摘要：

精细三维模型是智慧城市建设的关键空间基础信息，而视角变化、遮挡等因素导致基于航空影像生成的三维模型容易出现边缘不准确、孔洞以及建筑物立面纹理模糊等问题。地面影像可以很好地解决倾斜摄影建模底部缺失与区域遮挡的问题，因此，提出一种轻量化多层渐进式特征对齐网络优化的空地影像匹配方法，以实现空地影像的稳健匹配，为城市建模提供一定的技术支撑。设计多层渐进式匹配网络优化策略，利用EfficientNet-B3预训练模型的高层特征图进行双向匹配，取双向匹配的交集作为初始匹配点集。根据初始匹配点对，采用RANSAC策略计算初始单应矩阵，运用该矩阵对地面影像进行图像变换，得到近似空中视角的影像，从而完成特征匹配与粗差剔除。针对空中影像和近空视角影像，在前面多层特征图上进行匹配和优化。在每一层特征图上都计算该层特征图的匹配和对上层匹配点对的位置校正，最终得到精确的匹配点集。以无人机DJI-MAVIC2拍摄的航空影像及手持设备拍摄的地面影像等8组典型数据作为对象进行实验，结果表明，与SIFT、D2-net、DFM等方法相比，该方法具有良好的匹配性能，平均同名点匹配数量较次优方法提升了1.3倍。

关键词: 三维模型, 多层渐进式特征对齐网络, 空地影像, 渐进式匹配与优化, 图像匹配

Abstract:

Fine 3D models provide the key spatial basic information for smart city construction. However, factors such as perspective changes and occlusion lead to inaccurate edges, holes, and blurry building facade textures in 3D models generated from aerial images. Ground images can effectively solve the problems of missing bottoms and regional occlusion in oblique photography modeling. Therefore, a lightweight aerial-ground image matching method optimized by multi-layer progressive feature alignment is proposed to achieve robust matching of aerial-ground images and provide certain technical support for urban modeling. A multi-layer progressive matching network optimization strategy is designed, utilizing the high-level feature maps of the EfficientNet-B3 pre-trained model for bidirectional matching by taking the intersection of the bidirectional matching as the initial matching point set. Based on initial matching point pairs, the RANSAC strategy is used to calculate the initial homography matrix, thereby using it to transform the ground image, to obtain an image with an approximate aerial perspective, which completes feature matching and gross error removal. For aerial and near-field perspective images, matching and optimization are carried out on the previous multi-layer feature maps. In calculating the matching of each layer's feature map, the position of the upper layer's matching point pairs is corrected on each layer's feature map, to ultimately obtain an accurate set of matching points. Experiments are conducted on eight sets of typical data, including aerial images captured by the drone DJI-MAVIC2 and ground images captured by handheld devices. The results demonstrate that the proposed method has good matching performance compared to SIFT, D2-net, DFM, and other methods, with an average 1.3x increase in the Number of Correct Matches (NCM) compared to the suboptimal method.

Key words: 3D model, multi-layer progressive feature alignment network, aerial-ground image, progressive matching and optimization, image matching