针对倾斜航空摄影特点与匹配处理要求,提出了融合尺度不变特征转换SIFT与半全局匹配SGM的倾斜影像密集匹配方法,包括两个阶段:(1)引入局部二阶矩变换的SIFT倾斜影像稀疏匹配。利用与二阶矩特征值有关的Hessian-Gabor算子提取影像初始特征,通过窗口二阶矩变换及椭圆归一化处理去除特征邻域的仿射变化,使得改进SIFT的特征描述符仿射不变性大大增强,满足宽基线倾斜像对稳健匹配要求的同时并能获得较多的匹配特征作为后续SGM优化计算的路径约束条件;(2)路径受限优化下的SGM倾斜影像密集匹配。以倾斜影像SIFT成功匹配像素的互信息为约束,对SGM的动态规划路径进行分段纠正,减小错误匹配代价的传播并加速最优路径搜索过程;以匹配像素的离散视差信息为基础,基于TPS变换生成良好初始视差图,以提高SGM互信息计算的可靠性并提高计算效率;对理想水平像对下的摄影测量水平核线重排过程进行扩展,整体旋转摄影基线以消除分量BX,BY的影响并建立虚拟"水平"像空间辅助坐标系,从而利用相对定向参数生成沿扫描线方向的"水平"核线影像以满足SGM应用要求。倾斜影像密集匹配试验结果证明了算法的有效性,可为后续摄影测量DSM自动生成或3维快速重建工作提供可靠、逐像素的密集匹配点。

This paper presents our research on developing a dense image matching algorithm to generate a high-quality disparity map from airborne oblique stereo images. We conducted off-site measurements and analysis not on models but on the actual aerial photography of a site. The Oblique Photogrammetry System (OPS) was influenced by and significantly affected photogrammetry. Oblique images released by OPS are becoming an indispensable tool for general use, such as tax assessment and building deviation, urban and infrastructural planning, management of military and security operations, and critical infrastructural protection. However, processing oblique images is quite challenging in terms of automation and accuracy. For example, oblique image stereo matching is subjected to various obstacles, such as obvious illumination differences, serious occlusions, discontinuous object boundaries, and low or repetitive textures. To address such problems, an image matching approach based on the integration of improved SGM and SIFT algorithm is proposed to generate a dense disparity map from an airborne oblique image pair, which establishes a basis for automatic photogrammetry Digital Surface Model (DSM) generation or Computer Vision three dimensional visualization. The proposed approach is composed of two stages: (1) The first stage is sparse oblique image matching with improved SIFT algorithm, in which the affine invariance of the SIFT descriptor is enhanced by implementing local two-order moment transformation to feature neighbors detected by using Hessian-Gabor algorithm. Thus, robust matching results for various wide-baseline oblique image pairs are obtained and used to estimate epipolar geometry model or as path constraints incorporated into SGM in next stages. (2) The second stage is dense oblique image matching with constrained SGM algorithm, in which mutual information from matched features are used as an unchanged "anchor" to block the propagation of mistaken matching cost along some SGM paths and based on TPS transformation. Discrete parallax from matched features is used to generate a continuous disparity map for reliable SGM mutual information computation. In addition, to generate the epipolar image input into SGM, an ideal photogrammetry image leveling algorithm is extended to oblique image pairs with various wide-baseline conditions by globally rotating wide-aseline vector to be horizontal. Then, orientation parameters of oblique image pairs are estimated with a matched feature coordinate by using photogrammetry relative orientation process. Selected five airborne oblique images from OPS with cameras positioned in the configuration of a Maltese cross are tested by using the proposed approaches. Thus, a high-quality dense disparity map is obtained. The following results were obtained: (1) The incorporation of Hessian-Gabor feature detector and local two-order moment transformation into the SIFT algorithm greatly enhances its affine invariance. Thus, reliable and rich conjunctive pixels can be obtained in various wide-baseline oblique image pairs, and these pixels are intended to be used as strong constraint for further dense matching. (2) Improved SGM could block the propagation of mistaken matching and accelerate path searching by means of mutual information and discrete parallax knowledge from known conjunctive pixels. Thus, a high-quality dense match result from oblique images can be obtained.