Calculation of 3-D Pose of a Known Object in a Single Perspective View

摘要

The thesis describes an algorithm for 3-D pose calculation of known objects from a single perspective view. The pose calculation algorithm consists of four main stages: pose candidate generation, candidate verification, optimization and reliability analysis. In the first stage, pose candidates are generated by matching model and image triangles. An important sub-problem, the so-called 3-point perspective pose estimation problem, is solved numerically. Pose candidates are verified by matching model and image features. The first accepted candidate is then optimized. Three optimization methods are considered: Newton-Raphson, simulated annealing and robust M-estimate using Lorenzian distribution. The reliability estimates are based on covariance analysis. Uncertainties due to both model and image noise are taken into account. A structure optimization method to reduce modeling noise is also presented. The algorithm was implemented on two parallel computers, a Connection Machine-2 (CM-2) and a Hathi-2/16. To evaluate the speed and accuracy of the algorithm, several experiments were performed using both synthetic and real images.