Research on hand gesture recognition

Abstract
The thesis presents methods for modeling deformable objects (such as a hand) that perform a gesture in the image and their effective detection and tracking using a-priori knowledge about the structure of gestures performed. The ability to determine changes in position and shape of the object in time enables also to use these methods for interpretation of its motion, including the recognition of dynamic gestures. In the work the stochastic model encoding gesture structure is proposed that can be used to perform simultaneous gesture tracking and recognition. The Dynamic Bayesian Network is applied for formal representation of the model. The Deformable Templates method is utilized to encode the hand shape. Different algorithms based on Switched Kalman and Particle Filtering are evaluated for performing network inference. It is shown that the information provided by the inference algorithms is sufficient to perform gesture recognition and temporal segmentation of gestures in continuous gesture sequence. In the thesis there is given a method for modeling the space of hand shapes as well as learning the conditional probability distributions of the Dynamic Bayesian Network basing on the training set of gestures. There are also provided experimental results for gesture tracking and recognition using the proposed system.