Research On Several Key Issues In Image Sequence Of UAV Preprocessing And Matching

The UAV (Unmanned Aerial Vehicle) has been opened from military to civilian. Because the UAV has many advantages including light weight, small size, high performance, high image resolution, the low operating time and space requirements, effectively working in the danger zone in the absence of circumstances, etc. UAV technology has been widely used and rapidly developed in many areas. Integrating UAV and new remote sensing equipment will be a great complement of aerial remote sensing, its applications also continues to expand, expanded from the initial reconnaissance, early warning and other military areas to resource exploration, meteorological observations and deal with emergencies and other non-military areas.This thesis introduced the composition of low-altitude UAV remote sensing systems, described the key steps and precautions of using the UAV to obtain images, deeply analyzed several factors which is needed to be considered and the specific design methods of UAV route planning. Moreover, this paper conducted some quality assessments for the acquired UAV images. Because of the features of the UAV images including big quantity; large distortion and poor visual contrast of part of the region’s images, it’s very important to make necessary preprocessing of UAV images. This paper researched the distortion correction, image enhancement, dodging and other aspects of preprocessing the UAV images. Though studying the theoretical aspects of the UAV matching and image stitching, this paper proposed the polygon matching with considering distance, angle, weights and other aspects, and integrating into adjustment theory to image stitching. Ultimately, it generated a three-dimensional terrain of experimental area. Specifically, the research work and innovations in this dissertation are mainly the following several aspects:1. Due to the features like many houses, roads and other rules of graphics in original UAV images, this paper used the canny edge to contour extraction, and based on Hough transform theory get a lot of feature contours. On the basis of this, the anti-solving model contact the object and image coordinates was used, from the object side to push out the real image point coordinates with the step approximation method to correct and check the results. Meanwhile, there are many parallel and vertical characteristics between the feature contours, so quadrilateral connection method and parallel line restraint method to correct the image was proposed. Through the experiments comparing the coordinates of the control points before with after correcting distortion, the corrective effect in the maximum area can reach 6m, contrast the positional relationship of the building between before and after image correction, we can find out that a lot of contour lines have been greatly improved, the experimental results were quite satisfactory.2. According to the discrete gray information of the original UAV images, each discrete pixel gray value and the average gray scale of the entire image is calculated and the method of automatically divide the image into different regions is proposed. Then, the transfer functions based on the distance-weighted interpolation are used to process each region of the image in different degrees and correct the histogram of the enhanced images. Finally, the nuclear line constraint method is used to match the images and test the uniformity of corresponding points. The experimental results of part of the images indicated that, because the gradient difference of gray value increases in some areas has nearly doubled after image enhanced, the whole tie points on the image are increased nearly 80 percent, especially in the area of large distortion and the woodland which has small gray changes, the tie points has increased nearly 1.5 times than the original image. Moreover, by contrasting the uniformity of the tie points between before and after enhancement, it found that the uniformity has been improved to some extent; it is conducive to subsequent images processes.3. Due to the problems that the UAV images exist large and complex geometric deformation which may increase the probability of false match, this paper uses the Harris operator to extract and match feature points, proposed a polygon-based matching method to detect and eliminate false match corresponding points. This method can ensure the correctness of matching with considering the distance, angle, weigh and many other aspects. The experimental statistical results for each region of point matching accuracy and speed of operation show that, the matching success rate is over 90% in the obvious characteristics areas and more than 80% in the less change texture woodland area. Compared with the traditional methods, the proposed method improved nearly 10%. Meanwhile, the computing speed is almost the same of traditional method, which verified the reliability and validity of the algorithm.4. Because the UAV image stitching process may cause lots of accumulated error, it will face certain difficulties in fast and accurately getting a wide range of panorama. Based on this, this paper proposed a precise and efficient stitching method of sequence UAV images. This method effectively reduces the time of feature searching and matching through calculating the approximate area. During the process, it records the coordinates of center point, and then distinguishes the different area of plains, hills, mountains, etc. and set different weights for these points to correct the coordinate position of matching feature points by using the adjustment theory. For the small overlap rate between hang with and large attitude differences, the suggested method uses the stitching approach of "first hang with next hang, every hang closer to the middle" and reduces the overall area of the accumulated error, the final completion of the overall image stitching. Experimental results show that this method enhances the stitching speed nearly 15% compared with the direct stitching method; according to the last stitching renderings, it can also find out that the error rate has been greatly improved, many of the road and other obvious surface features are stitched very correct.