Graduation Year

2012

Document Type

Dissertation

Degree

Ph.D.

Degree Granting Department

Computer Science and Engineering

Major Professor

Miguel Labrador, Ph.D.

Committee Member

Wilfrido Moreno, Ph.D.

Committee Member

Rafael Perez, Ph.D.

Committee Member

Xiaoning Qian, Ph.D.

Committee Member

Kandethody Ramachandran, Ph.D.

Keywords

context-aware applications, human-centric sensing, machine learning, pervasive computing, ubiquitous sensing

Abstract

Delivering accurate and opportune information on people's activities and behaviors has become one of the most important tasks within pervasive computing. Its wide spectrum of potential applications in medical, entertainment, and tactical scenarios, motivates further

research and development of new strategies to improve accuracy, pervasiveness, and eciency.

This dissertation addresses the recognition of human activities (HAR) with wearable sensors in three main regards: In the rst place, physiological signals have been incorporated as a new source of information to improve the recognition accuracy achieved by conventional approaches, which rely on accelerometer signals solely. A new HAR system, Centinela, was born from such concept, employing structural feature extraction along with classier

ensembles, and achieving over 95% of recognition accuracy.

In the second place, real time activity recognition was enabled by Vigilante, a mobile HAR framework under the AndroidTM platform. Providing immediate feedback on the user's activities is especially benecial in healthcare and military applications, which

may require alert triggering or support of decision making. The evaluation demonstrates that Vigilante is energy ecient while maintaining high accuracy (i.e., up to 96.8%) and

low response time. The system features MECLA, a mobile library for the evaluation of classification algorithms, which is also suitable for further machine learning applications.

Finally, the activity recognition accuracy is improved by two new strategies for decision fusion and selection in multiple classier systems: the failure product and the precision-recall dierence. The experimental analysis conrms that the presented methods are benecial, not only for recognizing human activities, but also for many other classication problems.

Share

COinS