic stability was further investigated through MoS estimation using only three inertial sensors. Fully MIMU-based MoS computation is still largely unexplored, and this work contributes to establishing it as a feasible, unobtrusive technique for quantifying dynamic stability—a key predictor of fall risk, especially in older or frail individuals. Although preliminary, the results show good agreement between inertial- and marker-based measures in both healthy and Parkinson’s disease participants, highlighting the potential of this approach and the diagnostic importance of turning tasks. Overall, these contributions support the development of next-generation wearable technologies for early diagnosis, personalized rehabilitation, and long-term monitoring of motor function in real-world environments.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


