Schneider, O., "Using Gait as an Input Modality for Mobile Exercise Games,", M.Sc. Thesis, University of British Columbia, 2012.
To encourage and support physical activity in increasingly sedentary lifestyles, many are turning to mobile technology. Modern smartphones are equipped with a wealth of sensors, including Global Positioning Systems (GPS) and accelerometers, suggesting great potential to be integrated with fitness and exercise applications. So far, GPS-enabled devices have been used to support running, cycling, or even exercise games that encourage people to be physically active, but GPS- enabled devices lack fine-grained information about the user’s activity. Accelerometers have been used to some effect to detect step count and walking cadence (step rate), and even to classify activity (distinguishing walking from cycling, for example), but require a known carrying location and orientation. In this work, we examine the role of location in two application areas - real-time cadence estimation and gait classification - and develop algorithms to accommodate diverse carrying locations. In the first application area, real-time cadence estimation, our algorithm (Robust Real-time Algorithm for Cadence Estimation, or RRACE) uses a frequency- domain analysis to perform well without training or tuning, and is robust to changes in carrying locations. We demonstrate RRACE’s performance and robustness to be an improvement over existing algorithms with data collected from a user study. In the second application area, gait classification, we present a novel set of 15 gaits suitable for exercise games and other fitness applications. Using a minimal amount of training for participants, we can achieve a mean of 78.1% classification for all 15 gaits and all locations, an accuracy which may be usable now in some applications and warrants further investigation of this approach. We present findings of how our classification scheme confuses these gaits, and encapsulate insights in guidelines for designers. We also demonstrate that our classification performance varies dramatically for each individual even when trained and tested on that individual, suggesting strong individual differences in performing gaits. Our innovative methodology for simple and quick collection of accelerometer data is also described in detail. Future work includes planned improvements to both algorithms, further investigation of individual differences, and extension of this work to other application areas.