概率体素模型中建筑物精确三维结构提取

Accurate 3D Building Structure Extraction from Probabilistic Voxel Model

从图像中重建的三维体素模型通常存在噪声和建筑物结构不完整的问题,分层边缘拟合方法为解决该类问题提供了一种思路,然而实际场景中建筑物横截面轮廓复杂多样。为此,提出一种两步式由粗到细的边缘拟合算法,在层图像上拟合建筑物横截面轮廓。利用分层投影方法将三维模型投影到二维层图像上,并采用一种结合上下文信息的基于密度的聚类方法去除场景中的噪声,通过形状分类和形状拟合得到平面轮廓的精细拟合结果。最终的三维模型由层图像上边缘拟合结果组合而成。实验结果表明,与最初重建的模型相比,该方法可使精确的建筑物模型更加规则完整且几乎没有噪声,同时大幅减少存储空间。

3D voxel model reconstructed from images often suffers from problems of noise and incoml-Iete building structure. Layered contour fitting method provides a way to solve this kind of problem. However, the cross planar contours are very complicated in the realistic scene. In view of this, this paper proposes a two-step contour fitting method to fit the cross-sectional contours of a building on the layer image. The hierarchical projection method is used to project the initial 3D model into a two-dimensional layer image which is then de-noised by a contextual density-based clustering algorithm. Precise fitting results of planar contours are obtained by shape classification and shape fitting. The final 3D model is composed of the results from the upper edge of the layer image. Experimental result shows that compared with the original reconstruction model, the precise building model has more regular and complete shape with less noise, and substantially reduces storage space.