In this work, a framework based on Deep Learning (DL) for Human Activity Recognition (HAR) is designed, developed, and validated, together with the presentation of an experimental protocol and an acquisition system. Eighteen subjects from IRCCS Maugeri of Bari (Italy) were involved in the study to train a model able to recognize activities of daily living (ADLs) considering individual physical characteristics by means of a kinematic model. The acquisition system involves the use of four IMU sensors, placed on the right and left pelvis, right wrist, and sternum. The implemented Convolutional Neural Network (CNN) model was trained on separated repetitions and tested on the repetitions performed continuously, for each possible combination of sensors. The data will be collected, in this stage of progress of the project, on a sample of ten healthy people, trying to maximize the variability in terms of age, sex, build, and height. After a pre-processing process of the kinematic data, these will be input to a convolutional neural network which guarantees the recognition of the type of movement performed. The implementation of a HAR system of this type represents an advantage for the assisted patient, as they are subjected to continuous monitoring.
In questo lavoro viene progettato, sviluppato e validato un framework basato sul Deep Learning (DL) per la Human Activity Recognition (HAR), insieme alla presentazione di un protocollo sperimentale e di un sistema di acquisizione. Diciotto soggetti dell'IRCCS Maugeri di Bari (Italia) sono stati coinvolti nello studio per addestrare un modello in grado di riconoscere le attività della vita quotidiana (ADL) considerando le caratteristiche fisiche individuali attraverso un modello cinematico. Il sistema di acquisizione prevede l'utilizzo di quattro sensori IMU, posizionati sul bacino destro e sinistro, sul polso destro e sullo sterno. Il modello di rete neurale convoluzionale (CNN) implementato è stato addestrato su ripetizioni separate e testato sulle ripetizioni eseguite in continuo, per ogni possibile combinazione di sensori. I dati saranno raccolti, in questa fase di avanzamento del progetto, su un campione di dieci persone sane, cercando di massimizzare la variabilità in termini di età, sesso, corporatura e altezza. Dopo un processo di pre-elaborazione dei dati cinematici, questi verranno inseriti in una rete neurale convoluzionale che garantisce il riconoscimento del tipo di movimento eseguito. L'implementazione di un sistema HAR di questo tipo rappresenta un vantaggio per il paziente assistito, che viene sottoposto a un monitoraggio continuo.
MODELLO DI DEEP LEARNING PER HUMAN ACTIVITY RECOGNITION
PALAZZO, LUCIA
2022/2023
Abstract
In this work, a framework based on Deep Learning (DL) for Human Activity Recognition (HAR) is designed, developed, and validated, together with the presentation of an experimental protocol and an acquisition system. Eighteen subjects from IRCCS Maugeri of Bari (Italy) were involved in the study to train a model able to recognize activities of daily living (ADLs) considering individual physical characteristics by means of a kinematic model. The acquisition system involves the use of four IMU sensors, placed on the right and left pelvis, right wrist, and sternum. The implemented Convolutional Neural Network (CNN) model was trained on separated repetitions and tested on the repetitions performed continuously, for each possible combination of sensors. The data will be collected, in this stage of progress of the project, on a sample of ten healthy people, trying to maximize the variability in terms of age, sex, build, and height. After a pre-processing process of the kinematic data, these will be input to a convolutional neural network which guarantees the recognition of the type of movement performed. The implementation of a HAR system of this type represents an advantage for the assisted patient, as they are subjected to continuous monitoring.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_PalazzoLucia.pdf
File consultabile da postazioni biblioteca
Descrizione: Tesi Lucia Palazzo
Dimensione
4.85 MB
Formato
Adobe PDF
|
4.85 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14253/69278