Sampling frequency optimization and training model selection for physical activity classification with single triaxial accelerometer
Jiang, Chao (2015)
Jiang, Chao
2015
Master's Degree Programme in Science and Bioengineering
Luonnontieteiden tiedekunta - Faculty of Natural Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2015-10-07
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tty-201509221600
https://urn.fi/URN:NBN:fi:tty-201509221600
Tiivistelmä
Ambulatory monitoring system with accelerometers can provide a reliable, continuous, unsupervised and objective monitoring of human physical activities. The system can in many cases recognize the type of activity being performed, and calculate the duration and intensity. This kind of information can be utilized to help people to follow up their physical activities and remind people to be more active, because physical inactivity can cause some health problems. However, especially for mobile devices continuous sam-pling, signal processing and activity recognition rapidly depletes the system’s energy, which is a critically constrained resource.
In this thesis work, several methods for reducing energy consumption in physical activi-ty recognition were reviewed and discussed, i.e., 1) reducing the number of sensors used; 2) selecting low power sensors; 3) reducing the number of axes; 4) decreasing the sampling frequency; 5) adopting an adaptive sampling strategy. In this thesis, a single tri-axial accelerometer was utilized for sensing the accelerations, and sampling frequency was optimized in order to lower the energy consumption. The physical activity recognition was performed with different sampling frequencies and training strategies, with the target to reach good classification accuracies and low energy consumption.
Based on the obtained classification results, several conclusions were drawn. Firstly, personal models did not always achieve better classification accuracies over impersonal and hybrid models. However, personal models performed much better for some activi-ties, e.g., biking, lying, and rowing. Secondly, there was no uniform optimal sampling frequency for all activities. Sampling frequencies no larger than 10 Hz were enough to classify all activities.
To further optimize the energy consumption, adaptive sampling rate logic was designed and implemented. It adaptively used 1 Hz when sampling the accelerations from lying activity and 10 Hz for other activities. The results showed it worked effectively and efficiently.
In this thesis work, several methods for reducing energy consumption in physical activi-ty recognition were reviewed and discussed, i.e., 1) reducing the number of sensors used; 2) selecting low power sensors; 3) reducing the number of axes; 4) decreasing the sampling frequency; 5) adopting an adaptive sampling strategy. In this thesis, a single tri-axial accelerometer was utilized for sensing the accelerations, and sampling frequency was optimized in order to lower the energy consumption. The physical activity recognition was performed with different sampling frequencies and training strategies, with the target to reach good classification accuracies and low energy consumption.
Based on the obtained classification results, several conclusions were drawn. Firstly, personal models did not always achieve better classification accuracies over impersonal and hybrid models. However, personal models performed much better for some activi-ties, e.g., biking, lying, and rowing. Secondly, there was no uniform optimal sampling frequency for all activities. Sampling frequencies no larger than 10 Hz were enough to classify all activities.
To further optimize the energy consumption, adaptive sampling rate logic was designed and implemented. It adaptively used 1 Hz when sampling the accelerations from lying activity and 10 Hz for other activities. The results showed it worked effectively and efficiently.