3-D Reconstruction From Image Sequence Of UAV And Its Application On Vision-based Navigation

The backgrounds of the dissertation are 3-Dimension(3-D) reconstruction from Unmanned Air Vehicle(UAV) images sequence and its application in vision-based navigation. The researched contents include image feature matching, object's 3-D reconstruction from multi-view images and terrain contour matching based on the terrain recovered from images sequence. It is achieved that the automatic matching of image features, the robust and precise estimating of the 3-D reconstruction from multi-view images, and navigating information extracting form images. Besides, based on the research of the image matching, the three-line array CCD camera images registration of China'CE-1 satellite images is studied, and its super resolution reconstruction is fulfilled.The main content and contributions of this dissertation are presented as following:(1) A series of image features matching methods without the relative geometry information of imaging are proposed. They are as follows:â‘ An image features matching method based on the combination of the robust description of feature and the Least Square Matching(LSM) is proposed. The method employs the matching of feature's descriptive vector to get a rough position and rotation as the initial value for the LSM, and then implement precise matching by LSM.â‘¡A robust and high accuracy matching algorithm called MBVIFP(Matching Based on Valid Invariant Feature Part) is proposed. The algorithm improves the accuracy of the corrupt image matching by employing the invariant feature description and recognition of valid feature.â‘¢A matching method for the images of satellite borne three-line array CCD camera is proposed which represents deformation along the camera motion direction by the quadric function. Based on the matching method the super-resolution (SR) reconstruction of China' CE-1 lunar images is realized.(2) A series of image features matching methods with the relative geometry information of imaging are proposed. They are as follows:â‘ A feature matching algorithm based on the rectification of epipolar-line region is proposed. The algorithm rectifies the regions of the corresponding epipolar-lines which include the feature points to be matched through rotating. Matching on the rectified regions works automatically and gives accurately correspondence through LSM. The algorithm improves the robustness and the accuracy of feature matching method under epipolar constraint.â‘¡An algorithm for image matching uses both epipolar and planar homography constraints is proposed. Based on the principle that the features which are adjacent can form a local plane or quasi-plane, the algorithm uses the correspondences as seeds, and calculate the local homography of the features to be matched. Under the guide of the local homography matrix and the constraint of the epipolar line the normal correlation and LSM matching methods are employed to match the features accurately.(3) For the estimating of the Fundermatal Matrix and Homogrhpy matrix, an algorithm weighed by the matching measure of correspondences is proposed. The evaluation of correspondences is introduced, which is defined as matching measure function. Two matching measure functions are defined according to the correlating and feature-distance-comparing matching methods. The matching measure weight is applied to estimate the Fundamental Matrix(FM) and homography, meanwhile the RANSAC algorithm is employed to overcome the outliers.(4) A novel passive navigation approach by terrain contour match is presented which employs the reconstructed 3-D terrain from images sequence. The sparse control points are tracked in the whole image sequence, which serve as correspondences in solving the relation geometry. And then the optimal local homography estimating algorithm from the control points is presented, which guides the feature points matching by giving the predicted correspondences. At last, the 3-D terrain of the trajectory is reconstructed and matched to the referential terrain map. The approach improves the speed for the real time reconstruction and terrain contour matching to give the navigating information.(5) For the terrain contour matching at the case that there are scaled transformation between the referential map and the real time map, a method based on invariant feature description of the 3-D terrain map is proposed. The vertexes are detected from the terrain map firstly; then the terrain is represented by the invariant feature vector which describes the relative position distribution of the vertexes. The terrain matching approach is realized by comparing the distances of feature vectors which is robust for the similarity transformation of the terrain maps and develops the using field of the 3-D reconstruction from image sequence in the terrain contour matching.(6) Two methods for vision-based navigating information estimating are proposed. They are as follows:â‘ Aim at the problem of horizon detection in the vision-based automatic determination of UAV's gesture, the identification rule and detection algorithm of the horizon are proposed. With the horizon detection the yaw and pitch angle are estimated.â‘¡It proposes a method to estimate the off-plane distance of the objects on the landing plane using homography induced parallax. The quantitative relation between the homography induced parallax and the object's off-plane distance, the imaging parameters is studied. And then the method of calculating the object's off-plane distance is proposed, which gives a flatness estimation of the landing plane.A software system is developed according to methods in the dissertation, which produces the 3-D structure reconstruction from UAV image sequence. The research will expand the application of the 3-D reconstruction and vision-based navigation by UAV image sequence substantially.