3D hand tracking.

Abstract

The hand is often considered as one of the most natural and intuitive interaction modalities for human-to-human interaction. In human-computer interaction (HCI), proper 3D hand tracking is the first step in developing a more intuitive HCI system which can be used in applications such as gesture recognition, virtual object manipulation and gaming. However, accurate 3D hand tracking, remains a challenging problem due to the hand’s deformation, appearance similarity, high inter-finger occlusion and complex articulated motion. Further, 3D hand tracking is also interesting from a theoretical point of view as it deals with three major areas of computer vision- segmentation (of hand), detection (of hand parts), and tracking (of hand). This thesis proposes a region-based skin color detection technique, a model-based and an appearance-based 3D hand tracking techniques to bring the human-computer interaction applications one step closer. All techniques are briefly described below. Skin color provides a powerful cue for complex computer vision applications. Although skin color detection has been an active research area for decades, the mainstream technology is based on individual pixels. This thesis presents a new region-based technique for skin color detection which outperforms the current state-of-the-art pixel-based skin color detection technique on the popular Compaq dataset (Jones & Rehg 2002). The proposed technique achieves 91.17% true positive rate with 13.12% false negative rate on the Compaq dataset tested over approximately 14,000 web images. Hand tracking is not a trivial task as it requires tracking of 27 degreesof- freedom of hand. Hand deformation, self occlusion, appearance similarity and irregular motion are major problems that make 3D hand tracking a very challenging task. This thesis proposes a model-based 3D hand tracking technique, which is improved by using proposed depth-foreground-background ii feature, palm deformation module and context cue. However, the major problem of model-based techniques is, they are computationally expensive. This can be overcome by discriminative techniques as described below. Discriminative techniques (for example random forest) are good for hand part detection, however they fail due to sensor noise and high interfinger occlusion. Additionally, these techniques have difficulties in modelling kinematic or temporal constraints. Although model-based descriptive (for example Markov Random Field) or generative (for example Hidden Markov Model) techniques utilize kinematic and temporal constraints well, they are computationally expensive and hardly recover from tracking failure. This thesis presents a unified framework for 3D hand tracking, using the best of both methodologies, which out performs the current state-of-the-art 3D hand tracking techniques. The proposed 3D hand tracking techniques in this thesis can be used to extract accurate hand movement features and enable complex human machine interaction such as gaming and virtual object manipulation.