This thesis concerns the use of kinematic models and shape models to solve the problem of estimating the motion of articulated objects from image data. The data considered in the first part of the thesis are trajectories of 3D feature points (markers). In the second part, image sequences are considered.
In the first part of the thesis, the problems of estimating 1) the pose of single rigid body, and 2) the configuration of a system of connected rigid bodies, are treated in a unifying manner. An extended Kalman filter is applied to the problem of tracking the pose/configuration, and is shown to give accurate estimates. The Kalman filter framework also provides an improved solution to the common problem of temporary missing marker data. The method is applied to the case of estimating the motion of the distal limb of the horse, and provides the first ever non-invasive \invivo{} estimates of the detailed 3D motion of the distal limb. The estimation of the average center of rotation and axis of rotation is also treated. Four alternative methods are analyzed and contrasted. The methods are found to be biased, and a bias compensation algorithm is derived for one of the methods. Simulation tests show that the bias-compensated method gives a substantial improvement in accuracy.
In the second part, a 3D shape model and a framework for tracking the motion of body segments from multi-view image sequences are developed. The novel shape model, which consists of a stack of circular or elliptic discs, gives a good approximation to the shape of human segments, at a moderate computational cost. The tracking is based on minimizing a measure of the distance between the model contours and the set of edges extracted from the images.
A differentiable model and distance measure are derived, which improve the rate of convergence of the algorithm, and enable the pose to be tracked using an extended Kalman filter. The extended Kalman filter is shown to give reasonable estimates and eliminates the time-consuming minimization step. The thesis concludes with the development of a method for estimating the inertial parameters (mass, center of mass and moments of inertia) of body segments using the proposed stack-of-discs shape model.