Large motion based pose estimation method.

摘要

This thesis investigates pose estimation of a moving camera from 2D images of the 3D world suitable for Augmented Reality (AR) systems. Determining the position and orientation (pose) of an object or a user is one of the most basic problems in computer science. There are numerous approaches for solving this pose estimation problem. Most pose estimation methods involving vision require a priori 3D knowledge about the environment and correspondences between scene features and 2D images. Such information is difficult to acquire for large working volumes and it may not available at all. As a result, most vision-based pose estimation methods are limited to small and static working spaces.; The new method presented here neither imposes restrictions on the structure of the environment nor requires any 3D scene calibration. The pose of a camera is estimated through two 5 degree-of-freedom (DOF) motion estimations, which only require 2D-2D correspondences and natural features. The only requirement is capturing two or more reference images at the known poses prior to operations. To improve the accuracies of motion estimations, the new method suitable for an omnidirectional camera is developed for motions with large displacements and rotations, which are referred to as large motions. Translation direction estimates improve as the distances between camera positions increase. The new method estimate large-motions based on the Implicit Extended Kalman Filter (IEKF) and the Recursive Rotation Factorization (RRF). Using the new motion estimation method, a new Large Motion-based Pose Estimation (LMPE) method is developed. The LMPE method is tested for 50 meters using 15 reference images, two reference images of which are automatically selected for pose estimations. The new approach for pose estimations is unique. I am unaware of any pose estimation method that uses only 5DOF motion estimations. The results from simulated and real data demonstrate the LMPE method is suitable to wide area outdoor AR tracking.