Determining the position and orientation (t and R) of a known object from its image is called pose estimation (PE). PE is one of the central problems in computer vision, robotic manipulation and visual servoing. The difficulty in solving PE problems is mainly due to the fact that the sensory data from which the pose should be determined are imprecise. Choosing proper image features to represent objects is a key factor to estimate pose robustly and efficiently.;In this dissertation, pose estimation using points to regions correspondence is studied. Given the points on an object and the convex regions in which the correspondent image points lie, the estimate of t and R are found explicitly (we call this PE using forward constraints).;Considering there exists partial self-occlusion, a points to regions correspondence based PE algorithm using backward constraints is also put forward. Given the image points and the convex regions in which the corresponding object points lie, this method can find the concrete estimate of pose with satisfying accuracy even when the object is partially occluded.;The pose estimation algorithms have the following advantages: (1) Using regions that make up the objects makes the PE algorithm using forward constraints more robust to imprecise sensory data. (2) Being sensitive to occlusions is the main drawback of the available PE methods using global features. The PE algorithm using backward constraints overcomes this drawback. (3) Using the unit quaternion representation of a rotation matrix makes it possible to implicitly include the constraint that the estimate of R must be a matrix in SO(3). (4) PE by solving a convex LMI optimization is guaranteed to converge and avoids local minima. (5) Our method is the first one that can explicitly provide the estimate of R and t using points to regions correspondence. Most of the available PE methods only use image matching to verify the effect of their algorithms.
展开▼
