Algorithmic Research On Camera Self-Calibration

In the field of 3D reconstruction in computer vision, camera calibration plays an im-portant role on recovering 3D space information from 2D images. Camera self-calibrationmethods use the image data only, and determine the camera models upto a similarity trans-formation, without an of?ine calibration procedure.Camera self-calibration is closely relevant to stratified reconstruction. When a re-construction is refined to a metric reconstruction, camera models are updated to metriccameras. The root of camera self-calibration is that the absolute conic keeps stay underrigid body motion of camera.In the thesis, we introduce the whole procedure of self-calibration, including com-putation of the fundamental matrix, projective cameras and metric cameras. According tothe non-positive definiteness of DIAC, a semi-definite programming based self-calibrationmethod is introduced. Furthermore, we make use of the camera intrinsic parameters,andconstrain the absolute dual quadric during the solving process. The semi-definite pro-gramming based self-calibration method is essentially a constrained linear self-calibrationmethod, theoretically performs better than the linear method.We compare the traditional linear method and our semi-definite programmingmethod from synthesis data to real data, and illustrate the semi-definite programmingbased method is more robust than the linear method under noise.