Multi-modal human gesture recognition combining dynamic programming and probabilistic methods

Abstract

In this M. Sc. Thesis, we deal with the problem of Human Gesture Recognition using Human Behavior Analysis technologies. In particular, we apply the proposed methodologies in both health care and social applications. In these contexts, gestures are usually performed in a natural way, producing a high variability between the Human Poses that belong to them. This fact makes Human Gesture Recognition a very challenging task, as well as their generalization on developing technologies for Human Behavior Analysis. In order to tackle with the complete framework for Human Gesture Recognition, we split the process in three main goals: Computing multi-modal feature spaces, probabilistic modelling of gestures, and clustering of Human Poses for Sub-Gesture representation. Each of these goals implicitly includes different challenging problems, which are interconnected and faced by three presented approaches: Bag-of-Visual-and-Depth-Words, Probabilistic-Based Dynamic Time Warping, and Sub-Gesture Representation. The methodologies of each of these approaches are explained in detail in the next sections. We have validated the presented approaches on different public and designed data sets, showing high performance and the viability of using our methods for real Human Behavior Analysis systems and applications. Finally, we show a summary of different related applications currently in development, as well as both conclusions and future trends of research.