A Comparison Study of Classifier Algorithms for Cross-Person Physical Activity Recognition Articles uri icon

publication date

  • January 2017

issue

  • 1

volume

  • 17

International Standard Serial Number (ISSN)

  • 1424-3210

Electronic International Standard Serial Number (EISSN)

  • 1424-8220

abstract

  • Physical activity is widely known to be one of the key elements of a healthy life. The many benefits of physical activity described in the medical literature include weight loss and reductions in the risk factors for chronic diseases. With the recent advances in wearable devices, such as smartwatches or physical activity wristbands, motion tracking sensors are becoming pervasive, which has led to an impressive growth in the amount of physical activity data available and an increasing interest in recognizing which specific activity a user is performing. Moreover, big data and machine learning are now cross-fertilizing each other in an approach called "deep learning", which consists of massive artificial neural networks able to detect complicated patterns from enormous amounts of input data to learn classification models. This work compares various state-of-the-art classification techniques for automatic cross-person activity recognition under different scenarios that vary widely in how much information is available for analysis. We have incorporated deep learning by using Google's TensorFlow framework. The data used in this study were acquired from PAMAP2 (Physical Activity Monitoring in the Ageing Population), a publicly available dataset containing physical activity data. To perform cross-person prediction, we used the leave-one-subject-out (LOSO) cross-validation technique. When working with large training sets, the best classifiers obtain very high average accuracies (e.g., 96% using extra randomized trees). However, when the data volume is drastically reduced (where available data are only 0.001% of the continuous data), deep neural networks performed the best, achieving 60% in overall prediction accuracy. We found that even when working with only approximately 22.67% of the full dataset, we can statistically obtain the same results as when working with the full dataset.

subjects

  • Computer Science

keywords

  • physical activity recognition; classification; machine learning; deep learning; biomedical signal processing; time series analysis;wearable sensors; triaxial accelerometer; feature-selection; mobile phones; networks; health; smartphones; multiclass; children; features