Over the past several years, the use of wearable devices has increased and, thanks to their reliability, ease of use, affordability and smaller size, those devices are nowadays widely diffused in different fields. Of major interest is their employment in the health and fitness and medical fields. Recent studies focused their attention on the feasibility of using wearable devices for the recognition of daily activities and gestures: the variety of integrated sensors, such as movement sensors (i.e., accelerometers and gyroscopes), allow to acquire a huge amount of data that can be processed and used for gesture tracking. As concerns the medical field, the recognition of daily activities is important for different reasons, such as medical adherence. For example, considering elders and patients affected by chronic diseases, gesture tracking can used to monitor if the patient, once at home, takes medication as prescribed. Another fundamental aspect is the possibility to monitor if a person does physical activity or stays enough hydrated. The aim of this thesis was to assess the feasibility to perform gesture tracking using a ring and a wrist monitor integrating movement sensors. Therefore, a machine learning approach aimed at recognizing nine different gestures was implemented, paying particular attention to the specific gesture of “Drinking” water from a glass. Also, the possibility to quantify the amount of water intake was assessed. Twenty-six subjects were enrolled and provided written informed consent to participate in the study. A wristband (MMR, MbientLab) and a ring (CART, Sky Labs) placed on the index finger were worn simultaneously on the dominant hand during the acquisition. Both the devices integrate movement sensors: in particular, data from tri-axial accelerometer were used (fs=50Hz). This study focuses on the comparison between the two devices using three different classification methods: K-Nearest Neighbors, Support Vector Machines and Naive Bayes. However, in the estimation of the quantity of water intake, the Naïve Bayes model was replaced by a Linear Regression algorithm. The signals acquired were thus elaborated to identify recurrent patterns and then some features, that described the performed activities, were extracted from the training data. In a second step, these features were grouped in an array and given as input to the machine learning models to make the prediction on the test set. Comparing the results obtained with the two different devices, the wrist monitor seems to be the best solution for the activity detection. Indeed, the performances in term of accuracy, precision, recall, F-measure and specificity were higher for the three different machine learning approaches. As regards the detection of the amount of water intake, the results are similar both for the ring and the wrist monitors, but also in this case the second one gave values slightly higher. Moreover, the obtained results indicate that Naïve Bayes were the method that gave the better results in terms of accuracy for the activity detection, but they are not a good approach for the recognition of the “Drinking” moments. On the other hand, Support Vector Machines gave the best performances in recognizing the “Drinking” moments in term of accuracy, precision, recall, F-measure and specificity both for the ring and for the bracelet. Concerning the amount of water intake, the linear regression model gave the lowest committed error in the prediction, expressed as MAPE, with respect to K-Nearest Neighbors and Support Vector Machines.
Negli ultimi anni, l’utilizzo di dispositivi indossabili è cresciuto notevolmente ed è sempre più diffuso in diversi settori, di grande interesse sono le loro applicazioni nel mondo del fitness e in ambito medico. Questo è principalmente dovuto al fatto che sono facilmente accessibili ed utilizzabili, piccoli ed affidabili. Sempre più studi si sono concentrati sull’utilizzo di sistemi indossabili per il riconoscimento di attività tipiche della vita quotidiana: lo sviluppo dei sensori integrati, come quelli di movimento (accelerometri e giroscopi), permettono di raccogliere un grande numero di dati che vengono poi processati per identificare l’attività svolta dall’utente in un determinato istante. In ambito medico, il riconoscimento di attività risulta importante per differenti aspetti, come l’aderenza dei pazienti alle cure. Per esempio, per anziani o persone affette da malattie croniche, il riconoscimento di gesti quotidiani può essere usato per controllare se, in ambiente domestico, per esempio, assumono i farmaci prescritti, si idratano sufficientemente o svolgono regolare attività fisica. L'obiettivo di questa tesi è stato quello di valutare la possibilità di riconoscere gesti quotidiani tramite sensori di movimento integrati in un anello e un dispositivo da polso. Sono stati utilizzati algoritmi di machine learning per classificare nove differenti attività. Inoltre, è stata valutata la possibilità di determinare la quantità di acqua bevuta in un’acquisizione di due minuti. Ventisei soggetti hanno partecipato allo studio firmando un consenso informato. Un dispositivo da polso (MMR, MbientLab) e un anello (CART, Sky Labs), indossati sulla mano dominante, sono stati utilizzati contemporaneamente durante le acquisizioni. Entrambi contengono sensori di movimento, in particolare è stato utilizzato l’accelerometro tri-assiale per acquisire i dati di interesse (fs=50Hz). Questo studio si basa sul confronto tra i due sistemi indossabili utilizzando tre approcci di classificazione: K-Nearest Neighbors, Support Vector Machines and Naive Bayes. I dati ottenuti sono quindi stati elaborati, al fine di individuare sequenze di pattern ricorrenti, e sono stati estratti features, di tipo statistico, che descrivono le attività eseguite dai soggetti, dai dati di training. Successivamente, i features così ottenuti sono stati raggruppati in vettori e dati in input ai tre modelli di machine learning per la classificazione sui dati di test. Confrontando le performances ottenute per i segnali acquisiti con i due dispositivi, il dispositivo da polso sembra la scelta migliore per il riconoscimento di attività. Infatti, i valori di accuratezza, precisione, recall, F-measure e specificità raggiunti sono migliori per tutti e tre i modelli di machine learning utilizzati. Per quanto riguarda la predizione della quantità di acqua bevuta in una sessione di acquisizione i risultati ottenuti sono simili per entrambi i dispositivi, ma anche in questo caso con quello da polso appaiono leggermente più alti. I risultati ottenuti indicano inoltre che per la classificazione delle attività svolte, i Naive Bayes risultano il metodo che fornisce i valori più alti in termini di accuratezza. Le Support Vector Machines risultano invece il classificatore più efficace per riconoscere il gesto del bere, fornendo valori di accuratezza, precisione, recall, F-measure e specificità migliori rispetto agli altri modelli sia per le acquisizioni con anello che con il dispositivo da polso. Per la stima di quantità di acqua bevuta, invece, la regressione lineare è il metodo che ha prodotto risultati migliori in termini di errore commesso nella predizione, espresso come MAPE.
Classification of gestures by wrist and ring accelerometers using machine learning methods
BOSSI, FEDERICA FRANCESCA STEFANIA
2018/2019
Abstract
Over the past several years, the use of wearable devices has increased and, thanks to their reliability, ease of use, affordability and smaller size, those devices are nowadays widely diffused in different fields. Of major interest is their employment in the health and fitness and medical fields. Recent studies focused their attention on the feasibility of using wearable devices for the recognition of daily activities and gestures: the variety of integrated sensors, such as movement sensors (i.e., accelerometers and gyroscopes), allow to acquire a huge amount of data that can be processed and used for gesture tracking. As concerns the medical field, the recognition of daily activities is important for different reasons, such as medical adherence. For example, considering elders and patients affected by chronic diseases, gesture tracking can used to monitor if the patient, once at home, takes medication as prescribed. Another fundamental aspect is the possibility to monitor if a person does physical activity or stays enough hydrated. The aim of this thesis was to assess the feasibility to perform gesture tracking using a ring and a wrist monitor integrating movement sensors. Therefore, a machine learning approach aimed at recognizing nine different gestures was implemented, paying particular attention to the specific gesture of “Drinking” water from a glass. Also, the possibility to quantify the amount of water intake was assessed. Twenty-six subjects were enrolled and provided written informed consent to participate in the study. A wristband (MMR, MbientLab) and a ring (CART, Sky Labs) placed on the index finger were worn simultaneously on the dominant hand during the acquisition. Both the devices integrate movement sensors: in particular, data from tri-axial accelerometer were used (fs=50Hz). This study focuses on the comparison between the two devices using three different classification methods: K-Nearest Neighbors, Support Vector Machines and Naive Bayes. However, in the estimation of the quantity of water intake, the Naïve Bayes model was replaced by a Linear Regression algorithm. The signals acquired were thus elaborated to identify recurrent patterns and then some features, that described the performed activities, were extracted from the training data. In a second step, these features were grouped in an array and given as input to the machine learning models to make the prediction on the test set. Comparing the results obtained with the two different devices, the wrist monitor seems to be the best solution for the activity detection. Indeed, the performances in term of accuracy, precision, recall, F-measure and specificity were higher for the three different machine learning approaches. As regards the detection of the amount of water intake, the results are similar both for the ring and the wrist monitors, but also in this case the second one gave values slightly higher. Moreover, the obtained results indicate that Naïve Bayes were the method that gave the better results in terms of accuracy for the activity detection, but they are not a good approach for the recognition of the “Drinking” moments. On the other hand, Support Vector Machines gave the best performances in recognizing the “Drinking” moments in term of accuracy, precision, recall, F-measure and specificity both for the ring and for the bracelet. Concerning the amount of water intake, the linear regression model gave the lowest committed error in the prediction, expressed as MAPE, with respect to K-Nearest Neighbors and Support Vector Machines.| File | Dimensione | Formato | |
|---|---|---|---|
|
2019_12_Bossi.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Master Thesis
Dimensione
14.86 MB
Formato
Adobe PDF
|
14.86 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/151064