This Master’s thesis project was conceived and developed during an Erasmus+ mobility period at Katholieke Universiteit Leuven (Belgium), within the Department of Movement and Rehabilitation Sciences. The experience abroad represented an important opportunity for academic and personal growth, allowing collaboration with the Human Movement Bioengineering research group and deepening knowledge on topics related to human movement biomechanics. The work is part of the scientific collaboration between the University of Padua and KU Leuven, which promotes the exchange of knowledge and innovative methodologies in the field of biomedical engineering and movement sciences. In recent years, biomechanical measurement has progressively moved beyond the laboratory environment, largely thanks to the use of wearable sensors and markerless technologies. This transition toward “out-of-lab” analysis aims to obtain data that are more accessible, continuous, and representative of real-life movement conditions. In this context, estimating Ground Reaction Forces (GRF) without the use of traditional force plates—considered the gold standard for reliability and precision—remains an open challenge. This thesis addresses this challenge by estimating three-dimensional GRFs from data acquired through instrumented insoles, using deep learning techniques. The importance of this challenge clearly emerges in clinical and sport domains. GRFs represent a key indicator for assessing gait function and motor performance: they allow the identification of pathological patterns, the monitoring of musculoskeletal conditions over time, the evaluation of rehabilitative or reconditioning interventions, and they support the optimization of athletic technique and injury prevention through the analysis of mechanical loads acting on the locomotor system. Moreover, GRFs serve as an essential input for the development of realistic musculoskeletal models, which are crucial for estimating internal forces, joint moments, and muscle activations through optimization techniques or dynamic simulations. The ability to accurately estimate these forces outside the laboratory, using portable and low-cost devices, would therefore represent a decisive step toward continuous, personalized monitoring applicable to both clinical practice and high-level sports settings. The aim is to obtain force estimates comparable to those measured with laboratory instrumentation, while relying on portable and non-invasive devices. To this end, a supervised neural network based on a multilayer LSTM architecture was developed, chosen for its ability to capture the temporal dependencies characteristic of the gait cycle. In a first phase, the model was trained and validated on data collected with force plates, which were used as the standard reference. Subsequently, a cross-domain validation was carried out by applying the pre-trained model in inference only to insole data, with the goal of evaluating its generalization capability and its applicability to the acquisition domain of instrumented insoles. Preliminary results are encouraging: the network demonstrated high reliability in reconstructing the GRF components and solid generalization capabilities, both across different acquisition systems and across subjects not included in the training phase. The main contribution of this work lies not only in the development of an innovative model for estimating GRFs from wearable sensors, but above all in demonstrating its robustness and transferability through cross-domain validation. This approach represents a concrete step toward the implementation of portable biomechanical analysis systems that are accurate and applicable in real-world conditions, supporting the evolution of “out-of-lab” biomechanics toward increasingly widespread clinical and daily use.
Questo progetto di tesi magistrale è stato concepito e sviluppato durante il periodo di mobilità Erasmus+ presso la Katholieke Universiteit Leuven (Belgio), all’interno del Dipartimento di Movement and Rehabilitation Sciences. L’esperienza all’estero ha rappresentato un’importante opportunità di crescita accademica e personale, permettendo di collaborare con il gruppo di ricerca di Human Movement Bioengineering e di approfondire tematiche legate alla biomeccanica del movimento umano. Il lavoro si inserisce nell’ambito della collaborazione scientifica tra l’Università degli Studi di Padova e la KU Leuven, che promuove lo scambio di conoscenze e metodologie innovative nel campo dell’ingegneria biomedica e delle scienze del movimento. Negli ultimi anni, la misurazione biomeccanica si è progressivamente spostata oltre l’ambiente di laboratorio, grazie soprattutto all’impiego di sensori indossabili e tecnologie markerless. Questa transizione verso l’analisi “out-of-lab” mira a ottenere dati più accessibili, continui e rappresentativi delle reali condizioni del movimento umano. In questo contesto, la stima delle forze di reazione al suolo (Ground Reaction Forces, GRF) senza l’utilizzo delle tradizionali piattaforme di forza, considerate lo standard di riferimento per affidabilità e precisione, rappresenta ancora una sfida aperta. La presente tesi affronta questa sfida stimando le GRF tridimensionali a partire da dati acquisiti tramite solette sensorizzate, mediante l’applicazione di tecniche di deep learning. L’importanza di questa sfida emerge in ambito clinico e sportivo. Le GRF costituiscono infatti un indicatore fondamentale per la valutazione funzionale del cammino e della performance motoria: consentono di identificare pattern patologici, monitorare l’evoluzione di condizioni muscoloscheletriche, valutare l’efficacia di interventi riabilitativi e riadattivi, oltre a supportare l’ottimizzazione del gesto atletico e la prevenzione degli infortuni attraverso l’analisi delle sollecitazioni meccaniche sull’apparato locomotore. Inoltre, le GRF rappresentano un input imprescindibile per la costruzione di modelli muscoloscheletrici realistici, essenziali per stimare forze interne, momenti articolari e attivazioni muscolari tramite tecniche di ottimizzazione o simulazioni dinamiche. Riuscire a stimare queste grandezze al di fuori del laboratorio, mediante strumenti portatili e a basso costo, costituirebbe dunque un progresso decisivo verso un monitoraggio continuo, personalizzato e integrato nella pratica clinica e nello sport ad alto livello. L’obiettivo è ottenere forze comparabili a quelle misurate tramite strumentazione laboratoriale, pur operando con dispositivi portatili e non invasivi. A tal fine è stata sviluppata una rete neurale supervisionata basata su un’architettura LSTM multilivello, scelta per la sua capacità di catturare le dipendenze temporali caratteristiche del ciclo del passo. In una prima fase, il modello è stato addestrato e validato su dati provenienti da piattaforme di forza, assumendo tali misurazioni come riferimento standard. Successivamente, è stata condotta una validazione cross-dominio applicando in only inference il modello pre-addestrato ai dati delle solette, al fine di valutarne la capacità di generalizzazione e l’applicabilità al dominio di acquisizione delle solette sensorizzate. I risultati preliminari sono incoraggianti: la rete ha mostrato un’elevata affidabilità nella ricostruzione delle componenti della GRF e una solida capacità di generalizzazione, sia verso sistemi di acquisizione diversi sia verso soggetti non coinvolti nella fase di addestramento. Il contributo principale di questo lavoro risiede non solo nello sviluppo di un modello innovativo per la stima delle GRF tramite sensori indossabili, ma soprattutto nella dimostrazione della sua robustezza e trasferibilità attraverso una validazione cross-dominio.
Ground Reaction Force Estimation from Pressure Insoles Using LSTM Networks: Translating Biomechanics from Laboratory to Real-World Applications.
SCANAGATTA, GLORIA
2024/2025
Abstract
This Master’s thesis project was conceived and developed during an Erasmus+ mobility period at Katholieke Universiteit Leuven (Belgium), within the Department of Movement and Rehabilitation Sciences. The experience abroad represented an important opportunity for academic and personal growth, allowing collaboration with the Human Movement Bioengineering research group and deepening knowledge on topics related to human movement biomechanics. The work is part of the scientific collaboration between the University of Padua and KU Leuven, which promotes the exchange of knowledge and innovative methodologies in the field of biomedical engineering and movement sciences. In recent years, biomechanical measurement has progressively moved beyond the laboratory environment, largely thanks to the use of wearable sensors and markerless technologies. This transition toward “out-of-lab” analysis aims to obtain data that are more accessible, continuous, and representative of real-life movement conditions. In this context, estimating Ground Reaction Forces (GRF) without the use of traditional force plates—considered the gold standard for reliability and precision—remains an open challenge. This thesis addresses this challenge by estimating three-dimensional GRFs from data acquired through instrumented insoles, using deep learning techniques. The importance of this challenge clearly emerges in clinical and sport domains. GRFs represent a key indicator for assessing gait function and motor performance: they allow the identification of pathological patterns, the monitoring of musculoskeletal conditions over time, the evaluation of rehabilitative or reconditioning interventions, and they support the optimization of athletic technique and injury prevention through the analysis of mechanical loads acting on the locomotor system. Moreover, GRFs serve as an essential input for the development of realistic musculoskeletal models, which are crucial for estimating internal forces, joint moments, and muscle activations through optimization techniques or dynamic simulations. The ability to accurately estimate these forces outside the laboratory, using portable and low-cost devices, would therefore represent a decisive step toward continuous, personalized monitoring applicable to both clinical practice and high-level sports settings. The aim is to obtain force estimates comparable to those measured with laboratory instrumentation, while relying on portable and non-invasive devices. To this end, a supervised neural network based on a multilayer LSTM architecture was developed, chosen for its ability to capture the temporal dependencies characteristic of the gait cycle. In a first phase, the model was trained and validated on data collected with force plates, which were used as the standard reference. Subsequently, a cross-domain validation was carried out by applying the pre-trained model in inference only to insole data, with the goal of evaluating its generalization capability and its applicability to the acquisition domain of instrumented insoles. Preliminary results are encouraging: the network demonstrated high reliability in reconstructing the GRF components and solid generalization capabilities, both across different acquisition systems and across subjects not included in the training phase. The main contribution of this work lies not only in the development of an innovative model for estimating GRFs from wearable sensors, but above all in demonstrating its robustness and transferability through cross-domain validation. This approach represents a concrete step toward the implementation of portable biomechanical analysis systems that are accurate and applicable in real-world conditions, supporting the evolution of “out-of-lab” biomechanics toward increasingly widespread clinical and daily use.| File | Dimensione | Formato | |
|---|---|---|---|
|
Scanagatta_Gloria.pdf
Accesso riservato
Dimensione
27.43 MB
Formato
Adobe PDF
|
27.43 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/98953