Pose Estimation Of Non-cooperative Object Based On Monocular Vision

Machine Vision is one of highlights in the field of computer engineering researchdirections in the last few years. Compared with traditional methods of measurement,measurement method based on machine vision has advantages in speed, simplicity andcost so that it draws more and more attention from researchers. Pose information of theobject is commonly used in motion tracking and aircraft control, so the precision of poseestimation directly affects accuracy of the above problems. In order to achieve accuratepose measurement, this thesis focus on the problem of pose estimation and to be specificthis thesis is dedicated to two sub-problems, solving the projection equations andestablishing the correspondence between2D-3D coordinates. Besides research onalgorithm, a vision measurement system with Bumblebee2industrial camera is setup toverify precision of algorithm.In the aspect of solving the projection equations, this paper comes up with anon-linear optimization method based on POSIT (POS with Iteration). Since POSIT doesnot fully utilized orthogonality of R (rotation matrix), this leads to be a bad result whendepth changes largely. So our method uses the result of POSIT as initial value andorthogonality of R is significantly improved after optimization. The results suggest thataccuracy of the method in this paper is extraordinarily better than POSIT with a littleextra computation.In the aspect of establishing the correspondence between2D-3D coordinates, thispaper introduces local feature descriptor and one of the classical algorithm SIFT (ScaleInvariant Feature Transform). But when tilt is relatively large (more than40degree),SIFT easily gives a wrong matching result or even no matching. So when searchingcorrespondences, ASIFT which considers affine transform is good improvement of SIFT.The results illustrate more correspondence can be found by ASIFT than SIFT, especiallytilt greater than60degree.In the end of this paper, a vision measurement system based on Bumblebee2industrial camera and PC is described. To overcomes the disadvantages of computationcomplexity and provide a more accurate2D coordinate, a local image library of object isestablished. After calculating pose of object, a3D model is displayed by using OpenGL.Results of pose estimation of objects, plane model and human face, present that methods in this paper are able to achieve pose estimation with relative accuracy.