Available at: http://digitalcommons.calpoly.edu/theses/1665
Date of Award
MS in Computer Science
The use of supervised machine learning to extend the capabilities and overall viability of motion sensing input devices has been an increasingly popular avenue of research since the release of the Leap Motion in 2013. The device's optical sensors are capable of recognizing and tracking key features of a user's hands and fingers, which can be obtained and manipulated through a robust API. This makes statistical classification ideal for tackling the otherwise laborious and error prone nature of adding new programmer-defined gestures to the set of recognized gestures.
Although a handful of studies have explored the effectiveness of machine learning with the Leap Motion, none to our knowledge have run a comparative performance analysis of classification algorithms or made use of more than several of them in their experiments. The aim of this study is to improve the reliability of detecting newly added gestures by identifying the classifiers that produce the best results. To this end, a formal analysis of the most popular classifiers used in the field of machine learning was performed to determine those most appropriate to the requirements of the Leap Motion. A recording and evaluation system was developed to collect recordings of gestures that could then be used to train a classification prediction model, as well as calculate the training run time and performance statistics for each of the classifiers tested. It is from these measurements made under the framework of this study that a recommendation of classifiers can be made.