Study On Progressive Multi-feature Dense Matching Techniques For Multi-source High Resolution Satellite Image

With the development of the technologies of sensing, aerospace and data communication, the modern remote sensing (RS) technology, providing dynamic, fast, multi-platform, multi-temporal, high-resolution earth observation data, enters a new era. However, for many reasons such as natural conditions and technology restriction and so on, it is difficult to obtain a result in the practical application of the same phase image sequence acquired by the same sensor. Comparably, it may be more convenient to get high resolution satellite images with the same area of different resolutions, different viewing angles and different heights platform. Therefore, in recent years some scholars obey the principles of binocular vision and photogrammetry and RS theories, study the use of multi-source high resolution satellite images to achieve a three-dimensional reconstruction of the target objects, which will greatly expand the value of the use of remote sensing images and has important theoretical and practical significance.In the process of multi-source image three-dimensional reconstruction, the key technology is high-precision intensive matching for multi-source, high resolution satellite images. The study shows that in the same floor covering, with the improvement of the spatial resolution, the satellite image data is increasing dramatically and shows massive data characteristics. Thus it produces a lot of specific questions such as geometric noise increasing, viewing angle changing, aggravation of local deformation and feature variability decreasing, which make it difficult to achieve multi-source dense matching for high resolution with the method used in the low and middle resolution matching with the same (different) source image. The thesis proposes a progressive multi-feature intensive matching method for multi-source high resolution satellite image. The idea takes full advantage of multi-feature advantages through a progressive matching strategy to achieve the purpose of gradually constraint getting intensive matching point. The major research are as follows:(1) Feature matching based on the progressive local geometric constraint model. The thesis focus on the extraction of the homonymy points based on geometric constraint model of epipolars and corresponding lines, and takes full advantage of the initial homonymy points to extract the geometric constraint model, and constraint match of multi-feature is progressively implemented. With the increasing number of homonymy points, the accuracy of geometric constraint model as well as the ability of feature geometric constraints to features is continuously improved and the characteristics of feature information are fully considered in the thesis. Ultimately the problem of the dense image matching is effectively solved.(2) Building RFM inverse model. Based on RFM positive model, the thesis iteratively builds an inverse RFM model through the relationship between the image space and object space. Then the approximation epipolars are generated by the inverse RFM and projection tracking method, and it can be used to limited range of image matching;(3) Building line matching model. The thesis presents a matching method that combines the linear geometric characteristics with spectral characteristics. With the constraint of the approximation epipolar and linear gradient, the corresponding lines are obtained by the vertical distance and the overlapping range based on the rough matching model.(4) Construction, description and matching of virtual point feature set. The thesis presents a virtual feature point set construction method. Aiming at the difficult matching of the virtual feature points, the thesis proposes a feature descriptor method, so that matching is carried out based on local geometric constraint model and spectral measure model. The algorithm greatly improves the quantity and accuracy of the homonymy points.Experimental results show that based on the progressive multi-feature matching methods, the thesis achieves the intensive matching purposes for multi-source high resolution satellite images, and lay a data foundation for the three-dimensional reconstruction.