Researching Into Extracting And Matching Methods Of Point And Line Features For Wide Baseline Stereo Images

With the fast development of high resolution Charged Coupled Device(CCD) and compute technology, digital close-range photogrammetry based on non-metric digital camera has become a very new research field, and at the same time, digital close-range photogrammetry has also become one of an important way to obtain the 3D spatial information with untouched, flexible and quick style . So, some problems related to this field have become hot topics around home and abroad recently. Among these, camera calibration, feature extraction and matching (especially feature extraction and matching for wide baseline stereo pairs) are the key problems when using this technology, because these problems are the basis for following photogrammetric processes, such as orientating the images, 3D reconstruction and data analysis, etc. The research work of this dissertation focus on previous problems and involved in multiple disciplinary theories and method, such as digital image process, image matching, and pattern recognition, etc, mainly include:1. Recognizing and matching methods of circular signalized points and its application of digital camera calibration were studied. The method of recognizing and matching the circular signalized points based on object geometric constraints was proposed; Two practical algorithms to obtain the initial values of the exterior orientation elements based on the vanishing points theory and the two dimensional direct linear transformation were demonstrated and deduced by combining the collinearity equations in planar scenes; Mathematical model of the bundle block adjustment with self calibration was given and the method of determining the weight of every class of observations based on repeat calculating the QVVP was introduced to improve the calculating speed and the stability.2. Least squares matching methods for wide base-line stereo images based on SIFT features were proposed. In this algorithm, the optimal SIFT features with good spatial distribution and large information content were first selected, then these SIFT features were matched by using the maching method based on singular value decompositon algorithm, adaptive NCC(Normalized Cross Correlation) matching method based on scale and orientation information of SIFT features, then the fundamental and homography matrix can be estimated by using these initial correspondences; Under the dual geometric constraints by using the fundamental and homography matrix, extended matching methods by using weighted least squares matching and multiple scale template windows were developed, further, by compared to the location error of the original correspondences of SIFT feature points, the least squares matching results were determined to adopted or not; In addition, some novel strategies such as NCC computing with high speed based on integral image, multi-level pyrmid image matching, etc, were adopted to improve the matching speed and reliability.3. Line feature description with scale and rotation invariant and matching methods based on epipolar geometry constraint were proposed. In this algorithm, the overlap distances, orientations and line support regions of line segments pair were defined based epiploar geometry constraint, and then a line descriptor with scale and rotation invariant based on the mean and standard deviation of the gray values and gradient vectors was constructed; By fully considering the occlude and disparity discontinuity of the line segment pair, the weighted similarity measure for line segment matching based on the two side line support regions were adopted; Dimensionality reduction method based on kernel principal component analysis was introduced to improve the matching speed.4. Application programms of the previous theories and methods were realized and established base on Windows platform with Visual C++6.0 high level programming languages and OpenCV(Open Compute Vision library), and relative experimental results indicate that the methods proposed in the dissertation is true and effective.Reseach work in this paper can enrich the relative theories and methods concerned on digital close-range photogrammetry, and good basis are provided for 3D reconstruction of wide baseline photogrammetry.