| Human motion models are of great importance in many application domains. Examples include computer graphics, human-computer interaction, biomechanics, arts and entertainment.; Traditional approaches to modeling human motion use chains of articulated rigid bodies. These representations approximate single parts of the human body as rigid shapes. Although these models can describe effectively coarse full-body motion, they are unable to represent subtle human movements such as skin deformations, changes in facial expressions or non-rigid torso motion.; This thesis describes novel representations and methods for capturing 3D models of human body deformations from single-view video without the use of training data or any prior geometric model. Strong geometric constraints are shown to derive from the proposed non-rigid 3D shape representations. These constraints are used in two ways: first, to disambiguate the 2D motion of feature points in visual tracking; second, to reconstruct non-rigid 3D shape from the estimated 2D motion in the image sequence. The proposed methods are closely related to generic machine learning algorithms, such as factor analysis, probabilistic principal component analysis and identification of linear dynamical systems. The algorithms are demonstrated on challenging video sequences of non-rigid human motion under severe cases of occlusion and variable illumination. |
