Remote Sensing Image Matching And Target Location Based On Local Features And Topological Constraints

Ground targeting is an important application of UAVs, is a hot research topic in the field of network processing and pattern recognition. Traditional UAV system relies on GPS to navigate and locate. In the case of GPS failure, Inertial navigation, optical measurement methods can be used to locate. However, these methods are more dependent on the device and vulnerable to environmental impacts. So this thesis proposed a targeting algorithm based on matching aerial image to remote sensing images and verified the effectiveness and superiority of this algorithm. The work and achievements of this thesis is mainly reflected in the following aspects:This thesis proposed a targeting framework combining the use of local features and topological constraints. Local features is used to extract and describe the key points. The topology of the key points is used to remove the mismatched points. Finally, got the accurate target location through affine parameter. As traditional feature matching use the feature of the key points only, when there are key points with similar feature in the image, it will cause mismatching, and mismatching points will seriously affect the location accuracy. In order to improve the location accuracy, this thesis combined the local features and topology. Experiments show that this method can improve the final location accuracy.This thesis proposed detecting corner points using multi-scale FAST, and pre-matching using the SIFT descriptor. Multi-scale FAST overcomed the shortcoming of FAST that did not have scale invariance. The generation process of SIFT descriptor lead the feature vectors of key points invariant, and then pre-match the key points in two images based on euclidean distance. The experiments show that Multi-scale FAST has higher accuracy and stability.This thesis proposed using delaunay triangulation to constraint the matching points and remove the mistaken matching points. When the image is scaled and rotated, the relationship between the key points remains unchanged, the sides of the triangulation composed by the key points in matching image keep a certain ratio with the corresponding side in the reference image. Use the proportional relationship to remove the mismatching points.This paper presented estimating parameter model of the local area with RANSAC to get the location of the target. Since there may be distortion in the image during recording, the affine parameter is different in different regions. So estimate the affine model of the local area near the target point with RANSAC as the affine model of the target point. Ultimately target location coordinates in remote sensing image. As can be seen from comparative experiment, positioning use local RANSAC has higher accuracy than the global RANSAC.