Uav Remote Sensing Image Fast Processing Key Technology Research And Implementation

The application of unmanned aerial vehicle remote sensing is more and more widely with its’ high spatial resolution, high efficiency, low cost and high security. With the limit of the flight height and high spatial resolution, the coverage of a UAV image is very small, and we can’t get enough useful information for the whole research region by one image. Therefore we need to stitching together all the images. The traditional method of UAV remote sensing image mosaic is Aerial Triangulation. This method needs a large number of ground control points and human interaction, and it wastes a lot of manpower, financial resources and time. In response to this situation, the research and improvement in this paper is from the following aspects.In terms of feature extraction: Based on CUDA, we use GPU and CPU co-processing mode to achieve the SIFT algorithm. Through the use of GPU’s powerful parallel computing capability, we have greatly reduced the time of feature extraction.In terms of image registration:Based on the coarse matching of the SIFT feature, we use the feature point’s pixel coordinate to exclude the repeated feature points. And then, we removed part of matching points with obvious error by slope constraint. After that, epipolar constraint is used to exclude some wrong matching points. At last, based on Perspective transformation model, we use RANSAC algorithm to remove the error matching points and "out points". By these steps, we guarantee the accuracy of the UAV remote sensing image registration.In terms of image mosaic:In response to accumulated error of UAV remote sensing image mosaic, we formulate UAV remote sensing image stitching as a non-linear optimization problem based on Homography matrix in this work. First of all, we use the least squares method to provide the initialization parameters to our method. Secondly, according to the transformation of each UAV image’s projection center, we use Levenberg-Marquardt algorithm to achieve global optimization for all Homography matrixes. During optimization, all control points are iterated separately to accelerate the processing and improve the accuracy of mosaic. In terms of image fusion:We use A*algorithm to search a best stitching line. During searching the best stitching line, we present a new method. The method uses the center of the overlapping region to make a constraint for searching the best stitching line, so as to reduce the impact of the lens distortion. And then, multi-resolution spline is used to achieve to seamless mosaic based on the best stitching line.