Machine Learning Based Physical Activity Extraction for Unannotated Acceleration Data

Abstract

Sensor based human activity recognition (HAR) is an emerging and challenging research area. The physical activity of people has been associated with many health benefits and even reducing the risk of different diseases. It is possible to collect sensor data related to physical activities of people with wearable devices and embedded sensors, for example in smartphones and smart environments. HAR has been successful in recognizing physical activities with machine learning methods. However, it is a critical challenge to annotate sensor data in HAR. Most existing approaches use supervised machine learning methods which means that true labels need be given to the data when training a machine learning model. Supervised deep learning methods have outperformed traditional machine learning methods in HAR but they require an even more extensive amount of data and true labels.

In this thesis, machine learning methods are used to develop a solution that can recognize physical activity (e.g., walking and sedentary time) from unannotated acceleration data collected using a wearable accelerometer device. It is shown to be beneficial to collect and annotate data from physical activity of only one person. Supervised classifiers can be trained with small, labeled acceleration data and more training data can be obtained in a semi-supervised setting by leveraging knowledge from available unannotated data. The semi-supervised En-Co-Training method is used with the traditional supervised machine learning methods K-nearest Neighbor and Random Forest. Also, intensities of activities are produced by the cut point analysis of the OMGUI software as reference information and used to increase confidence of correctly selecting pseudo-labels that are added to the training data. A new metric is suggested to help to evaluate reliability when no true labels are available. It calculates a fraction of predictions that have a correct intensity out of all the predictions according to the cut point analysis of the OMGUI software.

The reliability of the supervised KNN and RF classifiers reaches 88 % accuracy and the C-index value 0,93, while the accuracy of the K-means clustering remains 72 % when testing the models on labeled acceleration data. The initial supervised classifiers and the classifiers retrained in a semi-supervised setting are tested on unlabeled data collected from 12 people and measured with the new metric. The overall results improve from 96-98 % to 98-99 %. The results with more challenging activities to the initial classifiers, taking a walk improve from 55-81 % to 67-81 % and jogging from 0-95 % to 95-98 %. It is shown that the results of the KNN and RF classifiers consistently increase in the semi-supervised setting when tested on unannotated, real-life data of 12 people.