Marker-free human motion capture in dynamic cluttered environments from a single view-point

Human Motion Capture is a widely used technique to obtain motion data for animation of virtual characters. Commercial optical motion capture systems are marker-based. This thesis is about marker-free motion capture. The pose and motion estimation of an observed person is carried out in an optimization framework for articulated objects. The motion function is formulated with kinematic chains consisting of rotations around arbitrary axes in 3D space. This formulation leads to a Nonlinear Least Squares problem, which is solved with gradient-based methods. With the formulation in this thesis the necessary derivatives can be derived analytically. This speeds up processing and increases accuracy. Different gradient based methods are compared to solve the Nonlinear Least Squares problem, which allows the integration of second order motion derivatives as well. The pose estimation requires correspondences between known model of the person and observed data. To obtain this model, a new method is developed, which fits a template model to a specific person from 6 posture images taken by a single camera. Various types of correspondences are integrated in the optimization simultaneously without making approximations to the motion or optimization function, namely 3D-3D correspondences from stereo algorithms and 3D-2D correspondences from image silhouettes and 2D point tracking. Of major importance for the developed methods is the processing time and robustness to cluttered and dynamic background. Experiments show, that complex motion with 24 degrees of freedom is track-able from a single stereo view until body parts get totally occluded. Further methods are developed to estimate pose from a single camera view with cluttered dynamic background. Similar to other work on 2D-3D pose estimation, correspondences between model and image silhouette of the person are established by analyzing the gray value gradient near the predicted model silhouette. To increase the accuracy of silhouette correspondences, color histograms for each body part are combined with image gradient search. The combination of 3D depth data and 2D image data is tested with depth data from a PMD camera (Photonic Mixer Device), which measures the depth to scene points by the time of flight of light