City Research Online

Abstract

Hand pose is emerging as an important interface for human-computer interaction. The problem of hand pose estimation from passive stereo inputs has received less attention in the literature compared to active depth sensors. This thesis seeks to address this gap by presenting a data-driven method to estimate a hand pose from a stereoscopic camera input, with experimental results comparable to more expensive active depth sensors. The frameworks presented in this thesis are based on a two camera stereo rig capture as it yields a simpler and cheaper set-up and calibration. Three frameworks are presented, describing the sequential steps taken to solve the problem of depth and pose estimation of hands.

The first is a data-driven method to estimate a high quality depth map of a hand from a stereoscopic camera input by introducing a novel regression framework. The method first computes disparity using a robust stereo matching technique. Then, it applies a machine learning technique based on Random Forest to learn the mapping between the estimated disparity and depth given ground truth data. We introduce Eigen Leaf Node Features (ELNFs) that perform feature selection at the leaf nodes in each tree to identify features that are most discriminative for depth regression. The system provides a robust method for generating a depth image with an inexpensive stereo camera.

The second framework improves on the task of hand depth estimation from stereo capture by introducing a novel superpixel-based regression framework that takes advantage of the smoothness of the depth surface of the hand. To this end, it introduces Conditional Regressive Random Forest (CRRF), a method that combines a Conditional Random Field (CRF) and a Regressive Random Forest (RRF) to model the mapping from a stereo RGB image pair to a depth image. The RRF provides a unary term that adaptively selects different stereo-matching measures as it implicitly determines matching pixels in a coarse-to-fine manner. While the RRF makes depth prediction for each super-pixel independently, the CRF unifies the prediction of depth by modeling pair-wise interactions between adjacent superpixels.

The final framework introduces a stochastic approach to propose potential depth solutions to the observed stereo capture and evaluate these proposals using two convolutional neural networks (CNNs). The first CNN, configured in a Siamese network architecture, evaluates how consistent the proposed depth solution is to the observed stereo capture. The second CNN estimates a hand pose given the proposed depth. Unlike sequential approaches that reconstruct pose from a known depth, this method jointly optimizes the hand pose and depth estimation through Markov-chain Monte Carlo (MCMC) sampling. This way, pose estimation can correct for errors in depth estimation, and vice versa.

Experimental results using an inexpensive stereo camera show that the proposed system measures pose more accurately than competing methods. More importantly, it presents the possibility of pose recovery from stereo capture that is on par with depth based pose recovery.

Publication Type:	Thesis (Doctoral)
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Departments:	Doctoral Theses School of Science & Technology > School of Science & Technology Doctoral Theses School of Science & Technology > Computer Science

Export

Downloads