BibTeX entry

@PHDTHESIS{200609Fabrice_Caillette,
  AUTHOR={Fabrice Caillette},
  TITLE={Realtime markerless human body tracking},
  SCHOOL={University of Manchester},
  MONTH=Sep,
  YEAR=2006,
  URL={http://www.bmva.org/theses/2006/2006-caillette.pdf},
}

Abstract

The ability to perform efficient human motion tracking is essential in a wide variety of applications such as human-computer interfaces, anthropological studies, entertainment, and surveillance. Markerless human body tracking involves recovering the parameters of a kinematic model from video sequences. This inference problem is made difficult by the noisy and ambiguous nature of camera images. The high dimension ality of the parameter space is also a major challenge, making human-body tracking a very active research area in the computer-vision community. This thesis presents algorithms for real-time human body-tracking based on multiple camera views. A robust volumetric reconstruction technique is first presented, combining shape and colour from multiple views in an hierarchical scheme. Background segmentation and volumetric reconstruction are merged into a single process, with benefits in performance and robustness. The appearance model, used to relate the kinematic parameters to image observations, is composed of Gaussian blobs. This blob-based model is automatically acquired from the data, and updated from the reconstructed volume in an Expectation-Maximisation framework. Our first proposed tracking algorithm recovers the pose of the kinematic model directly from the blobs, using a two-steps inverse kinematics procedure. A second proposed method approaches tracking as a Bayesian estimation problem. To guide the propagation of samples in the parameter space, we propose a predictive model based on the combination of local dynamics and learnt variable length Markov models of behaviour. The evaluation of the likelihood of the candidate model configuration is critical for computational efficiency. We propose a novel evaluation procedure based on the relative entropy between mixtures of Gaussian blobs. The robustness and performance of our system are demonstrated on challenging video sequences exhibiting fast and diverse movements.