Abstract
In this thesis, we proposed a hand gesture recognition algorithm based on hierarchical hand parsing from a single depth image. In the proposed system, we first normalize in-plane rotation of the hand pose. According to hand configuration, we propose to segment a hand into 11 non-overlapping parts with a novel 3-layer hierarchical Random Decision Forest (RDF) per-pixel classifier. In the first layer, the hand region is divided into two parts: palm and fingers. In the second layer, pixels are classified into different finger classes: thumb, index finger, middle finger, ring finger and pinky finger. In the third layer, a finger pixel is classified into upper and lower part. In each layer, per-pixel classification is executed to assign a set of posterior probabilities corresponding to different hand parts to each pixel based on depth-context features. To develop hand gesture recognition, the information of parsed hand is employed to compute three kinds of features including posture feature, finger angle feature and hand part ratio feature, for Support Vector Machines (SVMs). Our experiments show superior performance of hand parsing and gesture recognition by using the proposed algorithm compared to some previous methods on different real hand pose datasets.