Online pose estimation and model matching

Computation of the relative position and orientation (pose) between a camera and an object from images is a classical problem in photogrammetry and computer vision. Many solution methods have been proposed. Most of them assume that the problem is to be solved in static environments where object models are exact and the correspondences between object and image features are perfectly known. This dissertation addresses the problem of online pose estimation with noisy 3D model observations and with partial or no knowledge of the feature correspondences.; With uncertainties in both 3D object space and 2D image space, object model (structure) and pose must be estimated simultaneously. We present a new error modeling scheme in which error measures in both 3D models and 2D projection are fused in the 3D object space using backprojection. A new pose estimation method is developed based on alternating subspace minimization with which the pose estimation problem becomes a series of progressive absolute orientation problems. The theory and the algorithm are validated using statistical hypothesis tests against a typical 0.05 significance level.; Extensive experiments on controlled synthetic data indicate that the new method is much more efficient than previous nonlinear techniques and is much more tolerant to noise and outliers than linear methods under most conditions.; A robust estimation scheme based on outlier processes is introduced for rejecting outliers in pose estimation. A continuation method is proposed for minimizing the non-convex objective function resulting from robust estimators. Outlier processes are generalized to correspondence processes to solve model matching problems where feature correspondences are unknown.