Real-time gait classification for persuasive smartphone apps: structuring the literature and pushing the limits

  • Authors:

  • Oliver S. Schneider;Karon E. MacLean;Kerem Altun;Idin Karuei;Michael M.A. Wu

  • Affiliations:

  • University of British Columbia, Vancouver, British Columbia, Canada;University of British Columbia, Vancouver, British Columbia, Canada;Istanbul Kemerburgaz University, Istanbul, Turkey;University of British Columbia, Vancouver, British Columbia, Canada;University of British Columbia, Vancouver, British Columbia, Canada

  • Venue:
  • Proceedings of the 2013 international conference on Intelligent user interfaces
  • Year:
  • 2013

Quantified Score

Hi-index 0.00

Visualization

Abstract

Persuasive technology is now mobile and context-aware. Intelligent analysis of accelerometer signals in smartphones and other specialized devices has recently been used to classify activity (e.g., distinguishing walking from cycling) to encourage physical activity, sustainable transport, and other social goals. Unfortunately, results vary drastically due to differences in methodology and problem domain. The present report begins by structuring a survey of current work within a new framework, which highlights comparable characteristics between studies; this provided a tool by which we and others can understand the current state-of-the art and guide research towards existing gaps. We then present a new user study, positioned in an identified gap, that pushes limits of current success with a challenging problem: the real-time classification of 15 similar and novel gaits suitable for several persuasive application areas, focused on the growing phenomenon of exercise games. We achieve a mean correct classification rate of 78.1% of all 15 gaits with a minimal amount of personalized training of the classifier for each participant when carried in any of 6 different carrying locations (not known a priori). When narrowed to a subset of four gaits and one location that is known, this improves to means of 92.2% with and 87.2% without personalization. Finally, we group our findings into design guidelines and quantify variation in accuracy when an algorithm is trained for a known location and participant.