%0 Journal Article
%T 基于双特征的丘陵山区耕地低空遥感图像配准算法<br>Low-altitude remote sensing image registration algorithm based on dual-feature for arable land in hills and mountains
%A 宋飞
%A 杨扬
%A 杨昆
%A 张愫
%A 毕东升
%J 北京航空航天大学学报
%D 2018
%R 10.13700/j.bh.1001-5965.2017.0674
%X 摘要 针对丘陵山区耕地小型无人机航拍图像（低空遥感图像）中的尺度变化、几何畸变、图像重叠等问题，提出了基于双特征的丘陵山区耕地低空遥感图像配准算法。该算法鉴于丘陵山区耕地背景环境复杂、光照因素等影响，采用尺度不变特征SURF算法提取了遥感图像的特征点，并构建了能够稳健描述航拍图像几何特征的双特征描述子；在此基础上，以高斯混合模型（GMM）为核心，结合2个单一特征差异描述子（基于欧氏距离的全局特征和基于和向量的局部特征）构造的双特征描述子，得到了能够同时通过2种特征进行对应关系评估的双特征有限混合模型（DFMM），并通过再生核希尔伯特空间（RKHS），基于高斯径向基函数（GRBF）对待配准图像进行了全局与局部结构双约束的空间变换更新。为了验证本文算法的可行性及其性能，采用小型无人机航拍的丘陵山区坡耕地多视角遥感图像开展了实验，将本文算法与SIFT、SURF、CPD、AGMReg、GLMDTPS及PRGLS进行了比较。实验结果表明，本文算法不仅在不同坡度的坡耕地航拍图像多视角配准过程中，均具有较好的鲁棒性，也适用于部分复杂地形小型无人机航拍的多视角遥感图像配准。<br>Abstract：Small unmanned aerial vehicle (UAV) aerial images (low-altitude remote sensing image) of arable land in hills and mountains are confronted with multiple challenges of image processing due to its scale change, geometric distortion and image overlap. To address the problems, the low-altitude remote sensing image registration algorithm based on dual-feature for arable land in hills and mountains was proposed. Due to the complex environment and the influence of light factors in hills and mountains, the feature points of remote sensing images are extracted by using the scale-invariant SURF algorithm. And then the dual-feature descriptor using geometrical structure of aerial images was constructed. On this basis, by taking Gaussian mixture model (GMM) as the core and combining with two single feature difference descriptors (i.e., global distance descriptor based on euclidean distance and local structure descriptor based on sum vectors), the dual-feature finite mixture model (DFMM) was obtained, which can simultaneously evaluate the correspondence between two features. With the reproducing kernel Hilbert space (RKHS), the spatial transformation of the global and local structure of the registration image was carried out based on the Gaussian radial basis function (GRBF). In order to verify the feasibility and performance of the proposed algorithm, experiments were carried out by using UAV images with different viewpoints taken from sloping arable land in hills and mountains. Experimental results show that comparing with SIFT, SURF, CPD, AGMReg, GLMDTPS and PRGLS, our method provides better performances in most cases, and can apply to multi-view remote sensing image registration of other complex terrain by small UAV.
%K 图像配准
%K 小型无人机
%K 双特征
%K 有限混合模型
%K 再生核希尔伯特空间(RKHS)&lt
%K br&gt
%K image registration
%K small UAV
%K dual-feature
%K finite mixed model
%K reproducing kernel Hilbert space (RKHS)
%U http://bhxb.buaa.edu.cn/CN/abstract/abstract14592.shtml