Velocity based training is a workout methodology that uses the speed of the repetition as evaluation parameter for the performance. Also, devices’ capability of giving real-time feedback allows the personal trainers, coaches or physiotherapists to give immediate and concrete advices to the athlete. Even if the concept of Velocity based training started to be introduced some decades ago, nowadays there is an increasing use thanks to the development of the technology and wearable devices. In this demanding environment, Beast Technologies has developed a fitness tracker connected via Bluetooth to a mobile application which is able to measure the speed of each repetition and count them. Further detailed data are gathered and sent to the web portal for analysis and consulting. The limitation of this solution is the counting of “ghost” repetition: movements that are not effective repetitions but are counted as such (e.g. unracking barbell to start the exercise). My work deals with the detection of that ghost repetitions to both increase the accuracy of the data and to improve the user experience by reducing the interactions with the smartphone. We have chosen to use a machine learning algorithm, namely a Random Forest classifier, to detect ghost repetitions among the correct ones. We have collected about 10,000 labelled repetitions, which has been the more time-consuming and error-prone activity, divided in more than 1000 sets and coming from 66 exercises involving different body muscles, gym equipment and loads from both male and female athletes of any level of preparation. We have used this dataset with more than 40 features to train the Random Forest that gives a cross-validation accuracy of 93,4%. However, it reaches 98% accuracy on simpler exercises such as squat or biceps curls with many-reps sets while it goes down to 79% in more complex exercises such as hang power clean (Olympic discipline) with few-reps sets. In 6-days testing workout we achieved an average 22% and 32,5% accuracy improvement in repetitions counting and correctness of sets, respectively.
Velocity based training è una metodologia di allenamento che usa la velocità della ripetizione come parametro di valutazione per la performance. Inoltre, la capacità dei device di dare feedback in tempo reale permette ai personal trainers, allenatori o fisioterapisti di dare consigli immediati e concreti all’atleta. Anche se il concetto di “velocity based” ha iniziato ad essere introdotto qualche decade fa, oggi c’è un uso crescente grazie allo sviluppo delle tecnologie e dei wearable devices. In questo contesto, Beast Techologies ha sviluppato un fitness tracker connesso via Bluetooth a un’app per smartphone capace di misurare la velocità di ogni ripetizione e di contarle. Ulteriori dati sono raccolti ed inviati al portale web per l’analisi e la consulenza. Il limite di questa soluzione è il conteggio di ripetizioni “fantasma”: movimenti che non sono ripetizioni effettive ma vengono conteggiate come tali (es. spostare il bilanciere per iniziare un esercizio). Il mio lavoro si occupa della individuazione delle ripetizioni fantasma sia per incrementare l’accuratezza dei dati sia per migliorare l’esperienza utente riducendo le interazioni con lo smartphone. Abbiamo scelto di usare un algoritmo di machine learning, cioè un classificatore con una Random Forest, per individuare le ripetizioni fantasma tra quelle corrette. Abbiamo catalogato circa 10,000 ripetizioni, che è stata l’attività più lunga e prona ad errori, divise in più di 1000 sets e derivanti da 66 esercizi che includono differenti muscoli, attrezzi e carichi sia da atleti maschili che femminili di tutti i livelli di preparazione. Abbiamo usato questo dataset con più di 40 features per allenare la Random Forest che risulta in un’accuratezza in cross-validation di 93,4%. L’algoritmo raggiunge un’accuratezza del 98% per esercizi più semplici come squat o curl per bicipiti in set con molte ripetizioni mentre scende fino al 79% per esercizi più complessi come le girate (disciplina olimpica) in set con poche ripetizioni. In 6 giorni di allenamenti di test abbiamo ottenuto un miglioramento di accuratezza rispettivamente del 22% e del 32,5% per il conteggio delle ripetizioni e per la correttezza dei set.
A machine learning algorithm to recognize movement with IMU sensors during strength workout
GATTI, GIAN GIACOMO
2017/2018
Abstract
Velocity based training is a workout methodology that uses the speed of the repetition as evaluation parameter for the performance. Also, devices’ capability of giving real-time feedback allows the personal trainers, coaches or physiotherapists to give immediate and concrete advices to the athlete. Even if the concept of Velocity based training started to be introduced some decades ago, nowadays there is an increasing use thanks to the development of the technology and wearable devices. In this demanding environment, Beast Technologies has developed a fitness tracker connected via Bluetooth to a mobile application which is able to measure the speed of each repetition and count them. Further detailed data are gathered and sent to the web portal for analysis and consulting. The limitation of this solution is the counting of “ghost” repetition: movements that are not effective repetitions but are counted as such (e.g. unracking barbell to start the exercise). My work deals with the detection of that ghost repetitions to both increase the accuracy of the data and to improve the user experience by reducing the interactions with the smartphone. We have chosen to use a machine learning algorithm, namely a Random Forest classifier, to detect ghost repetitions among the correct ones. We have collected about 10,000 labelled repetitions, which has been the more time-consuming and error-prone activity, divided in more than 1000 sets and coming from 66 exercises involving different body muscles, gym equipment and loads from both male and female athletes of any level of preparation. We have used this dataset with more than 40 features to train the Random Forest that gives a cross-validation accuracy of 93,4%. However, it reaches 98% accuracy on simpler exercises such as squat or biceps curls with many-reps sets while it goes down to 79% in more complex exercises such as hang power clean (Olympic discipline) with few-reps sets. In 6-days testing workout we achieved an average 22% and 32,5% accuracy improvement in repetitions counting and correctness of sets, respectively.| File | Dimensione | Formato | |
|---|---|---|---|
|
A machine learning algorithm to recognize movement with IMU sensors during strenght workout.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
4.17 MB
Formato
Adobe PDF
|
4.17 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/10589/141811