无人机图像处理关键技术的研究与实现

发布时间：2018-04-15 11:23

本文选题：无人机 + 鱼眼矫正　；参考：《电子科技大学》2017年硕士论文

【摘要】：近年来,无人机凭借使用方便、成本低廉的优点,成为了各领域应用的“新宠”,如环境监测、媒体报道、电商快递等。由于外界环境复杂,无人机的GPS信号存在丢失或者干扰的可能,此时基于视觉的导航方式就显得尤为重要。目前已有许多针对视觉导航的研究,但已实地应用的方法较少。本文基于无人机自主飞行项目撰写,工作重点之一为研究图像畸变矫正方法和图像拼接技术,工作重点之二是自主降落阶段的地标设计与识别。针对无人机利用航拍相机采集的图片容易产生畸变的问题,本文首先分析了相机的成像原理,选用张正友标定法对相机进行标定。先制作棋盘格标定板,用航拍相机采集标定版不同方位角的图像,提取Harris角点,计算得到相机的内外参数矩阵和畸变系数,再利用极大似然估计对内外参数和畸变系数进行优化。最后,利用畸变系数矫正图像,得到较好的矫正结果。图像拼接时,本文采用基于SIFT特征的图像拼接技术,先提取待拼接地图的SIFT特征点,生成128维的特征向量描述子,计算两张待拼接图片的特征点匹配情况。为了过滤误匹配点,首先采用最小距离和次小距离的比值进行筛选,对筛选出来的候选匹配点对,再利用RANSAC算法进行优化,误匹配点比例降低了74.74%。根据选出的匹配点对计算得到代表图像变换关系的单应性矩阵,用单应性矩阵将待拼接图像像素变换到基准图像坐标系,生成拼接图。本文提出了平均融合法和加权平均融合法对图像的重合区域进行像素融合,均得到了较好的拼接结果。最后,本文实现了多幅地图的自动拼接算法。自主降落部分,本文设计了简单易识别的H地标,对获取的降落图像进行预处理得到灰度图,再将自适应阈值法应用到灰度图的二值化中,提取图像中的轮廓,依据轮廓的长度和面积等信息,过滤出非地标轮廓。最后在候选轮廓中,利用形状的Hu不变矩判断是否是降落标志轮廓。本项目改进了H地标,增加轮廓边缘,并且将自适应阈值法应用于图像二值化中,增强了地标在各种环境下的鲁棒性。本文还提出了一套能够实现低空小范围内视觉降落的四旋翼无人机软硬件方案,并在电子科技大学清水河校区验证了视觉自主降落系统的可行性、稳定性和准确性,实验表明降落精度能够达到设计要求的1米。
[Abstract]:In recent years, unmanned aerial vehicles (UAVs) have become "new favorites" for various applications, such as environmental monitoring, media reports, e-commerce couriers and so on, because of their advantages of convenience and low cost.Because of the complexity of the external environment, the GPS signal of UAV may be lost or interfered, so the vision-based navigation is particularly important.At present, there have been many researches on visual navigation, but few methods have been applied in the field.Based on UAV autonomous flight project, one of the key tasks is to study image distortion correction method and image splicing technology, and the second is to design and identify landmarks in autonomous landing stage.Aiming at the problem that the images collected by aerial camera are easily distorted, this paper first analyzes the imaging principle of the camera, and uses the calibration method of Zhang Zhengyou to calibrate the camera.First, the checkerboard was made, and the images of different azimuth angles of the calibration plate were collected by aerial camera. The Harris corner was extracted, and the internal and external parameter matrix and distortion coefficient of the camera were calculated.Then the maximum likelihood estimation is used to optimize the internal and external parameters and distortion coefficients.Finally, the distortion coefficient is used to correct the image, and a better correction result is obtained.In this paper, the image stitching technique based on SIFT features is used to extract the SIFT feature points of the map to be stitched, and a 128-dimensional feature vector descriptor is generated to calculate the matching of the feature points of the two images to be stitched.In order to filter the mismatched points, the ratio of the minimum distance to the sub-small distance is first used to screen the candidate matching points, and then the RANSAC algorithm is used to optimize the mismatch points. The ratio of the mismatched points is reduced by 74.74 points.According to the selected matching point pairs, the monogram matrix representing the image transformation relationship is obtained, and the pixels of the image to be stitched are transformed into the reference image coordinate system by the homotropic matrix to generate the splicing image.In this paper, an average fusion method and a weighted average fusion method are proposed for pixel fusion of the overlapped region of an image, and good results are obtained.Finally, this paper realizes the automatic mosaic algorithm of multiple maps.In the part of autonomous landing, we design a simple and easily recognizable H landmark, preprocess the landing image to get the gray image, and then apply the adaptive threshold method to the binarization of the gray image to extract the contour of the image.According to the length and area of the contour, the non-Landmark contour is filtered out.Finally, in the candidate contour, Hu invariant moment of shape is used to judge whether it is a descent mark contour.In this project, the H landmarks are improved, the contour edges are increased, and the adaptive threshold method is applied to the binarization of images, which enhances the robustness of the landmarks in various environments.This paper also proposes a software and hardware scheme of four-rotor UAV which can achieve visual landing in low altitude and small range, and verifies the feasibility, stability and accuracy of the visual autonomous landing system in Qingshuihe Campus of the University of Electronic Science and Technology.The experiment shows that the precision of landing can meet the design requirement of 1 meter.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：V279;TP391.41

【相似文献】