HI, here a short abstract for my diploma thesis and a link to download the thesis itslef
In this thesis, a method is developed to reconstruct a human silhouette and to detect real world parameters, e.g. orientation, based on a foreground segmentation extracted from monocular image sequences. The approach uses a 2-D model, which consists of a parametric torso and head shape description, as well as two rectangles to describe each arm and leg. The parametric model description is a newly-developed PCA-based approach, which separates the parameters into individual ones for describing the inter-person variations and general ones for the intra-person variations. This concept allows for a good restriction of the model adaptation to person specific characteristics. The silhouette alignment is based on a linear regression method, which adapts the general parameters of the torso and head, as well as the legs and the arms, to a given foreground segmentation. An additional step uses linear regression to calculate the real world parameters from the general model parameters. The derived method allows for a stable and flexible human silhouette reconstruction, even in noisy images, as well as an accurate orientation detection for a person in an upright pose.
The final thesis can be downloaded here (PDF).