Development of a Human Activity Recognition System Using Inertial Measurement Unit Sensors on a Smartphone

Abstract

Monitoring an individual’s mobility with a modern smartphone can have a profound impact on rehabilitation in the community. The thesis objective was to develop and evaluate a third-generation Wearable Mobility Monitoring System (WMMS) that uses features from inertial measurement units to categorize activities and determine user changes-of-state in daily living environments. A custom suite of MATLAB® software tools were developed to assess the previous WMMS iteration and aid in third-generation WMMS algorithm construction and evaluation. A rotation matrix was developed to orient smartphone accelerometer components to any three-dimensional reference, to improve accelerometer-based activity identification. A quaternion-based rotation matrix was constructed from an axis-angle pair, produced via algebraic manipulations of acceleration components in the device’s body-fixed reference frame. The third-generation WMMS (WMMS3) evaluation was performed on fifteen able-bodied participants. A BlackBerry Z10 smartphone was placed at a participant’s pelvis, and the device was corrected in orientation. Acceleration due to gravity and applied linear acceleration signals on a BlackBerry Z10 were then used to calculate features that classify activity states through a decision tree classifier. The software tools were then used for offline data manipulation, feature generation, and activity state prediction. Three prediction sets were conducted. The first set considered a “phone orientation independent” mobility assessment of a person’s mobile state. The second set differentiated immobility as sit, stand, or lie. The third prediction set added walking, climbing stairs, and small standing movements classification. Sensitivities, specificities and -Scores for activity categorization and changes-of-state were calculated. The mobile versus immobile prediction set had a sensitivity of 93% and specificity of 97%, while the second prediction set had a sensitivity of 86% and specificity of 97%. For the third prediction set, the sensitivity and specificity decreased to 84% and 95% respectively, which still represented an increase from 56% and 88% found in the previous WMMS. The third-generation WMMS algorithm was shown to perform better than the previous version in both categorization and change-of-state determination, and can be used for rehabilitation purposes where mobility monitoring is required.