Study On Pose Estimation Using Single Camera Based On Object Model

The problem of pose(position and orientation) estimation using single camera arises in manyareas of computer vision including object recognition, site inspection and updating, objecttracking and autonomous navigation. How to estimate the pose parameters quickly andaccurately is a hotspot problem which is very urgent to be solved.In chapter 2, the existing approaches for pose estimation using single camera aresummarized and analysed. Most of these approaches determine the pose parameters by assumingthat small sets of points' correspondence are setting manually, and the solution is gained undersolving the problem of perspective projection etc. In order to determine the solutionautomatically, many algorithms such as hypothesize-and-test approach and method of solvingglobal object function are developed to solve the pose and correspondence simultaneously.Until nowadays, no algorithm can solve the correspondence problem efficiently, which existsin current approaches of pose determination potentially or expressly. So a new method isdeveloped in chapter 3 to estimate the pose parameters avoiding the correspondence problemcompletely. In the proposed method, the problem of pose estimation is converted into theoptimization of energy by defining an energy function including the pose parameters. First, the3D model of object is created. Then an energy function is defined by the degree of superpositionbetween the model projection and object area segmented from images acquired by the camera.Finally, Hooke-Jeeves algorithm is used to search for the optimal solution of the energy function.The object is segmented from the image using mean-shift algorithm.The proposed method has some advantages as follows besides avoiding correspondenceproblem: (1) to determine pose parameters automatically; (2) to use as many correspondingpoints as possible to gain a more accurate solution.Finally, the performance of the proposed method is evaluated in a variety of simulations onsynthetic data. The results show that the method is accurate and effective.