BibTeX entry

@PHDTHESIS{200403Bjorn_Dietmar_Rafael_Stenger,
  AUTHOR={Bjorn Dietmar Rafael Stenger},
  TITLE={Model-Based Hand Tracking Using A Hierarchical Bayesian Filter},
  SCHOOL={University of Cambridge},
  MONTH=Mar,
  YEAR=2004,
  URL={http://www.bmva.org/theses/2004/2004-stenger.pdf},
}

Abstract

This thesis focuses on the automatic recovery of three-dimensional hand motion from one or more views. A 3D geometric hand model is constructed from truncated cones, cylinders and ellipsoids and is used to generate contours, which can be compared with edge contours and skin colour in images. The hand tracking problem is formulated as state estimation, where the model parameters define the internal state, which is to be estimated from image observations.

In the first approach, an unscented Kalman filter is employed to update the model’s pose based on local intensity or skin colour edges. The algorithm is able to track smooth hand motion in front of backgrounds that are either uniform or contain no skin colour. However, manual initialization is required in the first frame, and no recovery strategy is available when track is lost.

The second approach, tree-based filtering, combines ideas from detection and Bayesian filtering: Detection is used to solve the initialization problem, where a single image is given with no prior information of the hand pose. A hierarchical Bayesian filter is developed, which allows integration of temporal information. This is more efficient than applying detection at each frame, because dynamic information can be incorporated into the search, which helps to resolve ambiguities in difficult cases. This thesis develops a likelihood function, which is based on intensity edges, as well as pixel colour values. This function is obtained from a similarity measure, which is compared with a number of other cost functions. The posterior distribution is computed in a hierarchical fashion, the aim being to concentrate computation power on regions in state space with larger posterior values. The algorithm is tested on a number of image sequences, which include hand motion with self-occlusion in front of a cluttered background.