Research Of Camera Pose Estimation Based On Moving Viewpoint

Camera pose estimation is an important research topic in computer vision, and it is a key point of the theory and practice on robots localization and navigation, object tracking and recognition, virtual reality and motion estimation to solve. There are two methods can be used to estimate the camera pose parameters. One is the method based on sensors, the other is the method based on the vision information. But currently most research get the camera pose estimation result through the method based on sensors. While estimating the camera pose parameters based on vision was few introduced. In the thesis, a camera pose estimation method based on moving viewpoint is adopted, which only depends on the vision information of the image sequence, and finally form an camera pose estimation method based on vision.Camera pose estimation is based on the camera calibration, so it is necessary to calibrate the camera and get the camera intrinsic parameters, and Zhang’s method was used, then on basis of the calibration result, the improved PnP algorithm was introduced to estimate the camera pose of through the single image. But in order to get the camera pose estimation result based on the moving viewpoint, the feature tracking methods was researched in the thesis. First, the feature extracting methods were discussed, and the Harris corner detector was used to extract features after analyzing. Second, the feature matching methods were researched and normalized correlation gray matching method was used to matching the features coarsely, and then we used the RANSAC algorithm to refine the coarse matched features, deleting the mistake matching. Finally, the feature tracking methods were researched, in order to tracking the features efficiently, a feature storage structure merging the2D and3D information was presented.Based on the tracking feature point pairs, and the6pairs3D points by artificial were provided of the first frame of the image sequence, then the improved PnP algorithm was utilized to estimation the camera pose of the image sequence; but relying on the vision information in the image completely to estimate pose will introduce accumulative error because of the image noise, thus it is necessary to optimize the initial camera pose parameters, sparse bundle adjustment(SBA) method was used to optimize the reprojection error to get a robust camera pose estimation result.In this thesis, the method of camera pose estimation based on moving viewpoint need to be provided3D point coordinates in the first frame of the image sequence, then we can get all the pose parameters of the images. This method plays an important role in the area of object recognition and tracking, robots navigation and the vision odmotry and so on.