IoT Prism Lab
Research Lab at University of Bologna
Research Lab at University of Bologna
Theses
2018 |
Zuhouri, Ramy Al Human Activity Recognition in Sports Using the Apple Watch Masters Thesis 2018. Abstract | Links | BibTeX | Tags: Accelerometers, Apple Watch, Fitness Tracking, Human Activity Recognition, Machine Learning, Smartwatch @mastersthesis{amslaurea16106, title = {Human Activity Recognition in Sports Using the Apple Watch}, author = {Ramy Al Zuhouri}, url = {https://amslaurea.unibo.it/id/eprint/16106}, year = {2018}, date = {2018-01-01}, abstract = {With the recent spreading of Internet of Things, the availability of a wide variety of cheap devices brought human activity recognition (HAR) to the broad audience, thus eliminating the need of using costly and obstructive hardware, and to constraint the users to remain in controlled environments. Human activity recognition finds application in the field of health monitoring, safety, context-aware behavior and fitness tracking. This thesis is focused on fitness tracking, with the aim of finding a way to perform activity recognition with a smartwatch in order to automatize activity tracking, exempting the user from manually interacting with the device in order to manage the workout sessions. For this purpose, the intertial sensors of an Apple Watch, along with the GPS and the heart rate monitor were used to train and test an activity recognition model. 4 subjects collected data for 8 different types of aerobic activities, populating a dataset of 4,083 instances, corresponding to about 20 minutes of physical activity for each subject. 9 different machine learning algorithms were evaluated using the holdout validation, trying different combinations of sensor data and features in order to find the optimal configuration. Due to its simplicity, it was chosen to use a decision tree for further validation on unseen data. As expected, the accuracy of the decision tree was higher when validated on the test set, but dropped from 95.42% to 90.73% when tested on unseen data. The use of a history set increased the recognition accuracy up to 92.68%. More conclusions derived from validation: the models were able to recognize activities independently on the wirst location of the watch; moreover, the accelerometer and the gyroscope were enough to obtain a good recognition model, while the GPS and the heart rate monitor did not significantly increase the accuracy.}, keywords = {Accelerometers, Apple Watch, Fitness Tracking, Human Activity Recognition, Machine Learning, Smartwatch}, pubstate = {published}, tppubtype = {mastersthesis} } With the recent spreading of Internet of Things, the availability of a wide variety of cheap devices brought human activity recognition (HAR) to the broad audience, thus eliminating the need of using costly and obstructive hardware, and to constraint the users to remain in controlled environments. Human activity recognition finds application in the field of health monitoring, safety, context-aware behavior and fitness tracking. This thesis is focused on fitness tracking, with the aim of finding a way to perform activity recognition with a smartwatch in order to automatize activity tracking, exempting the user from manually interacting with the device in order to manage the workout sessions. For this purpose, the intertial sensors of an Apple Watch, along with the GPS and the heart rate monitor were used to train and test an activity recognition model. 4 subjects collected data for 8 different types of aerobic activities, populating a dataset of 4,083 instances, corresponding to about 20 minutes of physical activity for each subject. 9 different machine learning algorithms were evaluated using the holdout validation, trying different combinations of sensor data and features in order to find the optimal configuration. Due to its simplicity, it was chosen to use a decision tree for further validation on unseen data. As expected, the accuracy of the decision tree was higher when validated on the test set, but dropped from 95.42% to 90.73% when tested on unseen data. The use of a history set increased the recognition accuracy up to 92.68%. More conclusions derived from validation: the models were able to recognize activities independently on the wirst location of the watch; moreover, the accelerometer and the gyroscope were enough to obtain a good recognition model, while the GPS and the heart rate monitor did not significantly increase the accuracy. |