Yan Yiming, Su Nan, Zhao Chunhui, Wang Liguo
Department of information and communication engineering, Harbin Engineering University, Harbin 150001, China.
Sensors (Basel). 2017 Sep 19;17(9):2153. doi: 10.3390/s17092153.
In this paper, a novel framework of the 3D reconstruction of buildings is proposed, focusing on remote sensing super-generalized stereo-pairs (SGSPs). As we all know, 3D reconstruction cannot be well performed using nonstandard stereo pairs, since reliable stereo matching could not be achieved when the image-pairs are collected at a great difference of views, and we always failed to obtain dense 3D points for regions of buildings, and cannot do further 3D shape reconstruction. We defined SGSPs as two or more optical images collected in less constrained views but covering the same buildings. It is even more difficult to reconstruct the 3D shape of a building by SGSPs using traditional frameworks. As a result, a dynamic multi-projection-contour approximating (DMPCA) framework was introduced for SGSP-based 3D reconstruction. The key idea is that we do an optimization to find a group of parameters of a simulated 3D model and use a binary feature-image that minimizes the total differences between projection-contours of the building in the SGSPs and that in the simulated 3D model. Then, the simulated 3D model, defined by the group of parameters, could approximate the actual 3D shape of the building. Certain parameterized 3D basic-unit-models of typical buildings were designed, and a simulated projection system was established to obtain a simulated projection-contour in different views. Moreover, the artificial bee colony algorithm was employed to solve the optimization. With SGSPs collected by the satellite and our unmanned aerial vehicle, the DMPCA framework was verified by a group of experiments, which demonstrated the reliability and advantages of this work.
本文提出了一种新颖的建筑物三维重建框架，聚焦于遥感超广义立体像对（SGSPs）。众所周知，使用非标准立体像对无法很好地进行三维重建，因为当图像对在视角差异很大的情况下采集时，无法实现可靠的立体匹配，我们总是无法获得建筑物区域的密集三维点，也无法进行进一步的三维形状重建。我们将SGSPs定义为在约束较少的视角下采集但覆盖同一建筑物的两幅或多幅光学图像。使用传统框架通过SGSPs重建建筑物的三维形状甚至更加困难。因此，引入了一种动态多投影轮廓逼近（DMPCA）框架用于基于SGSPs的三维重建。关键思想是进行优化以找到模拟三维模型的一组参数，并使用一个二值特征图像，该图像能使SGSPs中建筑物的投影轮廓与模拟三维模型中的投影轮廓之间的总差异最小化。然后，由这组参数定义的模拟三维模型可以逼近建筑物的实际三维形状。设计了典型建筑物的某些参数化三维基本单元模型，并建立了一个模拟投影系统以获得不同视角下的模拟投影轮廓。此外，采用人工蜂群算法来解决优化问题。利用卫星和我们的无人机采集的SGSPs，通过一组实验验证了DMPCA框架，这些实验证明了这项工作的可靠性和优势。
