Computational Learning for Hand Pose Estimation

Rapid advances in human–computer interaction interfaces have been promising a realistic environment for gaming and entertainment in the last few years. However, the use of traditional input devices such as trackballs, keyboards, or joysticks has been a bottleneck for natural interactions between a human and computer as two points of freedom of these devices cannot suitably emulate the interactions in a three-dimensional space. Consequently, a comprehensive hand tracking technology is expected as a smart and intuitive option to these input tools to enhance virtual and augmented reality experiences. In addition, the recent emergence of low-cost depth sensing cameras has led to their broad use of RGB-D data in computer vision, raising expectations of a full 3D interpretation of hand movements for human–computer interaction interfaces. Although the use of hand gestures or hand postures has become essential for a wide range of applications in computer games and augmented/virtual reality, 3D hand pose estimation is still an open and challenging problem because of the following reasons: (i) the hand pose exists in a high-dimensional space because each finger and the palm is associated with several degrees of freedom, (ii) the fingers exhibit self-similarity and often occlude to each other, (iii) global 3D rotations make pose estimation more difficult, and (iv) hands only exist in few pixels in images and the noise in acquired data coupled with fast finger movement confounds continuous hand tracking. The success of hand tracking would naturally depend on synthesizing our knowledge of the hand (i.e., geometric shape, constraints on pose configurations) and latent features about hand poses from the RGB-D data stream (i.e., region of interest, key feature points like finger tips and joints, and temporal continuity). In this thesis, we propose novel methods to leverage the paradigm of analysis by synthesis and create a prediction model using a population of realistic 3D hand poses. The overall goal of this work is to design a concrete framework so the computers can learn and understand about perceptual attributes of human hands (i.e., self-occlusions or self-similarities of the fingers) and to develop a pragmatic solution to the real-time hand pose estimation problem implementable on a standard computer. This thesis can be broadly divided into four parts: learning hand (i) from recommendations of similar hand poses, (ii) from low-dimensional visual representations, (iii) by hallucinating geometric representations, and (iv) from a manipulating object. Each research work covers our algorithmic contributions to solve the 3D hand pose estimation problem. Additionally, the research work in the appendix proposes a pragmatic technique for applying our ideas to mobile devices with low computational power. Following a given structure, we first overview the most relevant works on depth sensor-based 3D hand pose estimation in the literature both with and without manipulating an object. Two different approaches prevalent for categorizing hand pose estimation, model-based methods and appearance-based methods, are discussed in detail. In this chapter, we also introduce some works relevant to deep learning and trials to achieve efficient compression of the network structure. Next, we describe a synthetic 3D hand model and its motion constraints for simulating realistic human hand movements. The section for the primary research work starts in the following chapter. We discuss our attempts to produce a better estimation model for 3D hand pose estimation by learning hand articulations from recommendations of similar poses. Specifically, the unknown pose parameters for input depth data are estimated by collaboratively learning the known parameters of all neighborhood poses. Subsequently, we discuss deep-learned, discriminative, and low-dimensional features and a hierarchical solution of the stated problem based on the matrix completion framework. This work is further extended by incorporating a function of geometric properties on the surface of the hand described by heat diffusion, which is robust to capture both the local geometry of the hand and global structural representations. The problem of the hands interactions with a physical object is also considered in the following chapter. The main insight is that the interacting object can be a source of constraint on hand poses. In this view, we employ pose dependency on the shape of the object to learn the discriminative features of the hand–object interaction, rather than losing hand information caused by partial or full object occlusions. Subsequently, we present a compressive learning technique in the appendix. Our approach is flexible, enabling us to add more layers and go deeper in the deep learning architecture while keeping the number of parameters the same. Finally, we conclude this thesis work by summarizing the presented approaches for hand pose estimation and then propose future directions to further achieve performance improvements through (i) realistically rendered synthetic hand images, (ii) incorporating RGB images as an input, (iii) hand personalization, (iv) use of unstructured point cloud, and (v) embedding sensing techniques