| The importance of estimating pose and anthropometric measurements from a monocular image sequence stems from its numerous applications such as: (1) performance measurement for human factors engineering; (2) posture and gait analysis or training athletes and physically challenged persons; (3) animation of the human body, hands, and face; (4) automatic annotation of human activities in video databases; (5) control in video games and virtual reality; and (6) teleoperation of anthropometric robots.; In this thesis, we present a monocular optical human motion tracking method for estimating the anthropometry and pose of an individual in motion. Specifically, our model-based method consists of two steps: (1) semi-automatic model initialization by estimating an up-to-scale Virtual Human Model to match the anthropometry and pose of a person in a single image, and (2) automatic tracking. The novelty of our approach is the use of anthropometric statistics to constrain the estimation process of both anthropometry and pose. Moreover, our model-based framework alleviates the complexity of the high-dimensional and nonlinear problem of pose estimation by, first, decomposing it into a hierarchical problem involving three articulated segments from the hips to the limbs and head, and second, using penalty factors to transform the pose problem into a convex problem for each image. Experimental results from challenging data sets are presented to highlight the accuracy, advantages, and limitations of our approach in coping with lighting variation, rapid motion, occlusion of body parts, and lack of camera parameters. |
