The work focuses on the development of a machine learning framework based on neural networks, which is able to predict the labels of 3D markers from a motion capture system. This should be implemented in an existing application devoted to the biomechanical analysis in the medical and sports field. Starting from a dataset of 3D files acquired in field and in laboratory from an heterogeneity of patients and athletes, containing specific movements, the purpose is to speed up the process of analysis during which somebody until now proceeds with a manual labeling of the 3D cloud of points, before the data can be analyzed. The neural networks framework implemented is based on a strict pre-processing and post-processing voted to clean the data in order to get a better result and to handle the presence of missing and extraneous markers, while the core of the algorithm is a LSTM Neural Network. The training performed on the network on a first computable sub-set has been a success, revealing a test accuracy of 95%.
Il lavoro si concentra sullo sviluppo di un framework di machine learning basato su reti neurali, in grado di predire le assegnazioni di marcatori 3D da un sistema di motion capture, per un’applicazione esistente dedicata all’analisi biomeccanica in campo medico e sportivo. Partendo da un dataset di file 3D acquisiti in campo e in laboratorio da una eterogeneità di pazienti e atleti, contenenti movimenti specifici, lo scopo è quello di accelerare il processo di analisi attualmente svolto manualmente che prevede un etichettatura manuale dei dati tridimensionali, prima che i dati possano essere analizzati. Il framework di reti neurali implementato si basa su una rigorosa pre-elaborazione e post-elaborazione votata alla pulizia dei dati per ottenere un risultato migliore e per gestire la presenza di marcatori mancanti ed estranei, mentre il cuore dell’algoritmo è una serie di reti neurali LSTM. Il training della rete performato su di un primo sub-set è stato un successo, rilevando un’accuratezza in fase di test del 95%.
Automatic Labelling of 3D Motion Capture Markers using Neural Networks
MONACO, EDOARDO
2021/2022
Abstract
The work focuses on the development of a machine learning framework based on neural networks, which is able to predict the labels of 3D markers from a motion capture system. This should be implemented in an existing application devoted to the biomechanical analysis in the medical and sports field. Starting from a dataset of 3D files acquired in field and in laboratory from an heterogeneity of patients and athletes, containing specific movements, the purpose is to speed up the process of analysis during which somebody until now proceeds with a manual labeling of the 3D cloud of points, before the data can be analyzed. The neural networks framework implemented is based on a strict pre-processing and post-processing voted to clean the data in order to get a better result and to handle the presence of missing and extraneous markers, while the core of the algorithm is a LSTM Neural Network. The training performed on the network on a first computable sub-set has been a success, revealing a test accuracy of 95%.File | Dimensione | Formato | |
---|---|---|---|
Monaco_Edoardo.pdf
accesso aperto
Dimensione
9.09 MB
Formato
Adobe PDF
|
9.09 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/20.500.12608/29056