Abstract
In this research, we present a novel method to classify hand shape given a sequence of images capturing a hand manipulating an object obtained by a RGB-D camera. A finite set of typical hand shapes and object shapes are defined and the problem is formulated as to jointly estimate a pair of hand shape and object shape classes in each frame. Hu's invariant moment is calculated for hand region and object region separately and they are concatenated to become a single feature vector. Support Vector Machine is used to estimate the class of hand shape and object shape in each frame locally. Then, dynamic programming is applied to estimate the globally optimal object shape as well as the transition of hand shapes. Experimental results show that the proposed method outperformed the conventional methods.
