| This dissertation introduces a unified data acquisition, processing and synthesis framework for the creation of biomechanically correct human avatars. Contributions include a methodology for the creation of geometric skeletons from video streams, techniques to improve the video capture conditions, and a mathematical model for the treatment of human skeletal motion.; Assuming a 3D environment populated with humans and a network of vision-based sensors monitoring it, the image data can be used to synthesize avatars. Specifics of the acquisition system, such as environment, capture volume, time, budget, required accuracy and desired processing speed are discussed. In this context, techniques for the selection of camera characteristics, camera placement and calibration are introduced. Subsequently, synthetically generated datasets are compared against real data, to validate that engineering precision demands can be satisfied.; Given multiple video streams, a sequence of volumes is computed, and based on their spatiotemporal analysis, a sequence of geometric skeletons approximating the captured subject is extracted. A geometric skeleton describes an object as a combination of geometric entities, together with topological information such as connectivity and extreme locations. In case the geometric skeleton is part of a sequence, temporal consistency is ensured. Furthermore, using a dimensional kinematic synthesis approach, articulated motion is recovered as a kinematic skeleton and a set of parameters describing its inverse kinematics. The kinematic skeleton compactly characterizes the motion of the avatar, allowing for its subsequent animation. The mathematical formulation is based on dual quaternions, supporting the synthesis of R-, S- and T-joints. The initial motion sequence is expressed as a collection of geometric skeletons, for which the coordinate frames corresponding to each limb are consistently aligned over time.; To validate the proposed kinematic synthesis methodology, a quantitative and qualitative analysis was conducted for a number of case studies, based on synthetic and real motion capture data. Algorithm accuracy and execution times are discussed and analyzed in detail, demonstrating the effectiveness of the approach for human motion recovery at interactive rates. |
