Quantitative gait analysis provides critical objective biomarkers for monitoring motor impairments in Parkinson’s Disease (PD), overcoming the intrinsic subjectivity of traditional clinical rating scales such as the MDS-UPDRS. While marker-based optoelectronic motion capture (OMC) remains the clinical gold standard, its extensive setup times induce severe physical and cognitive fatigue in parkinsonian patients, frequently altering their natural gait patterns and compromising the ecological validity of the trials. To address these limitations, this thesis presents the development and validation of an engineering pipeline that contrasts traditional marker-based stereophotogrammetry against an advanced, AI-driven markerless approach leveraging Meta's SAM3D foundation model. To overcome the pervasive issue of Identity Switching (ID swapping) inherent to standard computer vision models when tracking pathological gaits, a custom software subsystem was engineered. This architecture introduces four user-selectable tracking modes, culminating in a robust spatial centroid distance minimization algorithm. The optimized pipeline was successfully deployed to analyze video data from 6 Parkinsonian subjects across frontal and sagittal camera views. The resulting high-density 3D meshes were ingested into a custom MATLAB framework designed to isolate virtual anatomical landmarks, filter coordinates, and reconstruct synchronized bilateral kinematics. The extracted joint kinematics were compared against both the optoelectronic baseline and a healthy reference cohort to map significant biomechanical discrepancies and tracking offsets. The findings demonstrate that while AI-driven architectures still face inherent precision boundaries, the integration of custom spatial tracking constraints significantly mitigates systematic tracking losses. This work highlights the concrete potential of custom-enhanced markerless modeling as a low-cost, non-invasive, and robust tool capable of expanding objective motion analysis from standard laboratory configurations into broader practical and clinical applications.
L'analisi quantitativa del cammino fornisce biomarcatori oggettivi fondamentali per il monitoraggio dei deficit motori nel morbo di Parkinson (PD), superando la soggettività intrinseca delle scale di valutazione clinica tradizionali come la MDS-UPDRS. Sebbene la cattura del movimento optoelettronica (OMC) basata su marker rimanga il gold standard clinico, i suoi lunghi tempi di setup causano un forte affaticamento fisico e cognitivo nei pazienti parkinsoniani, alterando frequentemente i loro pattern di cammino naturali e compromettendo la validità ecologica delle prove. Per far fronte a queste limitazioni, questa tesi presenta lo sviluppo e la validazione di una pipeline ingegneristica che confronta la stereofotogrammetria tradizionale basata su marker con un approccio markerless avanzato, guidato dall'intelligenza artificiale, che sfrutta il modello di fondazione SAM3D di Meta. Per superare il problema pervasivo dell'Identity Switching (ID swapping), intrinseco ai modelli standard di computer vision durante il tracciamento di cammini patologici, è stato progettato un sottosistema software personalizzato. Questa architettura introduce quattro modalità di tracciamento selezionabili dall'utente, che culminano in un robusto algoritmo di minimizzazione della distanza del centroide spaziale. La pipeline ottimizzata è stata impiegata con successo per analizzare i dati video di 6 soggetti parkinsoniani attraverso inquadrature telecamera frontali e sagittali. Le mesh 3D ad alta densità risultanti sono state importate in un framework MATLAB personalizzato, progettato per isolare i punti di repere anatomici virtuali (landmark), filtrare le coordinate e ricostruire le cinematiche bilaterali sincronizzate. Le cinematiche articolari estratte sono state confrontate sia con il baseline optoelettronico sia con una coorte di riferimento sana, al fine di mappare discrepanze biomeccaniche e offset di tracciamento significativi. I risultati dimostrano che, sebbene le architetture guidate dall'IA presentino ancora limiti intrinseci di precisione, l'integrazione di vincoli di tracciamento spaziale personalizzati riduce significativamente le perdite sistematiche di tracciamento. Questo lavoro evidenzia il potenziale concreto della modellazione markerless ottimizzata in modo personalizzato (custom-enhanced) come strumento a basso costo, non invasivo e robusto, in grado di estendere l'analisi oggettiva del movimento dalle configurazioni standard di laboratorio a più ampie applicazioni pratiche e cliniche.
Comparative Kinematic Analysis of Parkinsonian Gait: Marker-Based Stereophotogrammetry Versus Custom-Enhanced SAM3D Markerless Modelling
TOMASELLO, FILIPPO
2025/2026
Abstract
Quantitative gait analysis provides critical objective biomarkers for monitoring motor impairments in Parkinson’s Disease (PD), overcoming the intrinsic subjectivity of traditional clinical rating scales such as the MDS-UPDRS. While marker-based optoelectronic motion capture (OMC) remains the clinical gold standard, its extensive setup times induce severe physical and cognitive fatigue in parkinsonian patients, frequently altering their natural gait patterns and compromising the ecological validity of the trials. To address these limitations, this thesis presents the development and validation of an engineering pipeline that contrasts traditional marker-based stereophotogrammetry against an advanced, AI-driven markerless approach leveraging Meta's SAM3D foundation model. To overcome the pervasive issue of Identity Switching (ID swapping) inherent to standard computer vision models when tracking pathological gaits, a custom software subsystem was engineered. This architecture introduces four user-selectable tracking modes, culminating in a robust spatial centroid distance minimization algorithm. The optimized pipeline was successfully deployed to analyze video data from 6 Parkinsonian subjects across frontal and sagittal camera views. The resulting high-density 3D meshes were ingested into a custom MATLAB framework designed to isolate virtual anatomical landmarks, filter coordinates, and reconstruct synchronized bilateral kinematics. The extracted joint kinematics were compared against both the optoelectronic baseline and a healthy reference cohort to map significant biomechanical discrepancies and tracking offsets. The findings demonstrate that while AI-driven architectures still face inherent precision boundaries, the integration of custom spatial tracking constraints significantly mitigates systematic tracking losses. This work highlights the concrete potential of custom-enhanced markerless modeling as a low-cost, non-invasive, and robust tool capable of expanding objective motion analysis from standard laboratory configurations into broader practical and clinical applications.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tomasello Filippo, PowerPoint Discussione.pdf
accesso aperto
Dimensione
3.03 MB
Formato
Adobe PDF
|
3.03 MB | Adobe PDF | Visualizza/Apri |
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/20.500.12608/115075